# Ranomics — Full Technical Library
Full text of all published blog articles from https://ranomics.com. Generated for AI crawlers and citation engines.
---
## 5 Developability Red Flags That Kill mAb Programs
> The five antibody developability red flags most commonly responsible for late-stage CMC failures, how each one shows up in the clinic, and which of them you can detect from sequence before synthesis.
Source: https://ranomics.com/5-developability-red-flags-that-kill-mab-programs/
Published: 2026-04-21
import InOurWork from "../../components/InOurWork.astro";
A failure in developability late in CMC is the most expensive kind of failure a monoclonal antibody program can suffer. Discovery has spent two years reaching a candidate with sub-nanomolar affinity, cell-based potency, and clean animal PK. The formulation team then reports viscosity at 150 mg/mL above pumpability, or a monomer fraction that collapses after four weeks at 40 C, or a mass shift in stability studies that traces back to a single Asn in the heavy chain CDR2. The program stalls, the scientists go back to re-engineer, and a year of clinical timing quietly disappears.
The inconvenient truth surfaced by the industry-wide Jain et al. 2017 panel of 137 clinical-stage antibodies (Proc. Natl. Acad. Sci. USA, 114:944 to 949) is that poor developability correlates visibly with the stage at which an antibody stops advancing. Many of those liabilities would have been visible from sequence before any wet-lab work. This post covers the five specific red flags that most reliably kill programs, how each one actually manifests, and which of them are detectable before synthesis.
## Red Flag 1: Deamidation Hotspots in CDRs (NG, NS, NT, NH)
Asparagine deamidates to aspartate or isoaspartate via a succinimide intermediate. The reaction is strongly dependent on the residue that follows the Asn. Asn followed by Gly is the textbook hotspot, with a half-life measured in days at physiological pH and temperature. Asn followed by Ser, Thr, or His is slower but still clinically relevant over months in storage. Once a CDR Asn has deamidated, the charge, local backbone geometry, and sometimes the potency of the antibody shift. You end up selling a heterogeneous product whose potency drifts over shelf life.
The clinical picture is well documented. The Vlasak et al. 2009 work on deamidation in Herceptin identified a CDR Asn that accumulated iso-Asp during storage and required careful formulation control. Several IgG4 programs have pulled back on CDR Asn residues after observing charge heterogeneity in capillary isoelectric focusing at release. Deamidation is one of the most detectable liabilities. Any sequence-scanning tool can find NG, NS, NT, and NH motifs in the six CDRs in seconds.
Detectability before synthesis: very high. Pure sequence motif search.
## Red Flag 2: Asp Isomerization Sites (DG, DS, DH in CDRs)
Aspartate isomerization runs through the same succinimide mechanism as Asn deamidation, converting Asp to iso-Asp in the backbone. The canonical motif is DG, with DS and DH also observed. The consequence is a backbone rearrangement that often sits directly inside the paratope. For an antibody that depends on a specific CDR conformation to contact its antigen, isomerization of a single Asp can drop affinity by an order of magnitude.
The Cacia et al. 1996 study on a recombinant antibody (Biochemistry, 35:1897 to 1903) showed that isomerization of a single CDR Asp produced a clinically relevant loss of potency during storage. This is one of the liabilities that converts a nanomolar binder into a double-digit nanomolar binder over six months on a pharmacy shelf, which is a regulatory problem and a commercial one.
Detectability before synthesis: very high. Same scan as Red Flag 1, different motif.
## Red Flag 3: N-Linked Glycosylation Sequons in Variable Regions (N-X-S/T)
The N-X-S/T sequon (where X is any residue except Pro) is the site of N-linked glycosylation in eukaryotic expression systems. In the Fc, this sequon is expected and functional. In the variable region, it is almost always unwanted. A glycan on the V region introduces size and charge heterogeneity, often reduces affinity (the glycan physically sits in or near the paratope), complicates CMC analytics (you now have a distribution of glycoforms where you wanted a single species), and can create immunogenicity concerns if the glycan is non-human.
In mammalian expression, V-region sequons are filled at variable occupancy, which means the released product is a mixture of glycosylated and unglycosylated species with different activity profiles. This is a known failure mode in antibodies derived from yeast display campaigns run without a post-campaign sequon scrub, and in de novo designed binders, where the sequon can appear purely by chance during ProteinMPNN sequence design.
Detectability before synthesis: very high. Simple regex for N-X-S/T in the variable region.
## Red Flag 4: Exposed Hydrophobic Patches
Hydrophobic surface patches drive two developability failures at once: non-specific binding to polyspecific reagents like cardiolipin, baculovirus particles, or membrane preparations, and aggregation in concentrated formulation. Polyspecificity is measured experimentally by the polyspecificity reagent (PSR) assay described by Xu et al. 2013 (Protein Eng. Des. Sel., 26:663 to 670) and by hydrophobic interaction chromatography (HIC) retention time. Both correlate with exposed hydrophobic surface area computed from a homology or predicted structure.
The Jain 2017 panel showed that clinical-stage antibodies with late-stage developability issues tended to cluster at the high end of HIC retention and PSR signal. Molecules that progressed cleanly tended to sit in the middle of the distribution. The mechanism is not surprising. Every time you concentrate to 150 mg/mL for subcutaneous delivery, hydrophobic patches on adjacent molecules find each other, and the result is reversible self-association, elevated viscosity, and elevated aggregation on long-term storage.
This is a structural property, not a sequence one. Catching it before synthesis requires a model of the antibody (a reasonable-quality homology model from ABodyBuilder, IgFold, or AlphaFold 3 is usually sufficient) and a solvent-exposed hydrophobic surface calculation. You can get a rough first pass from sequence by counting the fraction of CDR residues that are hydrophobic, but the reliable answer is the structural one.
Detectability before synthesis: moderate. Needs a structure model, not just sequence.
In our pipeline, the polyspecificity story usually shows up before any of the sequence-level red flags fire. A design that looks like a strong on-target binder on SPR or BLI can simultaneously be a sticky multi-target binder that grabs cardiolipin, baculovirus prep, every other plate well, and the BSA in the blocking buffer. The SPR result will not surface that. The display screen does, because the off-target counter-stain runs in the same experiment as the on-target stain, and we treat the off-target signal as a hard filter before any candidate is reformatted for recombinant purification.
## Red Flag 5: Aggregation-Prone Regions (High APR Score)
Aggregation-prone regions are stretches of sequence that are thermodynamically predisposed to assemble into cross-beta amyloid-like fibers or amorphous aggregates. The scoring approaches go back to TANGO and Aggrescan (Conchillo-Sole et al. 2007, BMC Bioinformatics, 8:65) and have been refined in SAP, CamSol, and SolubiS. All of them essentially ask: is this stretch of hydrophobic residues sitting in a context where it is solvent-exposed and flanked by the right charge and secondary structure to nucleate aggregation.
High APR scores correlate with the antibodies that quietly fail at phase 1 for reasons the clinical report describes as "manufacturing issues." The candidate expresses at reasonable titers, binds its target, works in animals, and then fails a stability study at the 12-month mark because the aggregate fraction crossed the specification. These failures are under-reported in the literature because programs do not publish failures at that stage, but they are common in industry experience and they have essentially the same signature every time: a CDR or framework loop with a run of hydrophobic residues that an APR calculator flagged early, and that was dismissed because it did not affect initial potency.
Detectability before synthesis: high for sequence-based APR (TANGO, Aggrescan, CamSol profile), moderate for structure-aware APR (SAP, spatial clustering of hydrophobicity on the modeled structure).
## How Much of This Is Catchable From Sequence Alone
Three of the five red flags above (deamidation, isomerization, V-region sequons) are pure sequence-level scans. They take milliseconds per antibody and they have no false negatives for the motif itself, though real reaction rates depend on local solvent exposure and flexibility. The remaining two (exposed hydrophobic patches, aggregation-prone regions) benefit from a structural model but can be approximated from sequence.
This means that for the cost of a model-building step you do not have to run (IgFold and ABodyBuilder2 take seconds per antibody, and AlphaFold 3 has made single-chain structure prediction close to free), the majority of the most expensive late-stage developability failures are visible before you order a single gene. The actual cost of this triage is not compute time. It is building the pipeline, picking defensible thresholds, and making the output readable to a scientist who is going to make a go or no-go call.
This is why we built [Developability Scout](/technology/developability-scout). It is a free tool that takes an antibody sequence and runs the five-red-flag scan plus the structure-aware hydrophobic and APR analysis in a single pass, with a scored report. It is designed for the stage after discovery has a hit list and before the synthesis order goes in.
## Where AI-Designed Binders Typically Land on These Dimensions
One practical note for teams running de novo or semi-de-novo antibody and nanobody campaigns with RFdiffusion, BindCraft, or ProteinMPNN. Those algorithms are optimized for different objectives than developability. RFdiffusion optimizes a diffusion objective that produces plausible backbones, BindCraft optimizes for AlphaFold-predicted interface metrics, and ProteinMPNN optimizes sequence recovery on training structures. None of these three loss functions penalize deamidation motifs, variable-region sequons, exposed hydrophobic paratopes, or APR hotspots.
Empirically, raw generative outputs tend to ship a meaningful fraction of sequences with at least one of the five red flags. Paratope-exposed hydrophobics are especially common because the training signal pushes toward interfaces with strong VDW contacts, which means burying a Phe, Trp, or Leu into the target surface with part of it facing solvent. Running the same five-red-flag scan on generative outputs before synthesis is the cheapest developability filter you can apply to a design campaign.
## A Practical Workflow for the Triage
The minimal workflow that catches these five red flags before wet-lab commitment, for either discovery-stage hits or de novo designs:
1. Run a sequence-level scan for NG/NS/NT/NH (deamidation), DG/DS/DH (isomerization), and N-X-S/T (V-region sequons). Flag everything in a CDR.
2. Fold the antibody (IgFold, ABodyBuilder2, or AlphaFold 3). Compute solvent-exposed hydrophobic surface area across the V region.
3. Run an APR score (Aggrescan, CamSol profile, or SAP) on the folded model.
4. Threshold and triage. For discovery-stage hits, decide which liabilities you can fix with point mutations and which require re-discovery. For de novo designs, filter out the worst offenders and regenerate if the pool is too depleted.
5. Advance the survivors to synthesis.
For teams without an in-house developability pipeline, [Developability Scout](/technology/developability-scout) runs steps 1 through 3 automatically and produces a per-antibody scorecard. It is free, browser-based, and does not require installing anything.
## Where This Fits in a Campaign
Catching these red flags before synthesis is the highest-leverage developability investment a discovery or design team can make. It does not substitute for a real CMC team, and it does not replace experimental confirmation by DSF, DLS, HIC, SEC, or stability studies. What it does is prevent the most common preventable failures from consuming six months of clinical timing at the far more expensive end of the pipeline.
The teams that run this kind of triage at every stage (discovery hits, affinity-matured leads, de novo designs, humanized candidates) rarely ship liabilities downstream. The teams that skip it rediscover the same five red flags in CMC at ten times the cost. In our work, the cheapest single intervention is running the full five-red-flag scan on every batch of AI-designed candidates before any gene is ordered. The pipeline-investment cost is small and it routinely catches a meaningful fraction of a generative pool before any wet-lab cost is committed.
If you want the full de novo design service with developability filtering baked into the pipeline by default, that is what the [AI Binder Sprint](/ai-binder-sprint) is scoped to deliver. If you have an existing hit list or a set of de novo designs and want to do the triage yourself, [Developability Scout](/technology/developability-scout) is the free tool for that. Either path is better than finding the problems in a phase 1 stability study.
{/* ranomics:related-services */}
## Related Ranomics services
- **[Developability Scout](/technology/developability-scout):** Free sequence and structural scan for the five common antibody developability red flags.
- **[AI Binder Sprint](/ai-binder-sprint):** Full de novo design program with developability filtering integrated into the pipeline.
---
## ai-design-vs-library-screening-when-to-use-which.mdx
Source: https://ranomics.com/ai-design-vs-library-screening-when-to-use-which/
import InOurWork from "../../components/InOurWork.astro";
import KeyTakeaway from "../../components/KeyTakeaway.astro";
This question comes up in almost every early conversation about binder discovery programs: should we use a library, or should we try computational design? The framing is wrong. These are not competing methods. They solve different problems. Choosing between them depends on your target, your timeline, and what you already have.
AI de novo design and library screening are not competing methods. Use library screening for targets with accessible epitopes and precedented scaffolds. Use de novo design for novel target classes, concave epitopes, or targets where prior libraries failed. The strongest programs couple both: design to find a scaffold, display to mature it.
## Library screening: mature, scalable, and sequence-space limited
Library-based binder discovery (phage display, yeast display, ribosome display) works by sampling diversity within a defined scaffold and selecting against a target. The approach is well-validated across decades of antibody and alternative scaffold discovery programs. Its advantages are well understood:
- Proven hit rates on soluble extracellular targets with accessible epitopes
- Straightforward path from hit to lead: affinity maturation, selectivity profiling, expression characterization
- Regulatory precedent for antibody scaffolds in clinical programs
- Predictable cost structure
The limitation is equally well understood: you can only find what your library contains. Library diversity is sampled from a finite, pre-defined sequence space, typically randomization within a known scaffold framework at positions selected for tolerance to variation. For a target with a well-characterized epitope and a precedented binding geometry, this is sufficient. For targets where the right answer is geometrically outside the library's distribution, it is not.
A useful rule from our binder campaigns: when the target is well validated and the design itself is high confidence (a CDR grafted onto a known scaffold, hotspots well mapped, mode of binding established), a small panel of 100 to 500 candidates is usually enough. When the epitope is bespoke and the binding mode is unknown, the uncertainty has to live somewhere, and the cheapest place to put it is a display library. Skipping the library at that stage is the failure mode we see most often.
## De novo design: strongest where libraries fail
Diffusion-based de novo design (RFdiffusion, BindCraft, Boltzgen) generates binder candidates computationally, guided by the three-dimensional structure of the target. The candidates are not derived from any prior scaffold. They are generated from scratch, conditioned on the epitope you specify.
This is strongest in the following situations:
**Novel targets with no existing binder scaffold.** If you are working on a target class where no antibody, nanobody, or alternative scaffold has been validated, library screening requires building (and validating) a library specifically for that target type. De novo design sidesteps the scaffold selection problem.
**Challenging epitopes.** Concave binding surfaces, protein-protein interface grooves, receptor pockets, and enzyme active sites are geometrically constrained in ways that conventional antibody frameworks do not naturally engage. Miniprotein and custom scaffold design allows you to specify the approach geometry directly.
**Cases where immunization is not practical or fast enough.** For research tool binders, diagnostic reagents, or programs where the timeline does not accommodate an immunization campaign, de novo design can start as soon as you have a target structure.
**Targets where multiple prior library campaigns have failed.** If three library campaigns have produced no confirmed binders, the problem is likely the scaffold's inability to access the relevant epitope, not a problem that another library will solve.
---
**Identify epitopes on your own target.** [Epitope Scout](https://scout.ranomics.com) scores and ranks surface patches on any PDB structure. Free to use.
---
## Decision framework
| Situation | Preferred approach |
|-----------|-------------------|
| Target has accessible epitope, precedented scaffold geometry | Library screening |
| Target is a novel protein class with no prior binders | De novo design |
| Timeline: immunization is an option | Library or hybridoma |
| Timeline: 6-8 weeks to first confirmed hits | De novo design |
| Epitope is a flat or accessible extracellular domain | Library screening |
| Epitope is a concave pocket, PPI interface, or recessed groove | De novo design |
| Prior library campaigns have failed | De novo design |
| Binder needs to be a full-length IgG | Library screening |
| Binder format is flexible (nanobody, miniprotein acceptable) | De novo design |
| Need high diversity around confirmed lead | Affinity maturation (post-screen) |
## The most powerful programs couple both
The distinction between de novo design and library screening is sharpest at the start of a discovery program, when no binder exists. Once you have a confirmed hit, even a weak one, the problem changes. Affinity maturation, stability optimization, and selectivity profiling are all better addressed by library methods (deep mutational scanning, focused diversity) than by returning to de novo generation.
The strongest integrated approach is: de novo design to identify a starting scaffold that engages the target, followed by display-based affinity maturation to optimize the lead. Our Sprint program uses this structure. In our campaigns, a single AI design round typically outputs on the order of 25,000 to 50,000 candidates, which is small enough to flow directly into yeast or mammalian display and far below the diversity ceiling phage was built for. The display screen simultaneously validates the design output and provides the selection pressure needed to enrich higher-affinity variants.
De novo design has collapsed the front of the discovery timeline. Experimental validation and optimization are still rate-limiting, and still require the same rigorous display and sequencing infrastructure that library campaigns depend on.
---
**Talk to us about which approach fits your target:** [Contact Ranomics](/ranomics-contact)
{/* ranomics:related-services */}
## Related Ranomics services
- **[Biotechnology services](/biotechnology-services):** Combined design + library programs matched to the target class.
- **[AI Binder Sprint](/ai-binder-sprint):** When de novo design is the right call: our flagship program.
---
## AI-Driven Protein Design: An Honest CRO Perspective
> What AI-driven protein design actually delivers in 2026: where it works, where it fails, and what wet-lab validation reveals.
Source: https://ranomics.com/ai-driven-protein-design/
Published: 2026-04-30
AI-driven protein design covers a specific set of methods — diffusion models, hallucination-based scoring, sequence design via inverse folding — applied to a specific problem: generating proteins with predefined function from scratch. The field went from research curiosity to production tool in about three years. This is what currently works and what doesn't, from a CRO that runs both the design and the wet-lab validation.
## What "AI-driven" actually means in protein design
Three model families do most of the work in 2026:
**Diffusion models** (RFdiffusion, RFantibody) generate protein backbones by reversing a noising process. They are conditioned on target hotspots and produce 3D coordinates of plausible binder structures. We've covered the operational details in [RFdiffusion in Practice](/resource-hub/how-rfdiffusion-works-protein-designers-guide).
**Hallucination + scoring pipelines** (BindCraft) iteratively propose backbone-and-sequence pairs and score them against a target. They produce fewer designs per run than diffusion, but each design is more structurally plausible because the scoring prunes during generation rather than after.
**Inverse folding models** (ProteinMPNN, ESM3) take a backbone and predict amino acid sequences likely to fold to it. They are the bridge between structure-generation and synthesizable DNA.
A typical campaign chains them: RFdiffusion generates 50,000 backbones → ProteinMPNN designs 5 sequences per backbone → AlphaFold2 or Boltz-2 predicts and confirms structure → developability filters cut the list to ~500 → wet-lab synthesis. The whole pipeline runs in 24-72 hours on H100 infrastructure.
## What works in 2026
**Hotspot-conditioned binder generation against well-defined targets.** When the target has a clear paratope (an extracellular receptor domain, an enzyme active site, a viral protein interface), RFdiffusion and BindCraft consistently produce binders that pass computational triage. We've designed binders against PD-L1, TIGIT, CD8a, CD3ε, and several membrane-protein extracellular domains; the design step is no longer the bottleneck.
**De novo scaffolds with custom topologies.** AI design produces folds that don't exist in the PDB, including helical bundles tuned to specific surface complementarity. This was effectively impossible with template-based methods.
**Sequence diversity at fixed structure.** Inverse folding lets you generate dozens of sequences that fold to the same backbone. This is useful for both campaign-level diversity and downstream developability triage.
## What doesn't (yet) work
**Flat targets without obvious hotspots.** When the target's surface lacks a clear interaction patch (transcription factors, intrinsically disordered regions, non-classical epitopes), AI design hits the same wall structural biologists hit decades ago. The diffusion model has to be told where to bind; if you don't know, the model doesn't either.
**Membrane proteins beyond the extracellular domain.** Diffusion models trained on soluble protein structures struggle when the binder needs to engage a transmembrane region or a lipid-embedded epitope. Some progress with conditioning on membrane-context priors, but production-grade results require careful target preparation that still takes structural-biology judgment.
**Wet-lab hit rates without aggressive filtering.** The headline numbers ("90% of designs bind!") in academic papers reflect carefully curated test sets. Real-world hit rates from raw RFdiffusion output, before developability filtering, fall well below the published headlines. Developability filtering ([RFdiffusion outputs need a developability check](/rfdiffusion-developability-check-before-wet-lab)) closes much of the gap, but the model doesn't filter for what wet labs care about.
## The validation gap
This is where AI-driven protein design becomes interesting commercially. Software-only AI-design companies generate designs and ship them to customers; the customer (or a CRO) then runs the experimental validation. The hit rate is the joint product of the design quality and the validation strategy. When the design pipeline and the screen are not co-designed, the joint hit rate is lower than either could achieve alone.
A concrete example: a diffusion model that produces structurally plausible but aggregation-prone binders looks great on AlphaFold2 confidence scores and pipeline pLDDT. The same designs fail expression in yeast or mammalian display. If the design loop doesn't get the wet-lab feedback, it doesn't learn — and the next campaign repeats the same failure mode.
Closed-loop campaigns — where design proposes, screen rejects, and design re-proposes against the screen's feedback — produce meaningfully higher hit rates than open-loop campaigns at comparable cost. This is the operational case for integrating AI design with experimental validation rather than treating them as separate vendor relationships.
## Practical recommendations for buyers
If you're scoping an AI-driven protein design campaign:
1. **Ask the vendor for their wet-lab hit rate** on a comparable target. Not the published rate from the original RFdiffusion paper. Their rate, on their last campaign.
2. **Ask whether developability filters are applied before delivery.** A list of 1,000 designs without filters is much less useful than 100 filtered designs.
3. **Decide who owns the validation.** If the AI vendor doesn't run wet-lab validation, you'll need to coordinate with a separate display/screening provider. The data-flow latency between vendors is the silent killer of these campaigns.
4. **Budget for at least two design rounds.** Single-shot AI design rarely produces leads good enough to skip iteration.
## The integrated approach
Ranomics designs binders with RFdiffusion, BindCraft, and Boltzgen, validates them with yeast and mammalian display, and feeds the experimental results back into the design model on the same campaign. The closed loop runs in 6-8 weeks for the [AI Binder Sprint](/ai-binder-sprint) and longer for [Custom Campaigns](/biotechnology-services). The hit rates we publish are end-to-end — design through validated binder, no curation.
That doesn't make AI-driven protein design easy. It makes the difficulty visible, which is the part that matters for project planning.
---
## ai-protein-design-display-screening-integrated-workflow.mdx
Source: https://ranomics.com/ai-protein-design-display-screening-integrated-workflow/
import InOurWork from "../../components/InOurWork.astro";
Most protein engineering programs draw a line between computation and experiment. One team runs RFdiffusion or BindCraft, hands off a set of designs, and waits. Another team clones, transforms, and screens. The two groups may not speak again until results come back weeks later.
This is how most campaigns are structured. It is also why many of them underperform.
The programs that consistently produce high-affinity, developable binders do not treat design and screening as sequential steps. They treat them as a single coupled system, where the output of each stage directly shapes the input of the next. The distinction matters because the failure modes at the interface between computation and experiment are different from the failure modes within either discipline alone.
## What Computational Design Actually Produces
Tools like RFdiffusion, BindCraft, and Boltzgen generate protein backbones and sequences optimized against structural and energetic objectives. RFdiffusion produces backbone geometries conditioned on a target interface. BindCraft refines these into full sequence designs with binding energy optimization. Boltzgen samples conformational diversity across the design landscape.
What these tools do not produce is experimental validation. A design with a predicted Rosetta interface energy (dG_separated) of -15 REU and a high pLDDT score is a hypothesis. It is a well-informed hypothesis, but it remains untested until it is expressed, displayed on a cell surface, and sorted against its target.
The practical implication: computational metrics like Rosetta binding energy, AlphaFold confidence, and shape complementarity are necessary filters, but they are not sufficient predictors of experimental success. We routinely see designs that score well computationally but fail to express, fail to display, or fail to bind in a yeast display assay. The reverse also happens: designs that rank modestly in silico perform well experimentally.
This is not a failure of the computational tools. It is a reflection of the gap between the objectives these tools optimize for and the full set of biophysical properties that determine experimental success.
## Three Ways to Lose Value Between Design and Screening
**Designs that do not express.** Computational tools optimize for structure and binding, not for translational efficiency, folding kinetics in the yeast secretory pathway, or compatibility with the Aga2p fusion context. A design can have a perfect predicted structure and still misfold when expressed as a surface display construct.
**Libraries that are too narrow.** Teams often select the top 50 or 100 designs by computational score and screen only those. This approach treats the computational pipeline as a precision tool when it is better understood as a diversity generator. The ranking function is noisy. Screening a narrow, top-ranked set discards designs that would have been hits.
**No feedback path.** The most common failure is structural: the team screens one batch of designs, identifies hits, and moves to affinity maturation without feeding experimental outcomes back into the design pipeline. The computational model never learns which of its predictions were right.
The campaigns we have seen go straight from design to recombinant expression and SPR or BLI miss the non-specific binding problem entirely. A design can be a strong on-target binder and also a sticky multi-target binder at the same time; SPR will report the on-target affinity and stay silent on the rest. Display-based screening lets us look at off-target binding in the same experiment as on-target binding, against a counter-screen population, and we treat that signal as a hard filter before any candidate advances to recombinant production.
## Designing Libraries for Screening, Not for Rankings
Generate more designs than you plan to screen and filter aggressively on expressibility proxies, not just binding metrics. Predicted solubility, aggregation propensity, and the presence of unpaired cysteines or N-linked glycosylation sites in the display context are all worth filtering on before cloning.
Maintain structural diversity in the screened set. If RFdiffusion produces designs with three distinct backbone topologies for the same epitope, carry all three into screening even if one topology scores lower on average. The computational energy function is not accurate enough to reliably distinguish between topologies. The screening assay is. In our pipeline, this is also where the multi-epitope rule pays off: when the design round samples three or four pre-selected epitopes rather than one, the screening pool gets both topology and epitope diversity in a single sort.
Include internal controls. Designs with known binding properties, non-binding scaffolds, and expression-positive but target-negative controls allow you to calibrate the assay and distinguish real signal from display level artifacts. This is standard practice in library screening but is often skipped when screening computationally designed sets because the batch sizes are smaller.
## The Feedback Loop That Changes Everything
The most valuable data in any design campaign is the first round of experimental results. Not because the first hits are the final product, but because the experimental outcomes recalibrate the entire computational pipeline.
Consider what a single round of yeast display screening tells you. FACS sorting against the target separates binders from non-binders. Display level measurements (anti-tag staining) separate designs that express and fold from those that do not. NGS on the sorted populations quantifies enrichment ratios for every design in the library.
This is a rich, multivariate dataset. It tells you which backbone topologies actually produce binders. Which sequence features correlate with expression. Which predicted metrics were informative and which were noise. This is exactly the information the computational tools need to improve.
Teams that feed this data back into the design pipeline before running a second round of design see measurably better outcomes. The second batch of designs is not just iterating on hits from round one. It is generated by a pipeline that has been recalibrated against experimental reality.
At Ranomics, this is the default workflow. Computational design and yeast display screening run as alternating cycles, not as a linear handoff. The experimental data from each round directly informs the design parameters, filtering criteria, and diversity strategy of the next.
## What Integrated Campaigns Look Like in Practice
A typical coupled workflow for a binder design campaign:
---
**Identify epitopes on your own target.** [Epitope Scout](https://scout.ranomics.com) scores and ranks surface patches on any PDB structure. Free to use.
---
**Round 1: Broad exploration.** Generate 5,000 to 10,000 designs across multiple backbone topologies, targeting multiple hotspots on the antigen surface using RFdiffusion, BindCraft, and Boltzgen in parallel. Each tool explores different regions of design space: RFdiffusion for backbone diversity, BindCraft for direct binding energy optimization, Boltzgen for conformational sampling. Filter to 2,000 to 4,000 on expression proxies and structural diversity. Clone as a pooled library. Screen by yeast display with FACS. Sequence enriched and depleted populations by NGS.
**Analysis: Recalibrate.** Identify which computational metrics predicted experimental outcomes. Retrain or re-weight the filtering criteria. Identify the top-performing backbone topologies. Flag sequence features associated with expression failure.
**Round 2: Focused refinement.** Generate new designs using the validated topologies. Apply the recalibrated filters. Include point variants of round 1 hits for affinity maturation. Screen again with tighter sorting gates.
**Round 3 (if needed): Optimization.** Narrow to the best 10 to 20 candidates. Characterize individually: SPR kinetics, thermal stability, cross-reactivity. Feed developability data back if moving to mammalian display for final filtering.
The total timeline is often comparable to a traditional directed evolution campaign, but the quality of the final candidates is higher because every round is informed by both computational prediction and experimental measurement.
## The Bottleneck Has Moved
Two years ago, the bottleneck in protein design was generating plausible structures. Tools like RFdiffusion and BindCraft have largely solved that problem. The computational side can now produce thousands of diverse, structurally plausible designs in hours.
The bottleneck is now at the interface: how efficiently you convert computational diversity into experimental data, and how effectively you feed experimental results back into the design pipeline. Teams that treat this interface as an engineering problem, not an administrative handoff, are the ones producing the best molecules.
The tools exist on both sides. The gap is in the workflow that connects them.
[Start a project with Ranomics](/ranomics-contact)
{/* ranomics:related-services */}
## Related Ranomics services
- **[Compute-to-clone](/compute-to-clone):** AI design coupled to yeast/mammalian display validation in a single program.
- **[AI Binder Sprint](/ai-binder-sprint):** Integrated design + display campaigns in 6–8 weeks.
---
## From Computational Protein Design to Validated Binders: What Actually Works
> What separates successful AI protein design campaigns from failed ones? A practical breakdown of the computational and experimental steps required to go from generative models to validated binders.
Source: https://ranomics.com/ai-protein-design/
Published: 2025-06-15
Generative models for protein design have matured rapidly. RFdiffusion, BindCraft, and Boltzgen can produce thousands of candidate backbones and sequences in hours. But generating designs is not the hard part. The hard part is knowing which ones will actually fold, bind, and express.
Most computational protein design campaigns fail not because the generative model was wrong, but because the pipeline between computation and experiment was incomplete. This post covers what that pipeline looks like when it works, and where it breaks down.
## Generative Design Is Necessary but Not Sufficient
RFdiffusion generates protein backbones through a denoising diffusion process conditioned on a target structure. For binder design, you define a target surface (hotspot residues on your antigen), and the model generates backbone geometries predicted to make contacts at those positions. BindCraft takes a different approach: it optimizes sequences directly against an AlphaFold2 confidence objective, producing binders that are jointly optimized for structure and binding.
Both methods work. Published benchmarks show experimental hit rates in the range of 1% to 15% for naive RFdiffusion campaigns and higher for BindCraft when the target is well suited (rigid, structured epitopes). At Ranomics, we see BindCraft hit rates around 46% after computational filtering, though that number reflects filtered candidates entering experimental validation, not the full generated library.
The failure mode most teams encounter is not that these tools produce bad designs. It is that they produce designs that look good in silico and fail in the lab. The gap between a predicted binding interface and a functional protein is where most campaigns stall.
## The Sequence Design and Validation Stack
Once you have a backbone (from RFdiffusion or Boltzgen), you need a sequence. ProteinMPNN is the standard tool here. It takes a fixed backbone and generates amino acid sequences predicted to fold into that structure. The key parameter is temperature: lower temperatures produce more conservative, higher confidence sequences. Higher temperatures explore more sequence space but increase the risk of misfolding.
After sequence design, you need structure validation. This is where ESMFold, ColabFold, and Boltz-2 come in. The question you are answering is simple: does the predicted structure of my designed sequence match the intended backbone? If your ProteinMPNN output, when folded by an independent predictor, reproduces the target backbone with an RMSD below 1.5 to 2 Angstroms, you have a candidate worth testing. If it does not, the sequence is not encoding the structure you designed.
This "self-consistency" filter is the single most important computational quality gate. Campaigns that skip it waste weeks of experimental time on designs that were never going to fold correctly.
## Why Computational Filters Are Not Enough
Even with self-consistency filtering, computed metrics do not capture everything that matters experimentally. A design can pass every in silico filter and still fail to express in your host organism, aggregate in solution, or bind an off-target surface.
Three failure modes dominate:
**Expression failure.** The protein does not express or is insoluble. Generative models do not optimize for expression. Codon usage, glycosylation sites, and hydrophobic surface patches all affect expression and are largely invisible to structure prediction tools.
**Aggregation.** The designed protein folds but forms oligomers or aggregates. This is especially common with de novo scaffolds that have exposed hydrophobic surfaces. Solubility predictors catch some of these, but not all.
**Off-target binding.** The binder contacts the target at the intended interface but also binds other surfaces. Specificity is harder to design than affinity, and most generative models do not explicitly optimize for it.
These failure modes are why experimental validation is not optional. It is the rate-limiting step in every serious design campaign.
## Experimental Validation: Yeast Display and Beyond
Yeast display is the workhorse platform for validating computationally designed binders. You express your designed library on the yeast cell surface, incubate with fluorescently labeled target, and sort by binding signal using FACS. In a single experiment, you can screen 10,000+ designs and recover the ones that bind.
The advantage of yeast display is throughput and quantitation. You get binding signal, expression level, and relative affinity data in one experiment. Coupled with deep sequencing, you can map the full fitness landscape of your designed library.
Mammalian display offers a complementary path when your protein requires mammalian post-translational modifications or when you need to validate in a more physiologically relevant context. It is slower and lower throughput than yeast, but for certain targets (particularly glycoproteins), it eliminates an entire class of false negatives.
Deep mutational scanning (DMS) comes after initial hit identification. Once you have binders that work, DMS tells you which positions tolerate substitution and which are critical. This data feeds directly back into the next round of computational design, creating a closed loop between experiment and model.
## What Separates Successful Campaigns from Failed Ones
After running dozens of binder design campaigns, the pattern is clear. Success correlates with three factors:
**Target characterization.** Campaigns succeed when the target structure is high quality and the binding site is well defined. Designing against a homology model with a disordered loop at the epitope is a recipe for failure. If your target structure is not confident, fix that first.
**Aggressive computational filtering.** The ratio of designs generated to designs tested experimentally should be at least 100:1, often 1000:1. Self-consistency, predicted binding energy, solubility scores, and interface metrics all contribute. No single filter is sufficient. The combination is what matters.
**Fast experimental iteration.** The first round of designs rarely produces a final lead. Successful campaigns plan for two to three rounds of design, test, and redesign. Each round uses experimental data to refine the computational model. Teams that treat computation as a one-shot prediction and experimentation as a one-shot validation consistently underperform.
## The Takeaway
Computational protein design tools are powerful, but they are tools, not solutions. The difference between a successful binder campaign and a failed one is almost never the choice of generative model. It is the quality of the pipeline connecting computation to experiment: the filtering, the validation, the iteration speed, and the willingness to let experimental data override computational predictions.
The teams that win are the ones that treat this as an engineering problem, not a prediction problem.
[Start a project with Ranomics](/ranomics-contact)
{/* ranomics:related-services */}
## Related Ranomics services
- **[Compute-to-clone](/compute-to-clone):** End-to-end AI design → validated clone pipeline.
- **[AI Binder Sprint](/ai-binder-sprint):** Multi-algorithm binder design with 100% binder guarantee.
---
## Beyond Antibodies: Using Surface Display to Engineer Enzymes and Receptors
> While surface display is the go-to platform for antibody discovery, applications extend far beyond. This guide covers strategies for engineering enzymes and receptors using yeast and mammalian display.
Source: https://ranomics.com/beyond-antibodies-using-surface-display-to-engineer-enzymes-and-receptors/
Published: 2025-09-10
When protein engineers think of surface display, they almost invariably think of antibody engineering. But to limit these powerful platforms to antibodies is to overlook the fundamental theory that any protein of interest can be displayed on cell surfaces. This includes enzymes and receptors.
## Engineering Enzymes: Screening for Catalysis, Not Just Binding
The fundamental challenge with enzyme engineering via surface display is measuring a dynamic catalytic reaction within the static snapshot of a FACS experiment. The solution is to convert enzymatic activity into a stable fluorescent signal on the cell surface.
## Strategies for Capturing the Signal at the Cell Surface
**1. Leveraging Intrinsic Product Properties**
If the enzyme cleaves a substrate to reveal a charged or hydrophobic moiety, the product associates with the cell surface. This approach works well with fluorogenic substrates.
**2. Biotin-Streptavidin Capture**
The substrate is synthesized with a biotin tag. The product stays localized on the surface, and fluorescently-labeled streptavidin "paints" the active cells for sorting.
**3. Substrate Tethering**
The substrate is chemically cross-linked to the cell surface. The enzyme acts on the substrate attached to its own cell, enabling covalent retention of the fluorescent product.
**4. Droplet Encapsulation**
Cells are encapsulated in picoliter water-in-oil droplets using microfluidics. Each droplet functions as a microscopic test tube. The fluorescent product accumulates within the droplet, and droplets are sorted by FACS.
## Engineering Receptors: Screening for Function Beyond Affinity
**Affinity and Specificity**
Receptor affinity screening is analogous to antibody screening. Counter-screening with dual fluorophores (red target + green off-target) allows collection of red-positive/green-negative cells for specificity selection.
**Enhanced Stability and Expression**
Sorting for the brightest cells using an anti-expression-tag antibody enriches for improved folding and expression. This can be combined with heat or pH stability challenges to select for robustness.
**Downstream Signaling**
For receptors like GPCRs, the signaling cascade can be converted into a fluorescent signal using reporter genes (e.g., GFP downstream of the receptor's signaling pathway).
{/* ranomics:related-services */}
## Related Ranomics services
- **[Enzyme engineering](/applications/enzyme-engineering):** Directed evolution and display-based selection for enzyme activity and stability.
- **[Receptor targeting](/applications/receptor-targeting):** Engineering receptors and receptor-directed binders on display platforms.
---
## MACS for Library Pre-enrichment: When FACS Becomes the Bottleneck
> Magnetic-activated cell sorting (MACS) pre-enriches naive yeast and mammalian display libraries with billions of variants — the throughput-first complement to FACS.
Source: https://ranomics.com/beyond-facs-an-introduction-to-magnetic-activated-cell-sorting-macs-for-library-pre-enrichment/
Published: 2025-09-16
Fluorescence-Activated Cell Sorting (FACS) is the undisputed gold standard for precision sorting in surface display campaigns. Its ability to perform quantitative, multi-parameter analysis on single cells is unmatched. However, when faced with a naive library of billions of variants, FACS can become a significant bottleneck.
## How MACS Works: From Fluorescence to Ferromagnetism
1. **Labeling:** The library is incubated with biotinylated target antigen.
2. **Tagging:** Streptavidin-conjugated magnetic microbeads (50-100 nm) bind to the biotinylated antigen on positive cells.
3. **Separation:** The library is passed through a column within a permanent magnet. Tagged cells are retained; non-binders flow through.
4. **Enrichment:** The magnet is removed to elute the enriched population.
The entire process is completed in under an hour, requires minimal equipment, and is gentle on cells.
## The Strategic Role of MACS
A standard high-speed cell sorter can process tens of millions of cells per hour. In contrast, MACS can process tens of billions of cells in the same timeframe.
Attempting to screen a 10^10 variant library with FACS alone would be impractical.
## MACS vs. FACS Comparison
- **Throughput:** MACS >10^9 cells/hr vs FACS ~10^7-10^8 cells/hr
- **Precision:** FACS offers unparalleled quantitative multi-parameter gating; MACS cannot
- **Purity:** FACS yields much higher purity; the MACS population still contains low-affinity and non-specific binders
- **Cell Stress:** MACS is very gentle with no high pressure, shear forces, or laser interrogation
## The Hybrid Workflow
1. **Round 1 (MACS):** Naive library (10^9-10^10 cells) -> eliminates >99.9% non-binders -> enriched pool of ~10^6-10^7 cells
2. **Rounds 2+ (FACS):** Precision sorting for affinity maturation, antigen titration, and expression normalization
By integrating MACS as a pre-enrichment step, researchers can overcome the throughput limitations of FACS and efficiently tackle larger and more diverse libraries.
{/* ranomics:related-services */}
## Related Ranomics services
- **[Yeast surface display services](/yeast-display):** MACS pre-enrichment + FACS rounds for billion-variant libraries.
- **[NGS analysis](/technology/ngs-analysis):** Enrichment tracking across MACS/FACS rounds for clean hit calling.
---
## bindcraft-vs-rfdiffusion.mdx
Source: https://ranomics.com/bindcraft-vs-rfdiffusion/
import TryToolCallout from "../../components/TryToolCallout.astro";
import InOurWork from "../../components/InOurWork.astro";
import KeyTakeaway from "../../components/KeyTakeaway.astro";
Both BindCraft and RFdiffusion produce de novo protein binders. Both are open source, both generate structures conditioned on a target, and both can deliver binders that work in wet lab. They are not interchangeable. In our hands, they fail in different ways, succeed on different targets, and chain in different orders. This is where each one wins.
RFdiffusion is the better default for unfamiliar targets and flat surfaces, because it samples binding geometry broadly. BindCraft is the better default for well-characterized targets with one obvious hotspot, because it filters weak designs during generation. For production campaigns, run both: the compute is cheap relative to the wet-lab time it saves.
## The two methods in one paragraph each
**RFdiffusion** is a denoising diffusion model trained on protein structures. It generates protein backbones by progressively denoising random coordinates conditioned on a target structure and hotspot residues. The output is a stream of 3D coordinates — sequence is added afterward by ProteinMPNN. RFdiffusion produces high backbone diversity at scale; we routinely run 10,000-50,000 backbones per campaign.
**BindCraft** combines hallucination-based scaffold generation with integrated sequence scoring during generation. Each design step proposes a candidate, scores it against multiple metrics (interaction quality, foldability, secondary structure), and accepts or rejects. The output is a backbone-plus-sequence pair, not just a backbone. BindCraft produces fewer designs per unit compute than RFdiffusion, but each one has already passed several internal filters.
## The architectural difference matters
RFdiffusion samples broadly. BindCraft samples narrowly but with internal vetoes. This is the core trade-off.
If your target has many viable binding modes and you want to discover them, RFdiffusion's diversity is the advantage. If your target is constrained to one or two reasonable interaction geometries, BindCraft's filtering during generation cuts the candidate-evaluation overhead at the cost of design diversity.
We see this most clearly on flat targets versus targets with deep pockets. On a target with a deep pocket and one obvious binding hotspot, BindCraft converges quickly on candidates that fit; RFdiffusion produces many backbones that approach the pocket from sub-optimal angles. On a flat target with multiple plausible interaction surfaces, RFdiffusion's broader sampling produces more candidates that turn out to bind in unexpected (and useful) modes after wet-lab validation.
The gain we get from each generator is much higher when we run several parameter strategies within the model rather than a single default sweep. For RFdiffusion that means varying sampling temperature and trying multiple hotspot configurations. For BindCraft that means alternating tight-scaffold, balanced, and quality-focused filter profiles. A single parameter strategy biases the output toward whichever fold or CDR archetype the model prefers under those conditions, and the union of several strategies covers more of the design space the model can actually access. The cost is a few extra runs; the upside is screening pools that are not all variations of the same archetype.
## Computational cost (production scale)
| Metric | RFdiffusion | BindCraft |
|---|---|---|
| Backbones per H100-hour | ~250-400 | ~50-80 |
| Sequences per backbone | 5-8 (via ProteinMPNN) | 1 (integrated) |
| Wall-clock for 10K candidates | ~25-40 hours on 1xH100 | ~125-200 hours on 1xH100 |
| Memory footprint | ~22 GB VRAM | ~16 GB VRAM |
| Failure rate (NaN, OOM, divergence) | Low | Moderate |
These numbers come from our own runs on PD-L1, TIGIT, and CD8a target structures. Performance varies with target size, contig length, and hotspot count.
## When RFdiffusion wins
**Targets with a large hotspot region or multiple interaction sites.** If you can specify five or more hotspot residues spanning a non-trivial surface area, RFdiffusion's diversity gives you binders that approach from multiple angles. Some of those will be the binders you wanted; others reveal binding modes you hadn't considered.
**Campaigns where you'll filter aggressively post-hoc.** RFdiffusion produces many candidates fast. If your downstream pipeline includes [developability filters](/rfdiffusion-developability-check-before-wet-lab), AlphaFold2 confidence scoring, and quick experimental triage, the volume becomes an asset rather than a liability.
**Scaffold-grafting and partial-diffusion workflows.** RFdiffusion supports partial diffusion (denoising only a region of an existing scaffold), which is the right tool for grafting a designed binding loop onto a stable framework. BindCraft has less mature support for this workflow.
**Membrane proteins where binding mode is uncertain.** On the extracellular domain of a GPCR, for instance, you may not know which face of the protein is the right interaction surface. RFdiffusion's diversity surfaces multiple plausible options; BindCraft's filtering may converge on whichever face the scoring function happens to prefer.
## When BindCraft wins
**Small targets with one obvious interaction hotspot.** If the target has a single binding site (a peptide, a small extracellular domain, a defined epitope), BindCraft converges faster and the integrated scoring filters out junk during generation. You spend less compute and less wet-lab budget on candidates that were never going to work.
**When you want one good binder per ~100 candidates rather than ten good binders per 10,000.** For a Pilot-scale campaign with limited screening budget, BindCraft's higher quality-per-design is worth the lower throughput.
**Compact backbone topologies.** BindCraft's scoring penalizes designs that don't fold cleanly to a small, stable topology. If your application needs a binder you can express in E. coli or yeast at high yield, the bias toward compact, well-folded outputs is a feature.
**Iterative refinement against a single design intent.** When you've decided what kind of binder you want (a 60-residue helical bundle hitting a specific patch) and you want many sequence variants, BindCraft's sample-and-filter loop is more efficient than running RFdiffusion plus separate ProteinMPNN runs.
## How we chain them
Most of our [AI Binder Sprint](/ai-binder-sprint) campaigns use both methods, not one. A common pattern:
1. **RFdiffusion sweep**, 10,000-30,000 backbones, broad sampling. Goal: find the right interaction modes against the target.
2. **Cluster the RFdiffusion outputs** by binding-mode geometry (typically 3-5 distinct modes survive AlphaFold2 confidence triage).
3. **For the most promising mode, run BindCraft** with hotspot conditioning that locks the geometry RFdiffusion discovered. Goal: generate higher-quality variants of the discovered solution.
4. **Wet-lab screen the union** — RFdiffusion winners plus BindCraft winners. Hit rates from BindCraft variants typically run 1.5-2x higher per design, but the absolute number of viable hits is comparable because RFdiffusion produces many more candidates.
This chain costs more compute than running either tool alone, but the hit rate per dollar of wet-lab validation is meaningfully higher because the candidates that reach screening are pre-filtered.
## The cases where neither wins
Both methods struggle on:
- **Very small targets** (peptides under 15-20 residues). The diffusion conditioning loses traction with little structural context.
- **Intrinsically disordered targets**. Neither model has a stable structure to anchor binding-mode prediction.
- **Targets that require allosteric binders rather than active-site binders**. The methods optimize for direct contact, not allosteric effect.
For these cases, the right tool isn't either one — it's directed evolution from a focused library, or DMS-guided rational design. AI binder design isn't the universal answer.
## Decision summary
If you can choose only one tool: RFdiffusion is the better default for unfamiliar targets; BindCraft is the better default for well-characterized targets. For production campaigns, run both, and run each across several parameter strategies. In our work, the compute is cheap relative to wet-lab time, and the marginal hit rate from sampling multiple parameter profiles per tool is consistently worth the extra GPU-hours.
If you're scoping a campaign and want a second opinion on which tool fits your target, see our [AI protein binder design services](/ai-protein-binder-design), [start a Binder Pilot](/binder-pilot), or reach out via the [contact page](/ranomics-contact).
---
## Binder Design on a Grant Budget: Scoping a Single-Target Campaign
> What to prioritize, what to cut, and what actually determines cost when a PI or postdoc is running a single-target de novo binder design campaign on a defined budget.
Source: https://ranomics.com/binder-design-on-a-grant-budget/
Published: 2026-04-20
import InOurWork from "../../components/InOurWork.astro";
Most published benchmarks for AI protein binder design assume a well-resourced industrial pipeline: GPU clusters, in-house yeast display, FACS, NGS, and enough budget to run several hundred to several thousand designs through the full experimental funnel. Most academic labs running their first de novo binder campaign do not have that setup. This post is a practical scoping guide for running a single-target binder campaign on a defined grant budget — what to prioritize, what to cut, and what actually determines cost.
## The Cost Drivers in a Binder Design Campaign
Before deciding what to cut, it helps to understand where the money actually goes in a binder campaign. In rough order of contribution:
1. **Gene synthesis of the filtered candidate pool.** This scales linearly with the number of designs you carry into the wet lab. A 1,500-design pool is a different cost category from a 300-design pool.
2. **FACS or MACS selection rounds.** Each sorting round has instrument time, reagents, and bench-scientist overhead. Multi-round campaigns compound.
3. **NGS sequencing of sorted populations.** One MiSeq or NextSeq run per sorted population is the standard readout.
4. **Display platform prep.** Yeast display construction and transformation is a relatively fixed cost per library.
5. **GPU compute for generative design.** RFdiffusion and BindCraft campaigns are real GPU hours, but they are usually a smaller line item than gene synthesis for a moderate-sized pool.
6. **Bioinformatics and reporting.** Hit calling, enrichment analysis, and writeup — bounded by design pool size.
Computational design is often the smallest line item. The big cost drivers are the synthesis and experimental follow-through.
## What to Cut: Scope, Not Rigor
The single most effective way to bring a binder campaign inside a grant budget is to cut pool size, not methodology rigor. A single-round campaign on a 200-500 design pool with a clean NGS readout is defensible, publishable, and answers a specific scientific question. A 2,000-design multi-round campaign is a production pipeline, not a focused experiment.
Specifically, at grant scale you can typically run:
- **One algorithm, not three.** Pick RFdiffusion plus ProteinMPNN sequence design, or pick BindCraft. Running RFdiffusion, BindCraft, and Boltzgen in parallel is a flagship-scale choice.
- **One selection round, not two or three.** A single FACS round on a well-designed library gives you enough enrichment signal to rank candidates. Multi-round affinity maturation is a separate experimental question and a separate campaign.
- **Yeast display, not mammalian display.** Yeast display is cheaper, faster, and well-validated for extracellular target binding. Mammalian display is worth the cost only when post-translational modifications on the displayed protein are the scientific question.
- **NGS hit calling, not characterization.** SPR, BLI, and ITC affinity measurements are expensive. A ranked NGS hit list plus one or two purified-protein validations of top hits is usually enough for a first paper.
This is how the [Binder Pilot](/binder-pilot) program is scoped. It is not a diminished version of the flagship; it is a deliberately focused campaign for teams where the output is one solid ranked hit list plus the NGS and sequence data to back it up.
## What Not to Cut: Computational Quality Gates
The computational side of a binder campaign is the cheapest part. Do not cut it.
Specifically, keep:
- **Self-consistency filtering.** Before you synthesize any designs, fold each candidate with an independent predictor (ESMFold, ColabFold, or Boltz-2) and compare to the intended backbone. Designs with RMSD > 2 Å from the intended fold go in the bin. This single filter typically removes 50-80% of raw generative output and drastically improves wet-lab hit rates.
- **Hotspot selection.** Where you direct the design matters more than how many designs you generate. Spend an hour on [Epitope Scout](/technology/epitope-scout) or manual PyMOL inspection identifying surface patches that are hydrophobic, rigid, accessible, and ideally contain known hotspot residues (Trp, Tyr, Arg, Phe). Bad hotspot selection makes a 2,000-design pool perform worse than a 300-design pool on the right epitope.
- **Structure quality triage on the target.** If your starting structure is an AlphaFold model, check pLDDT locally around the intended binding region. Disordered or low-confidence regions are a warning sign. Crystal structures are preferred; high-confidence models are usable; low-confidence models are a coin flip.
## Choosing Your Target
The most common failure mode for academic binder campaigns is not computational — it is target choice. Some targets are hard even for well-resourced pipelines. At grant scale, pick a target where the biology is on your side:
- **Soluble, extracellular, folded.** Yeast display works best for targets that can be surface-exposed and bound by designed proteins in a largely native conformation.
- **Ideally with a known binding partner structure.** Even if you are not engineering an antibody fragment, a co-crystal structure of any protein bound to your target tells you which surfaces are bindable.
- **Rigid and structurally well-defined.** Flexible loops, disordered termini, and conformational ensembles are where computational design struggles and where hit rates drop.
- **Not a PTM-dependent epitope.** If the native binding partner only recognizes a glycosylated, phosphorylated, or acetylated version of your target, a yeast-display de novo binder campaign is the wrong tool.
In our work, well-validated targets like TNF-α are the most forgiving for first campaigns: deep PDB coverage, known epitope geometry, and decades of published binder structures give you ground truth to compare your designs against. Novel-class targets are scientifically exciting, but they punish first-time campaigns when something goes wrong because you have no reference for what a "normal" hit rate or affinity range should look like.
## Timeline Expectations
A single-round, single-algorithm, grant-scale binder campaign typically runs 4-6 weeks end-to-end from kickoff to delivered hit list, depending on synthesis turnaround. That is roughly half the timeline of a flagship multi-algorithm program, and it is compatible with most grant or dissertation cycles.
The slowest steps are almost always the physical ones: gene synthesis (1-2 weeks) and NGS run scheduling. The computational side runs in days.
In our work, the slowest step in a yeast display campaign is almost always DNA synthesis, not the assay or the sort. Two to four weeks of gene synthesis wait dominates the wall-clock timeline, and the actual screening once DNA arrives runs in days. Scope grant timelines around the synthesis vendor lead time, not the assay throughput, because that is where the schedule actually slips.
## What You Should Publish
A focused, grant-scale binder campaign is not a diminished result. Published work in the RFdiffusion and BindCraft literature routinely reports single-round campaigns against single targets with 2-12 confirmed binders per 1,000 screened candidates. A clean single-round Pilot-scale campaign hitting that range, with the NGS enrichment traces to back it up, is a publishable dataset in its own right.
The things to include in a paper or preprint:
- The generative algorithm, version, and seed parameters.
- The hotspot selection procedure and justification.
- The filtering pipeline (self-consistency thresholds, expression filters, any physics-based triage).
- The library size at synthesis and at sort.
- The NGS enrichment distribution for the top hits, not just the headline hit count.
- At least one orthogonal validation of the top-ranked hit (purified protein binding assay on a handful of candidates).
## If You Are Deciding Between Running It In-House or Outsourcing
This is a judgment call, not a formula. Run it in-house when you have a yeast display pipeline, a FACS core, and a scientist or postdoc with bandwidth. Outsource when any of those is missing and when the project timeline cannot absorb a six-month ramp to set up display from scratch.
A well-scoped outsourced Pilot at grant scale typically comes in at roughly the cost of gene synthesis plus one FACS sort plus one NGS run, plus the bioinformatics overhead. There is no line item for GPU compute that you would not already be paying in cloud or shared-cluster time if you ran it yourself. The main thing you are buying is experimental throughput and reproducible methodology, not access to the algorithms themselves — RFdiffusion, BindCraft, and ProteinMPNN are all open-source.
## Starting Points
If you have a target structure, an idea of the binding region, and a grant or core-facility budget, the fastest path to scoping a campaign is:
1. Run [Epitope Scout](/technology/epitope-scout) on your target (free) to get a ranked list of candidate epitope patches.
2. Pick one patch. Send us the structure and the selection. We will assess feasibility and scope a [Binder Pilot](/binder-pilot) around it.
3. Expect a kickoff call to define deliverables, then 4-6 weeks to a ranked hit list.
For multi-target pipelines, multi-round optimization, or teams that need the 100% binder guarantee that the flagship [AI Binder Sprint](/ai-binder-sprint) includes, the Pilot is not the right scope. For a first single-target campaign on a grant, it usually is.
{/* ranomics:related-services */}
## Related Ranomics services
- **[Binder Pilot](/binder-pilot):** Grant-scale starter program — single-round, smaller design pool.
- **[Epitope Scout](/technology/epitope-scout):** Free self-serve epitope identification to de-risk the target before you spend.
---
## Titrating Display Levels for Reliable Affinity Data
> Learn how to optimize display-level titration in yeast and mammalian display systems to obtain accurate affinity data and avoid avidity effects when determining Kd.
Source: https://ranomics.com/correctly-titrating-display-levels-for-reliable-affinity-data-in-yeast-and-mammalian-systems/
Published: 2025-08-28
## The Affinity Data Dilemma
You've spent weeks optimizing your yeast or mammalian surface display experiment. Your flow cytometry data looks clean, the resulting binding curve is a perfect sigmoid, and you calculate an impressive, low-nanomolar Kd. But when you repeat the experiment, the value shifts significantly.
## The Core Principle: Why Display Level is Critical
**Affinity (KD):** The intrinsic binding strength between a single binding site and a single epitope. A monovalent interaction.
**Avidity:** The overall accumulated strength of multiple simultaneous interactions. Think of it like Velcro: a single hook-and-loop pair is weak (affinity), but the combined strength of thousands of pairs is immense (avidity).
## The Goal: The Stoichiometric Binding Regime
The experimental sweet spot where the surface density of your displayed protein is low enough that, on average, each labeled antigen molecule can only interact with a single displayed protein at a time.
## Step-by-Step Guide: Titration in Yeast Display
1. **Control Your Expression Level.** Time course with GAL1 promoter: 2, 4, 8, 16 hours at 20C
2. **Set Up the Labeling Experiment.** Fixed antigen concentration near expected KD, co-label with anti-c-myc PE + antigen-AF647
3. **Flow Cytometry & Gating.** 2D plot: X=Expression (anti-c-myc PE), Y=Binding (Antigen-AF647)
4. **Data Analysis.** Plot Antigen MFI vs Expression MFI. **Linear = good (1:1 regime). Curve/plateau = avidity.**
## Step-by-Step Guide: Titration in Mammalian Display
Leverage expression heterogeneity within a single transiently transfected sample.
1. Include a viability dye (DAPI, PI, or Zombie Dyes). This is critical.
2. Gate: Live Cells -> Singlets -> Analysis
3. Single-Sample Power Plot: antigen binding vs expression level
4. Linear relationship = proof of 1:1 regime. Plateau at high expression = avidity. Gate to exclude plateaued cells.
## Common Pitfalls
1. **Using Too Much Antigen.** Use sub-saturating concentration at or below expected KD.
2. **Not Normalizing to Expression Level.** Always use the Antigen MFI vs Expression MFI plot.
3. **Ignoring Cell Viability.** Dead cells are "sticky" and create false positives. A viability dye is non-negotiable.
## Conclusion: Confidence in Your Curves
Generating an accurate KD value is more than just fitting a curve to a set of data points. It's about ensuring that the data itself reflects the true molecular interaction.
For a worked example of titration-aware sorting across a real six-cycle campaign, see our case study on [pH-dependent antibody engineering via yeast surface display](/ph-dependent-antibody-engineering) — 640 clones, convergent hotspot residues, and the enrichment-score methodology that makes ranking reliable.
{/* ranomics:related-services */}
## Related Ranomics services
- **[Yeast surface display services](/yeast-display):** FACS/MACS screening with titration-aware protocols and NGS hit calling.
- **[Cell engineering services](/cell-engineering):** Custom yeast and mammalian display libraries with clean affinity measurement.
- **[Case study: pH-dependent antibody engineering](/ph-dependent-antibody-engineering):** 14-page technical walkthrough of a real client campaign.
---
## de-novo-protein-design-failure-modes.mdx
Source: https://ranomics.com/de-novo-protein-design-failure-modes/
import InOurWork from "../../components/InOurWork.astro";
The headline numbers in AI protein design papers — "90% of designs bind!" — are correct in their context and misleading out of it. In production binder design campaigns, the wet-lab hit rate for raw RFdiffusion or BindCraft output is a small fraction of the published rate. Aggressive developability filtering closes much of the gap. Hit rates approaching the method-paper headlines are exceptional and usually mean the target was particularly cooperative. Here are the failure modes that account for the gap, and the filters that close it.
## What the published hit rates measure
When a method paper reports "90% binding," it typically means: of designs that passed multiple curation steps internal to the method, 90% bound the target in a screening assay calibrated to that method. The denominator excludes:
- Designs that didn't pass AlphaFold2 confidence filters
- Designs with low pLDDT or pTM scores
- Designs the authors discarded for obvious structural issues
- Designs that failed expression and were therefore never tested
- Designs that bound but to off-target sites the method-development screen didn't detect
The published rate is a useful method-development metric. It is not a project-planning metric. The number a customer cares about — designs ordered, fraction that produced a usable binder — is what we mean by "wet-lab hit rate" in this article.
## Failure mode 1: aggregation and expression failure
The single largest failure category in our campaigns is designs that fail to express in soluble form. They aggregate during yeast or mammalian expression, or they fail to reach the cell surface in display experiments, or they form inclusion bodies in *E. coli*-based downstream production. The structural model the AI generated may be plausible, but the protein doesn't fold to that model in cells.
Why it happens: AI design models are trained on PDB structures, which are by definition successfully expressed and crystallized proteins. The training set has selection bias. Designs that resemble the failure cases — proteins that didn't make it to the PDB — are not penalized by the model because the model never saw them.
How to filter: predicted aggregation propensity (TANGO, AGGRESCAN, or developability-aware variants), surface hydrophobicity scores, and net surface charge. Any design with a hydrophobic patch larger than ~400 Ų should be flagged. We documented the specific filters at length in [RFdiffusion outputs need a developability check](/rfdiffusion-developability-check-before-wet-lab).
## Failure mode 2: low affinity (Kd > 1 µM)
A design that "binds" the target on AlphaFold2 confidence metrics may bind with Kd in the high micromolar range. This passes binary "does it bind" assays but fails any quantitative selection. The failure isn't in the structural model; it's that the model's scoring function is correlated with affinity but not equivalent to it. A geometrically perfect interface can still have insufficient enthalpic and entropic optimization.
Why it happens: AI design optimizes for "designability" — the probability that a sequence folds to a given backbone — and "binding mode plausibility." It does not directly optimize for affinity. Some methods incorporate affinity-correlated proxies, but the proxy is imperfect.
How to filter: dock-and-rescore approaches (Rosetta interface analysis, AlphaFold2-multimer with confidence-on-docked-pose), interface buried surface area, and shape complementarity. Designs with buried surface area below ~600 Ų rarely produce sub-µM affinity binders.
In our binder campaigns, IPTM and IPSAE work as negative filters and almost nothing else. Low scores reliably rule out high-affinity binders, but high scores do not predict affinity. We treat model confidence outputs as triage to discard the worst candidates, not as a ranking on the best. The real ranking has to come from downstream display screening and biochemical readouts.
## Failure mode 3: off-target binding (polyspecificity)
Designs that bind the intended target also bind everything else: serum albumin, off-target cell-surface receptors, plate plastic. This appears as "non-specific binding" in flow cytometry and as panreactivity in PSR (poly-specificity reagent) assays. For therapeutic applications, polyspecificity is a developability dealbreaker.
Why it happens: hydrophobic patches and electrostatic asymmetries that drive on-target affinity also drive off-target binding. The design model sees the on-target benefit; it doesn't penalize the off-target cost.
How to filter: PSR scoring during candidate triage, baculovirus particle (BVP) ELISA before commitment to soluble expression, and aggressive negative-selection rounds during yeast or mammalian display screening. We've documented the developability red flags in detail at [5 Developability Red Flags](/5-developability-red-flags-that-kill-mab-programs).
## Failure mode 4: binding the wrong epitope
The design binds the target — at a site other than the one specified by the hotspot conditioning. This is a partial success: you have a binder, but it doesn't do what the customer needs (block a specific receptor interaction, neutralize a specific function). The wet-lab assay reports "yes, binds." The functional assay reports "no effect."
Why it happens: hotspot conditioning is a soft constraint, not a hard one. The model can satisfy the hotspot condition partially while finding a stronger binding mode elsewhere on the target. Diffusion models that "almost" hit the right epitope are surprisingly common output.
How to filter: AlphaFold2-multimer redocking with hotspot-residue contact monitoring, competition assays during display screening (titrate against a known epitope-blocker), and structural biology validation of leads (HDX-MS, cryo-EM if available, or competition with structurally characterized binders).
In our pipeline, we never let a campaign exit on a binary "binds or does not bind" readout. Every confirmed binder has to clear a downstream functional or competition assay before it counts as a deliverable, because the cases where binding does not translate to function are common enough that the binary gate is a confidence trap.
## Failure mode 5: post-translational modification incompatibility
The design assumes a structure that depends on disulfide bonds, glycosylation, or specific PTMs that don't form correctly in the chosen expression system. Yeast-displayed designs that depend on mammalian-style glycosylation are the most common case. Bacterial expression of designs that need disulfide formation is another.
Why it happens: AI design models don't model PTM machinery. They predict structure assuming the relevant PTMs are installed correctly. When the expression system doesn't deliver, the design doesn't fold.
How to filter: PTM site prediction during the design phase, expression-system fit-checking (don't display in yeast a design that assumes mammalian glycosylation), and the [two-platform approach](/the-two-platform-approach-using-yeast-display-for-affinity-and-mammalian-display-for-developability) for designs where PTM dependence is unclear.
## The closed-loop fix
The pattern that closes the gap between published hit rates and production hit rates is iteration with feedback. Single-shot AI design produces a candidate list with a known hit-rate distribution. The hit rate for the next campaign improves only if the design model learns from the wet-lab outcome.
In practice, "learning" doesn't mean retraining the diffusion model — that's not feasible at customer-campaign scale. It means:
1. **Filter rules learned from this campaign feed the next campaign's pre-screen filters.** If a particular hotspot configuration produced 80% aggregators in the last run, deprioritize that configuration in the next.
2. **Sequence patterns that produced binders inform ProteinMPNN sampling temperature and constraint.** Designs that worked share statistical signatures the model can be coaxed toward.
3. **Target-specific developability thresholds replace generic ones.** A specific target's binders may tolerate higher hydrophobicity than the generic threshold; a different target may need stricter filters.
The closed loop is what shifts hit rates from "random pass-through of design pipeline" to "candidates already vetted against the target's specific failure modes."
## Practical recommendations
If you're running an AI binder design campaign and want to maximize hit rate:
1. **Apply developability filters before wet-lab synthesis, not after.** Synthesizing 1,000 candidates and discarding 950 in screening is more expensive than synthesizing 100 pre-filtered candidates and screening all of them. Synthesis cost dominates wet-lab budget for AI campaigns at this scale.
2. **Plan for at least two design rounds.** Single-round designs produce leads that are good enough to validate the workflow but rarely good enough to deliver to customers. The second round, with feedback from the first, is where the campaign-quality output happens.
3. **Run yeast or mammalian display, not just ELISA.** Display screening surfaces polyspecificity and aggregation issues that ELISA assays miss. The triage value is worth the extra weeks.
4. **Budget for at least 1,000 designs synthesized per round.** Below this, the per-round hit rate noise dominates the design-quality signal.
5. **Don't trust the published hit rate of any AI-design vendor without seeing their last campaign's numbers.** Method papers are the floor; vendor claims should match or exceed those numbers on real customer projects.
In our practice, the underlying mindset matters more than any single filter on this list. We treat AI design outputs the way we treat any other protein design: the developability rules, hydrophobic-patch checks, charge balance, and structural sanity that apply to grafted CDRs and scaffold-designed proteins still apply to model output. The model is the starting point of the design process, not the finished design.
---
If you're scoping an [AI protein binder design](/ai-protein-binder-design) campaign and want help anticipating which failure modes matter for your target, [start a Binder Pilot](/binder-pilot) or reach out via the [contact page](/ranomics-contact). We design and validate de novo binders end-to-end and publish hit rates without curation.
---
## Deconvoluting Polyclonal Hits: Strategies for Characterizing Enriched Library Pools
> Your yeast display screen is finished, but choosing the most abundant clone from NGS data can lead to costly mistakes. A strategic framework for deconvoluting polyclonal hits using enrichment ratios and convergent evolution.
Source: https://ranomics.com/deconvoluting-polyclonal-hits-strategies-for-characterizing-enriched-library-pools/
Published: 2025-09-22
Your [yeast display](/yeast-display) or mammalian display screen is finished, but now you face a complex set of NGS data where simply choosing the most abundant clone can lead to costly mistakes. This guide provides a strategic framework for deconvoluting polyclonal hits by moving beyond simple frequency to analyze enrichment ratios and patterns of convergent evolution.
## The Foundation: Why Simple Abundance is Not Enough
Relying solely on the final frequency of a clone is a flawed strategy because it ignores the history of the selection process.
Reasons a clone can be highly abundant without being a top performer:
- **"Jackpot" Effect:** Over-represented in the initial library due to synthesis or cloning bias
- **PCR Amplification Bias:** Some sequences are amplified more easily than others
- **Modest Binders with High Display Levels:** A million copies of a mediocre binder can be brighter than a thousand copies of an elite binder
The key is not to ask "Which clone is most common at the end?" but rather, "Which clone showed the most significant and consistent improvement throughout the selection?"
## The Primary Metric: Calculating Enrichment Ratios
**Enrichment Ratio = (Frequency of Variant in Final Round) / (Frequency of Variant in Unselected Library)**
You **must** deep-sequence both your final enriched pool and your initial, unselected (Round 0) library. A variant that started at a frequency of 0.001% and ended at 1% (a 1000x enrichment) is often far more interesting than a variant that started at 0.5% and ended at 2% (a 4x enrichment).
## Identifying Convergent Evolution: The Power of Sequence Families
Key patterns to look for:
- **Are entire families enriching?** If a cluster of 20 related sequences all show high enrichment ratios, it provides immense confidence that this structural solution is robust and effective.
- **Are there consensus mutations?** Aligning sequences within an enriching family identifies key consensus mutations driving improved function.
- **Are there shared motifs across different families?** Sometimes, different sequence families will independently discover the same solution at a key position.
## Putting It All Together: A Candidate Selection Framework
Selection matrix criteria:
- **Enrichment Ratio:** Quantitative measure of selection success
- **Final Abundance:** Abundant enough to be real, not a sequencing artifact
- **Family Convergence:** Part of a larger enriching family (confidence score)
- **Sequence Liabilities:** Red flags like glycosylation sites, deamidation motifs, or unpaired cysteines
**Hypothetical case:**
- **Candidate A:** #1 most abundant (5% final frequency), modest 15x enrichment, orphan clone
- **Candidate B:** #30 most abundant (0.5% final frequency), massive 800x enrichment, lead member of a family of 25 enriching variants sharing a key mutation at position H52
Candidate B is a far more compelling and well-validated hit than Candidate A.
## Conclusion: From Data to Discovery
Deconvoluting a polyclonal NGS dataset is an investigative process that blends quantitative analysis with biological intuition.
For a worked example of enrichment-ratio ranking and convergent-hotspot deconvolution from a real campaign, see our case study on [pH-dependent antibody engineering via yeast surface display](/ph-dependent-antibody-engineering) — 640 clones, six FACS sorts, and the convergent residues (designated H-1 and H-2) that the analysis surfaced.
{/* ranomics:related-services */}
## Related Ranomics services
- **[NGS analysis](/technology/ngs-analysis):** Enrichment-ratio hit calling and clonal deconvolution across sort rounds.
- **[Antibody engineering](/applications/antibody-engineering):** Polyclonal-to-monoclonal workflows including validation and characterization.
- **[Case study: pH-dependent antibody engineering](/ph-dependent-antibody-engineering):** 14-page technical walkthrough of a real client campaign.
---
## Deep Mutational Scanning: Mapping Protein Fitness Landscapes
> Deep mutational scanning (DMS) combines saturation variant libraries, functional selection, and NGS to measure the functional consequences of thousands of mutations in one experiment — producing a complete map of a protein's fitness landscape.
Source: https://ranomics.com/deep-mutational-scanning-a-high-throughput-approach-to-mapping-protein-fitness-landscapes/
Published: 2025-08-21
Deep mutational scanning (DMS) is a technique that systematically measures the functional consequences of thousands of mutations across a protein in a single experiment. By combining high-diversity variant libraries, functional selection, and next-generation sequencing (NGS), [DMS](/technology/deep-mutational-scanning) produces a comprehensive map of a protein's fitness landscape — revealing which substitutions are beneficial, neutral, or detrimental with single-residue resolution.
Classic approaches rely on site-directed mutagenesis to investigate one position at a time. While informative, that approach is like mapping a continent one footstep at a time. The fitness landscape of even a modestly sized protein spans an astronomically large sequence space; DMS is the only experimental method that can characterize it at scale. The data is invaluable for antibody engineering, enzyme optimization, viral evolution surveillance, and [directed evolution](/a-technical-guide-to-directed-evolution-for-enhancing-protein-stability-and-function) campaign design.
## The DMS Workflow: A Four-Step Process
At its core, a DMS experiment is a process of selection and quantification. It systematically measures how mutations affect a protein's function by counting the frequency of each variant in a population before and after a functional challenge.
**Step 1 - Library Generation:** A comprehensive library of genetic variants of the target protein is created. This library can be designed to contain all possible single amino acid substitutions (a saturation mutagenesis library), or it can be generated more randomly using methods like error-prone PCR.
**Step 2 - Functional Selection:** The core of the experiment. The entire library of variants is subjected to a selection pressure that links their genetic code (genotype) to a specific function (phenotype). For an antibody, this might be binding to an antigen. For an enzyme, it could be its catalytic activity.
**Step 3 - Deep Sequencing:** Using NGS, the DNA from the library is sequenced before and after the functional selection. This step generates millions of reads, providing a quantitative count of each specific variant in both populations.
**Step 4 - Data Analysis & Fitness Scoring:** The read counts from the before and after-selection pools are compared. Variants that are beneficial for the selected function will be more frequent in the after-selection pool, while detrimental variants will be depleted. This ratio is used to calculate an "enrichment score" or "fitness score" for each mutation, effectively mapping the entire landscape.
## Critical Technical Considerations Researchers Often Overlook
### 1. Library Quality and Bias are Not Trivial
The foundation of any DMS experiment is the mutant library. An ideal library has even distribution of all intended variants. However, biases from oligonucleotide synthesis and cloning can lead to some variants being over- or underrepresented from the start. [Library size and diversity tradeoffs](/the-numbers-game-library-diversity-ngs) interact directly with sequencing depth requirements — both need to be scoped together before the experiment begins.
**What's often overlooked:** Failing to deeply sequence the input library to quantify this initial bias. Without this baseline, you can't distinguish between a mutation that's truly detrimental and one that was simply rare in the starting pool. Furthermore, synthesis errors can introduce truncations or frameshifts, creating a population of non-functional proteins that act as dead weight and can skew normalization.
### 2. The Nuance of Selection Pressure
The selection assay is not a simple binary filter. The outcome data is a continuous spectrum of variant activity and many variants will be in the "grey" zone. The stringency of the selection pressure is a critical variable that dictates the quality of the fitness landscape you can resolve.
**What's often overlooked:** Applying a selection pressure that is too strong or too weak. An overly stringent selection may only identify the top few elite variants, masking the subtler effects of moderately beneficial or slightly deleterious mutations. Conversely, a selection that is too weak will fail to distinguish between functional variants and non-functional ones, leading to a noisy, flat landscape. Optimizing selection conditions (for example, by titrating the concentration of a target antigen in a yeast display experiment) is a non-negotiable step for generating high-quality data.
### 3. Sequencing Depth and Error Correction
DMS relies on accurate counting. If your sequencing depth is insufficient, you won't be able to reliably quantify rare variants, which can be just as informative as common ones.
**What's often overlooked:** The impact of PCR and sequencing errors. Standard NGS protocols have error rates that can introduce false mutations into your reads. This noise can obscure the signal, making it appear as though certain mutations exist when they don't. A robust approach involves incorporating **Unique Molecular Identifiers (UMIs)**, short, random DNA sequences attached to each initial DNA molecule. By collapsing reads that share the same UMI, you can computationally correct for PCR and sequencing errors, dramatically cleaning up the final dataset.
## Applications: From Basic Science to Drug Development
When executed correctly, DMS is a transformative tool. It allows researchers to:
- **Map antibody-antigen interfaces** with single-residue resolution, identifying critical binding hotspots.
- **Predict viral evolution** by identifying mutations that allow viruses like influenza or SARS-CoV-2 to escape the immune system.
- **Guide protein engineering efforts** by providing a complete roadmap of which mutations will enhance stability, activity, or binding affinity.
- **[Engineer enzymes](/leveraging-ai-deep-mutational-scanning-engineer-enzymes)** by identifying thermostability-compatible substitutions, substrate-specificity hotspots, and tolerance profiles for industrial bioprocessing conditions.
- **Understand drug resistance** by revealing mutations in a target protein that abolish its interaction with a small molecule inhibitor.
## Conclusion
Deep Mutational Scanning has moved the field of protein engineering from a process of educated guesses to one of data-driven design. It provides an unprecedented view into the rules that govern protein function. However, the quality of this view is entirely dependent on the quality of the experiment. By paying close attention to library quality, carefully optimizing selection pressures, and implementing robust error-correction strategies, researchers can unlock the full potential of this powerful technique and build a truly predictive understanding of the protein universe.
**Reference:** Melamed D, Young DL, Miller CR, Fields S (2015) Combining Natural Sequence Variation with High Throughput Mutational Data to Reveal Protein Interaction Sites. PLoS Genet 11(2): e1004918.
{/* ranomics:related-services */}
## Related Ranomics services
- **[Deep mutational scanning](/technology/deep-mutational-scanning):** Thousands of variants characterized per experiment with NGS readout.
- **[Protein engineering services](/protein-engineering):** DMS-driven engineering for stability, activity, and expression.
---
## Deep Mutational Scanning for Antibody Affinity Maturation
> How deep mutational scanning maps an antibody's affinity landscape and reveals which residues to optimize without breaking developability.
Source: https://ranomics.com/dms-for-antibody-affinity-maturation/
Published: 2026-05-11
Affinity maturation moves an antibody from a discovery hit (typically 10–100 nM) to a clinical lead (typically sub-nanomolar). The historical methods — error-prone PCR, CDR walking, parsimonious mutagenesis — are productive but inefficient. They search sequence space by random sampling and selection, which means most cycles spend most of their budget on neutral or deleterious mutations. The yield is measured in rounds, not in mutations per round.
Deep mutational scanning replaces the random walk with a systematic map. Every single-amino-acid substitution across the CDRs is measured for its effect on binding, expressed, displayed, or any other phenotype that can be linked to a sequencing readout. The output is a fitness landscape that tells you, position by position, which substitutions improve affinity, which are neutral, and which break the antibody. The next-round library is built from validated information, not random sampling.
This article is a practical guide to running DMS for affinity maturation: how to design the library, how to read it out, how to use the landscape, and how to respect the developability constraint that random affinity maturation tends to ignore.
## The affinity-maturation problem
A discovery campaign delivers a hit. Maybe it came from yeast display, maybe from a transgenic mouse, maybe from AI design followed by experimental validation. The hit binds in the high-nanomolar range with acceptable specificity, and the program needs sub-nanomolar affinity for clinical relevance.
Traditional affinity maturation has three problems.
**Most mutations don't help.** Across a typical CDR, ~80% of substitutions are neutral or deleterious. Random libraries spend ~80% of their diversity on unproductive mutants and ~20% on the candidates worth screening. Throughput screening can absorb this inefficiency, but it taxes the budget linearly.
**Epistatic combinations are hard to find.** Two mutations that improve affinity by 2× each may improve it by 3× or by 10× together. Or one may erase the other. Random sampling at fixed mutation rates rarely covers the combinatorial space densely enough to discover positive epistasis. The "ceiling" hit rate for traditional methods is set by what single mutations alone can deliver.
**Developability isn't selected.** Affinity is one fitness dimension. Aggregation propensity, stability, solubility, hydrophobic patch area, charge distribution are others. Pure affinity selection routinely drives the lead toward higher-affinity-but-less-developable space. The downstream cost (reformatting, failed expression, polyspecificity) is invisible at the selection step.
## DMS principles for antibodies
A DMS experiment for an antibody works by:
1. Constructing a library that systematically introduces every single-amino-acid substitution across the residues of interest (CDR positions, framework hotspots, or whole-domain coverage).
2. Coupling each variant to a measurable phenotype — for antibodies, this is almost always display level + antigen binding on yeast or mammalian display.
3. Reading out the population by NGS before and after a selection or sort, then computing the enrichment factor for each variant.
4. Mapping enrichment back to position-and-substitution to produce a per-residue, per-amino-acid fitness score.
We covered the general DMS workflow in [deep mutational scanning fitness landscapes](/deep-mutational-scanning-a-high-throughput-approach-to-mapping-protein-fitness-landscapes). For antibody affinity maturation specifically, three design choices matter.
**Single mutants vs combinatorial.** Single-mutant DMS (200 variants per 10-residue CDR) is the minimum viable design. It reveals which positions are tolerant and which are forbidden, and it identifies the top single substitutions per position. Combinatorial DMS (pairs or triples of mutations) reveals epistatic interactions but costs an order of magnitude more in library size. Run single mutants first; build a focused combinatorial library from the top single-mutant hits.
**Yeast vs mammalian display readout.** Yeast display is the default — fast, quantitative, normalized per-cell. Mammalian display is the right choice when developability is the bottleneck and the affinity-maturation library must respect glycosylation and complex disulfides. The [two-platform approach](/the-two-platform-approach-using-yeast-display-for-affinity-and-mammalian-display-for-developability) — yeast for affinity, mammalian for developability — generalizes to DMS-driven maturation.
**Sort strategy.** Affinity selection at decreasing antigen concentrations (10 nM → 1 nM → 100 pM across three sort rounds) generates a per-variant enrichment trace. Variants that enrich at low concentration are higher-affinity; variants that drop out are lower. The slope of the enrichment trace, not the round-3 endpoint alone, gives the most reliable affinity rank.
## Library design — saturation per CDR vs combinatorial
For most antibody affinity maturation campaigns, we run the following ladder:
**Phase 1 — Single-mutant scan, CDR-H3.** 200 variants for a 10-residue CDR-H3, plus 200 for CDR-H2, plus 100 for CDR-H1 (typically shorter). Sequenced before and after a single-round sort at the discovery-hit Kd. Output: position-by-position fitness map for the heavy chain.
**Phase 2 — Single-mutant scan, light chain.** If the heavy-chain scan didn't deliver enough affinity gain, scan CDR-L1 and CDR-L3. Light-chain contributions are smaller on average but non-zero, and the position-by-position map reveals whether the light chain has unexploited room.
**Phase 3 — Combinatorial top hits.** Take the top 5–10 single mutants per CDR. Build a combinatorial library of all pairwise combinations (~25–100 doubles) plus triples of the best singles (~50–200 triples). Sort under tightened stringency. The output is the candidate lead.
Each phase costs ~$5–15K in synthesis (Twist oligo pools) plus 2–4 weeks of yeast display and NGS. Total campaign cost is dominated by sorting and sequencing reagents, not by synthesis.
## Readout — yeast/mammalian display plus NGS
The NGS readout for antibody DMS is delicate because the variant counts are small (200–1,000 per library) and the read depth needs to cover the rarest variant with statistical confidence. The standard:
- **Pre-sort sequencing**: 500K–1M reads. Establishes the library composition. Confirms each variant is present at ≥100 reads.
- **Post-sort sequencing**: 1–5M reads. Higher depth because the population has been compressed and the variants of interest are now over-represented.
- **Enrichment calculation**: log2(post-sort frequency / pre-sort frequency). Variants with log2 enrichment > +1 are candidate gain-of-function mutants. Variants with log2 enrichment < −1 are candidate loss-of-function mutants. Variants between are neutral.
Software: enrich2, DiMSum, or a custom pipeline. For most campaigns, enrich2 is fine; the bookkeeping (read pairing, variant calling, statistical filtering) is what matters more than the algorithm choice.
For depth math and the relationship between library diversity, read count, and statistical confidence, see [calculating library diversity with NGS](/the-numbers-game-a-practical-guide-to-calculating-and-validate-library-diversity-with-ngs).
## From fitness landscape to optimized clones
The fitness landscape — a heatmap of position × substitution colored by enrichment score — is the deliverable that informs the next-round library. Three patterns to look for:
**Conserved positions.** Residues where only the wild-type tolerates substitution. These are typically structural — they form the CDR backbone or pack against framework residues. Don't waste combinatorial budget here.
**Tolerant positions.** Residues where most substitutions are accepted. These positions have engineering room: pick substitutions that improve affinity (the top of the heatmap column) and that also satisfy other constraints (developability, immunogenicity).
**Beneficial substitutions.** Specific (position, substitution) pairs with enrichment >+2. These are the candidate single-mutant gain-of-function hits. The top 5–10 per CDR feed the combinatorial library.
The combinatorial library is built by combining beneficial single mutants. Positive epistasis (combined effect greater than the sum of singles) is more common between mutations at positions that are spatially proximal in the CDR loop; negative epistasis is more common between distant mutations. Both happen; both are useful information.
After the combinatorial sort, the top 10–50 clones are reformatted to Fab or IgG, expressed, and characterized by SPR or BLI for absolute Kd. The conversion rate from "top NGS hit" to "validated high-affinity lead" in a well-designed DMS campaign is typically 60–85% — substantially higher than from a random-walk library, where conversion runs 10–30%.
## Developability constraint
Affinity selection alone optimizes for affinity. The DMS framework lets you add other selection axes without redesigning the experiment.
**Co-selection for stability.** Sort the post-affinity-selection library a second time on display level alone (without antigen). Variants that display well after antigen selection are stable AND high-affinity. Variants that display poorly are high-affinity but biophysically compromised.
**Co-selection for low polyspecificity.** Sort against a polyspecificity reagent (PSR) as a counter-screen. Variants that bind PSR drop out; variants that don't survive.
**Co-selection for thermal stability.** Heat-treat the displayed library before antigen binding. Variants that retain display after heat are thermostable; the rest enrich poorly even if their wild-type affinity was higher.
These co-selections double the experimental burden but deliver a candidate lead that's already past the developability triage step. For programs heading to therapeutic development, the upstream cost is repaid 5–10× in saved downstream re-engineering.
## When DMS isn't the right tool
DMS is a powerful affinity-maturation method, not a universal one. It struggles when:
- **The starting affinity is too weak.** If the discovery hit binds at Kd > 1 µM, the single-mutant landscape is dominated by noise — too few variants register signal above background. Improve affinity to below 500 nM by other methods first, then run DMS.
- **The library can't be displayed.** Membrane proteins, very large constructs, post-translationally modified targets that yeast can't make — these block display-based readouts. Alternative readouts (mammalian display, mRNA display) work but cost more.
- **Combinatorial optimization needs more than 3–4 simultaneous mutations.** DMS scales to pairs and triples; quadruples and higher need different methods (directed evolution, ML-guided sequence sampling).
For the cases where DMS does fit — most therapeutic antibody affinity maturation programs — it is the most efficient way we know to move from a discovery hit to a clinical-grade lead.
## Decision summary
If you have a discovery hit at 10–500 nM and need to reach sub-nanomolar: DMS-guided maturation is the right starting point. Single-mutant scan on CDR-H3 first, then expand based on what the landscape shows.
If your hit is at sub-nanomolar already and you need to improve developability: skip the affinity scan and run a stability-selected DMS instead. Same library design, different selection axis.
If your starting material is a library of candidates from AI design: DMS is the natural complement — AI generates the diversity, DMS measures it.
---
Ranomics designs and runs DMS campaigns for antibody affinity maturation end-to-end. If you're scoping a program and want help designing the library and selection strategy, see our [affinity maturation services](/applications/affinity-maturation), [deep mutational scanning services](/technology/deep-mutational-scanning), or reach out via the [contact page](/ranomics-contact).
{/* ranomics:related-services */}
## Related Ranomics services
- **[Affinity maturation](/applications/affinity-maturation):** Antibody affinity maturation campaigns combining DMS, display screening, and developability triage.
- **[Deep mutational scanning](/technology/deep-mutational-scanning):** Systematic mapping of protein fitness landscapes.
- **[Yeast surface display services](/yeast-display):** Display platform for DMS readouts.
---
## Avidity Artifacts in Yeast Display: Detecting False Positives
> A practical guide to detecting and eliminating avidity artifacts in yeast display screening. Covers off-rate selection, soluble competition assays, display level titration, and FACS gating strategies with specific concentrations and timescales.
Source: https://ranomics.com/eliminating-false-positives-mastering-avidity-effects-in-yeast-display-screening/
Published: 2025-08-11
## Avidity Is an Apparent Affinity Problem
Every yeast cell in a display library presents 10,000 to 100,000 copies of a displayed protein on its surface. When a fluorescently labeled target binds to one copy and dissociates, it does not diffuse into bulk solution. It encounters another displayed copy within nanometers and rebinds almost immediately. The result: the observed off rate from the cell surface is orders of magnitude slower than the true monovalent off rate. The displayed protein appears to bind with low nanomolar affinity when the true monovalent KD may be mid to high micromolar.
This is avidity. It is not a vague concern. It is a quantitative, predictable artifact that inflates apparent affinity by 100 to 1000 fold in standard yeast display experiments. The mechanism is straightforward: multivalent display creates a high local concentration of binding sites, which drives rapid rebinding after each dissociation event. The observed dissociation rate constant (koff,apparent) reflects the probability of complete escape from the cell surface, not the intrinsic koff of a single binding interaction.
## What Happens When You Ignore It
If you sort a yeast display library by equilibrium binding signal alone, you select for two populations: genuine high affinity binders and avidity dependent false positives. These false positives look identical during FACS. They stain brightly, they sort cleanly, and they enrich over multiple rounds.
The failure appears downstream. You express the hits as soluble monomeric proteins, measure binding by SPR or BLI, and find that half or more of your enriched clones show no detectable binding at the concentrations you expected. The campaign has consumed weeks of sorting, sequencing, and expression work on clones that never had meaningful monovalent affinity. This is not a rare edge case. It is the default outcome of naive equilibrium sorting on high copy number display systems.
## Detection: Titration Curves Reveal the Signature
The most reliable diagnostic for avidity is a binding titration curve. Label your target at a series of concentrations (typically 0.1 nM to 1 µM in half log steps) and measure the fraction of cells that stain positive at each concentration. Plot the resulting curve.
A genuine high affinity interaction produces a sigmoidal curve with a sharp transition centered near the true KD. An avidity driven interaction produces a broad, shallow curve that shifts left (toward lower apparent KD) as display level increases. If you see binding at concentrations 100 fold below what SPR or ITC would predict for the monovalent interaction, avidity is dominating your signal.
You can confirm this by comparing titrations on high expression and low expression subpopulations within the same library. If the apparent KD shifts with display level, the binding is avidity dependent.
## Elimination Strategy 1: Off Rate Selection
Off rate selection is the most effective single method for eliminating avidity artifacts, and understanding why requires understanding the mechanism.
Avidity inflates apparent affinity primarily through rebinding. A target molecule dissociates from one displayed copy and immediately rebinds a neighboring copy before it can diffuse away. This means avidity affects the apparent off rate, not the intrinsic on rate. When you select for slow dissociation under competition conditions, you specifically select against the rebinding phenomenon that drives avidity.
The protocol: label cells at saturating target concentration (at least 10x the expected KD, typically 100 nM to 1 µM of labeled target). Wash once to remove unbound target. Resuspend cells in a large excess of unlabeled soluble target (50 to 100x the labeled target concentration) to act as a sink for any dissociated labeled molecules. Incubate at room temperature. The key variable is the competition time.
For binders in the low nanomolar range (1 to 10 nM KD), a competition time of 1 to 4 hours is appropriate. For binders in the sub nanomolar range, extend to 8 to 24 hours, checking for cell viability at each timepoint. For initial enrichment rounds where you expect a broad affinity distribution, start with a short competition (30 to 60 minutes) to remove the weakest binders without losing moderate affinity clones.
After competition, sort cells that retain fluorescent signal. These are clones where the labeled target remained bound through the competition period, indicating a slow intrinsic off rate that does not depend on rebinding.
## Elimination Strategy 2: Display Level Titration
Reducing the number of displayed copies per cell directly reduces the local concentration that drives rebinding. If each cell displays 500 copies instead of 50,000, the spacing between copies increases from ~10 nm to ~100 nm, and the rebinding probability drops dramatically.
In practice, you achieve this by titrating the inducer concentration. For the Aga2p system in S. cerevisiae, reduce galactose from the standard 2% to 0.05 to 0.2% and induce for a shorter period (4 to 8 hours instead of overnight). Confirm display levels by co staining with an anti tag antibody and gating on the lowest 10 to 20% of expressors during FACS.
This approach has a tradeoff: lower display levels reduce signal to noise, making it harder to detect weak binders. Use display level titration in later rounds of selection when the library is already enriched and you are discriminating between moderate and high affinity clones.
## Elimination Strategy 3: Soluble Competition Assays
Soluble competition directly measures monovalent binding by removing the surface from the equation. Pre incubate cells with varying concentrations of unlabeled soluble target (serial dilutions from 10 µM to 1 pM) for 1 hour at room temperature. Then add a fixed concentration of labeled target (at or near the expected KD) and incubate for another hour. Measure the fraction of cells that stain positive.
Genuine high affinity binders will resist competition: the pre bound soluble target occupies the binding site and prevents labeled target from binding. Avidity dependent binders will show competition at much lower soluble target concentrations than their apparent surface KD would predict, because the soluble interaction reflects true monovalent affinity.
The IC50 from this competition curve approximates the true monovalent KD. Compare it to the apparent KD from surface titration. A discrepancy of more than 10 fold is diagnostic of significant avidity contribution.
## FACS Gating for Off Rate Selection
During off rate sorts, your gating strategy determines whether you successfully eliminate avidity artifacts or simply re enrich them.
Gate on a two dimensional plot of binding signal (target fluorescence) versus display level (anti tag fluorescence). Draw your sort gate on cells that show high binding signal relative to their display level. This ratiometric approach normalizes for expression variation and specifically enriches clones where binding strength does not depend on copy number.
Exclude the highest 10 to 20% of expressors entirely. These cells are most susceptible to avidity, and including them reintroduces the artifact you are trying to eliminate. Set your binding threshold stringently: collect only the top 1 to 5% of the binding/display ratio distribution.
## The Decision Framework
For early rounds of selection (rounds 1 and 2), use equilibrium sorting with moderate stringency to reduce library diversity. Avidity is acceptable here because you are casting a wide net.
For intermediate rounds (rounds 3 and 4), introduce off rate selection with short competition times (30 to 60 minutes) and begin gating against high expressors. This removes the most egregious avidity dependent clones while preserving moderate affinity binders.
For final rounds (rounds 5 and beyond), use stringent off rate selection (2 to 24 hour competition) combined with low display level induction and ratiometric gating. At this stage, every clone that survives should have genuine monovalent affinity.
Validate your final hits as soluble monomeric proteins by SPR or BLI before committing to downstream characterization. The correlation between surface apparent KD and solution KD should be within 3 to 5 fold if avidity has been adequately controlled.
## The Takeaway
Avidity artifacts are not a minor technical nuisance. They are the primary source of false positives in yeast display binder campaigns. Off rate selection eliminates them because it targets the rebinding mechanism directly. Combined with display level control and ratiometric FACS gating, these methods convert yeast display from a system that enriches for surface affinity into one that reports true monovalent binding.
Every binder campaign we run at Ranomics incorporates these controls from the first sort. The result is a hit list where surface enrichment data predicts solution phase binding, and downstream validation confirms what the screen already told us.
If you are running binder campaigns and seeing poor hit rates in solution phase validation, the problem is likely upstream. [Start a project](/ranomics-contact) and we will design a screening strategy that eliminates avidity artifacts before they waste your downstream effort.
For a worked example of these controls applied across a six-cycle campaign, see our case study on [pH-dependent antibody engineering via yeast surface display](/ph-dependent-antibody-engineering) — display-level normalization and ratiometric gating from a real 640-clone library.
{/* ranomics:related-services */}
## Related Ranomics services
- **[Yeast surface display services](/yeast-display):** Avidity-aware screening design with off-rate selection and soluble competition.
- **[Binder discovery](/applications/binder-discovery):** End-to-end binder discovery campaigns with reliable affinity ranking.
- **[Case study: pH-dependent antibody engineering](/ph-dependent-antibody-engineering):** 14-page technical walkthrough of a real client campaign.
---
## enzyme-thermostability-engineering.mdx
Source: https://ranomics.com/enzyme-thermostability-engineering/
import InOurWork from "../../components/InOurWork.astro";
Industrial enzymes are run at temperatures that natural enzymes were never optimized for. Cellulases in biofuel processing operate at 50–60 °C. Transaminases in pharmaceutical intermediate synthesis run at 40–60 °C for days. PCR polymerases see 95 °C in every cycle. Wild-type enzymes from mesophilic hosts denature at these temperatures within minutes; the cost of replacing them dominates process economics. Thermostability engineering is the solution.
This article is a playbook for the methods that, in our work, consistently deliver +10–20 °C Tm shifts while preserving activity. It covers consensus design, directed evolution, DMS-guided strategies, and the validation that distinguishes a genuine thermostability gain from an in-vitro artifact.
## Why thermostability matters
The economic argument is direct. An industrial enzyme that loses half its activity per hour at the process temperature requires either continuous addition (raising reagent cost) or temperature reduction (slowing reaction rate). A thermostable variant that retains 90% activity after 24 hours runs the same process at much lower enzyme cost.
The corollary arguments:
- **Higher operating temperature accelerates reaction rates** (typically 2× per 10 °C). A more thermostable enzyme allows a faster, more compact process.
- **Higher operating temperature suppresses microbial contamination.** Stable enzymes can run at temperatures that prevent bacterial fouling without sterilization steps.
- **Storage and shipping costs drop.** A thermostable enzyme can be stored at ambient temperature for months; a marginal-stability enzyme requires cold chain.
For the high-volume industrial applications (detergent proteases, food enzymes, biofuel cellulases), thermostability engineering is the difference between a viable product and a research curiosity.
## Strategy 1 — Consensus design
The simplest method, often the first to try. Compile a multiple-sequence alignment (MSA) of homologs across the family. At each position where the wild-type residue differs from the consensus (most common residue across the alignment), test the consensus substitution.
The logic: evolutionarily conserved residues are conserved for stability reasons. A residue that appears in 80% of homologs at a given position has been selected across hundreds of millions of years against destabilization. Substituting the wild-type residue with the consensus often delivers a +1 to +3 °C Tm gain at zero design cost.
Practical limits:
- Works best when the MSA has 100+ diverse homologs. Smaller alignments are noisier and false positives are common.
- Consensus substitutions at active-site positions risk killing activity. Filter the consensus list by spatial proximity to active site (>8 Å away as a starting cutoff).
- Combinatorial consensus (mutating 5–10 positions simultaneously) often delivers super-additive gains, but it also fails super-additively when the substitutions are co-dependent. Test individually first, then combine top hits.
A consensus design pass typically delivers a +5 to +10 °C Tm shift in 3–6 weeks at minimal cost. It is the right starting point for any thermostability campaign.
## Strategy 2 — Directed evolution with thermochallenge
The classical method. Build an error-prone PCR library or a structural-loop-shuffled library. Express the library in a host capable of survival selection at elevated temperature, or screen the library by heat-treating before activity assay.
The standard workflow:
1. Generate a library of ~10^5 variants by error-prone PCR.
2. Express each variant (typically in 96-well plate format for plate-based screening, or in fluorescent display format for FACS-based screening).
3. Heat-challenge each variant at a temperature above the wild-type Tm (e.g., 70 °C for 30 min if wild-type Tm is 60 °C).
4. Assay residual activity. Variants with ≥50% residual activity are candidates.
5. Sequence the top 10–50 candidates. Confirm gains in liquid-format Tm measurements.
We covered the directed evolution mechanics in [the technical guide to directed evolution](/a-technical-guide-to-directed-evolution-for-enhancing-protein-stability-and-function). For thermostability specifically:
- **Iterative rounds compound.** Most successful campaigns run 5–10 rounds, each adding 1–3 stabilizing mutations. The final variant has Tm shifted +10–25 °C above wild-type.
- **Watch for activity loss.** Pure thermostability selection can drive the enzyme toward higher-stability-but-inactive variants. Co-select for activity at every round by including an activity assay at the screening step.
- **Heat-challenge stringency matters.** Too lenient (low temperature, short duration) and you don't enrich; too harsh and the surviving population is too small to find the rare gain-of-function mutants.
## Strategy 3 — DMS-guided stability engineering
The most efficient method when high-throughput screening exists. Build a single-mutant scanning library covering the protein, sort under thermal challenge, and read out by NGS.
The output is a per-residue, per-substitution fitness map under thermal stress. Stabilizing substitutions register as positive enrichment; destabilizing as negative. The top 10–30 stabilizing single mutants then feed a combinatorial library.
We've covered DMS in detail in [DMS for protein engineering](/deep-mutational-scanning-a-high-throughput-approach-to-mapping-protein-fitness-landscapes). For thermostability specifically:
- Single-mutant DMS reveals which positions tolerate substitution AND which substitutions stabilize. Both pieces of information are essential.
- Combinatorial assembly of top single mutants captures positive epistasis. Stabilizing mutations often combine super-additively when they're in different structural regions.
- Co-selection for activity (sort first under thermal challenge, then under activity selection) avoids the activity-loss failure mode.
For programs where the enzyme can be displayed and where high-throughput thermal challenge is feasible, DMS-guided design is the highest-information-per-experiment method. We cover the AI-augmented variant in [leveraging AI and DMS to engineer enzymes](/leveraging-ai-and-deep-mutational-scanning-to-engineer-enzymes).
In our enzyme work, we have found that custom models trained on bespoke deep-mutational-scanning fitness maps consistently outperform off-the-shelf open-source generators. The training data in the public structural databases is too sparse on enzyme fitness for general-purpose models to do the job. The path that has been most reliable for us is to generate the DMS data first, train a small model on the resulting sequence-to-fitness map, and then sample stabilizing combinations that pure random sampling would not have reached in the same library size.
## Strategy 4 — Structure-based design
When a high-resolution structure exists, computational methods can identify candidate stabilizing mutations:
- **Rosetta ddG calculations** identify mutations that decrease folding free energy. False positives are common but the predictions narrow the search space.
- **B-factor minimization** targets residues in high-B-factor regions for stabilization. The logic: flexible regions are the first to unfold under thermal stress.
- **Disulfide engineering** introduces cysteine pairs that constrain the fold. Effective when the geometry permits, but few candidate sites typically exist per protein.
- **Salt-bridge optimization** at the protein surface adds favorable electrostatic interactions. The contribution per salt bridge is modest (+0.5 to +1.5 °C) but additive.
Structure-based design delivers candidate lists that feed into experimental campaigns. Treat the computational predictions as a focused library, not as final answers.
## Validation — Tm by DSF/DSC, activity at temperature
Thermostability claims must be validated by orthogonal measurements:
- **Differential Scanning Fluorimetry (DSF)** or **nanoDSF**: the workhorse Tm measurement. Throughput-friendly, requires small protein quantities, gives a single melting transition for most enzymes.
- **Differential Scanning Calorimetry (DSC)**: the gold standard. Reports the full unfolding thermogram, distinguishes single vs multi-domain unfolding, gives ΔH directly. Lower throughput but unambiguous.
- **Activity assay at temperature**: Tm and operational thermostability are correlated but not identical. The variant that survives 70 °C for 30 minutes may not catalyze the target reaction at 70 °C. Always run activity assays at the intended operating temperature.
The combination that we use as the validation standard: nanoDSF for ranking (high-throughput, ~50 variants per day) plus activity assay at the operating temperature for top 10–20 candidates.
## Trade-offs — stability vs activity
The stability-activity trade-off is real but bounded. Three patterns recur:
**Surface mutations.** Stabilizing substitutions on the protein surface rarely affect activity. They alter solvent interactions without changing the active site. Almost all consensus-derived gains are surface or near-surface.
**Core-packing mutations.** Stabilizing substitutions in the hydrophobic core (filling cavities, optimizing van der Waals contacts) usually preserve activity unless they propagate to the active site. Most computational ddG predictions target core positions.
**Active-site mutations.** Mutations within 6–8 Å of the active site can be stabilizing but very often reduce kcat. These mutations need careful activity testing.
The rule: prioritize surface and core mutations first. Only return to active-site stabilization if surface and core changes don't reach the Tm target.
## Decision summary
For a new thermostability campaign, run the methods in order of cost:
1. Consensus design (1–2 months). Cheap, often delivers +5–10 °C without extensive experiments.
2. DMS-guided stability engineering (3–4 months). High-information output, scales well with a good screening setup.
3. Directed evolution with thermochallenge (3–6 months). Lower information density per cycle, but no requirement for high-throughput screening infrastructure.
4. Structure-based predictions (1 month, layered on top of 1–3). Improves candidate selection in any of the above methods.
Most successful campaigns combine all four. The combinations compound; the trade-off question is which method to start with given the available infrastructure and timeline.
---
If you're scoping an enzyme thermostability campaign, see our [enzyme engineering services](/applications/enzyme-engineering) or reach out via the [contact page](/ranomics-contact). For multi-objective protein engineering combining stability with activity and substrate scope, see [protein engineering services](/protein-engineering).
{/* ranomics:related-services */}
## Related Ranomics services
- **[Enzyme engineering](/applications/enzyme-engineering):** Thermostability, activity, and substrate-scope campaigns.
- **[Directed evolution](/directed-evolution-protein-engineering):** Iterative mutagenesis and selection for protein engineering objectives.
- **[Deep mutational scanning](/technology/deep-mutational-scanning):** Systematic fitness mapping for stability and activity.
---
## From an AlphaFold Model to Your First Binder: A Walkthrough for Teams Without Structural Biology Expertise
> A practical, step-by-step guide for small biotech and academic teams who have an AlphaFold model of their target but no structural biologist on staff — what to check, what to decide, and how to move into a binder design campaign.
Source: https://ranomics.com/from-alphafold-model-to-first-binder/
Published: 2026-04-20
import TryToolCallout from "../../components/TryToolCallout.astro";
import InOurWork from "../../components/InOurWork.astro";
If your team has an AlphaFold model of a target protein and no structural biologist on staff, you are in the same position as most seed-stage biotech companies that want to run a de novo binder design campaign. AlphaFold and AlphaFold 3 have made structural models easy to get. The harder question is: can you use this model to actually design binders, and if so, where do you target them? This post walks through the decisions in plain language, without assuming a structural biology background.
## Step 1: Decide Whether Your Model Is Usable
Not every AlphaFold model is a valid starting point for binder design. The key number to look at is pLDDT — AlphaFold's per-residue confidence score, from 0 to 100. Roughly:
- **pLDDT above 90**: high confidence. Treat this region as you would a crystal structure.
- **pLDDT 70-90**: reasonable confidence. Usable, but check the backbone against any available experimental data.
- **pLDDT 50-70**: low confidence. Usable only for rough context; do not design against this region.
- **pLDDT below 50**: effectively no structural information. This region is almost certainly disordered or poorly predicted.
For binder design, what matters is the pLDDT **locally, around the region you want to target**, not the average over the whole protein. A target with overall pLDDT of 85 but pLDDT of 55 in the exact surface patch you want to bind is not a usable starting point for that patch.
How to check this without a structural biologist: open the PDB file in a viewer that color-codes by pLDDT (ChimeraX does this automatically; PyMOL can be scripted; Mol\* in a web browser is the fastest). Look at the regions near the surface you care about. If they are red or orange (low pLDDT), pick a different region or invest in experimental structure determination first.
A second check: AlphaFold's PAE (predicted aligned error) plot. If you are using AlphaFold 3, the PAE plot shows how confident the model is about relative positions between different parts of the protein. Large off-diagonal values mean the model has high internal uncertainty about how two domains sit relative to each other. For binder design, you want low PAE in the region you are targeting.
## Step 2: Choose Where on the Protein to Target
This is the single most consequential decision in the entire campaign. It is more consequential than which generative model you use, how many designs you generate, or which display platform you screen on.
A good binder target surface has five properties, drawn from the protein-protein interaction literature:
1. **Hydrophobic enough to drive desolvation.** Binder-target interfaces form because hydrophobic surfaces prefer each other to water. Pure polar or charged patches are hard to bind de novo.
2. **Structurally ordered.** Beta-strands and alpha-helices are rigid, preorganized, and predictable. Loop-dominated surfaces move too much.
3. **Rigid.** Low B-factor in a crystal structure, high pLDDT in a model. Flexibility costs entropy at binding and is the single most common reason computational designs fail experimentally.
4. **Accessible.** The designed binder has to physically reach the surface. Deep pockets, grooves, and buried sites are harder than flat, exposed surfaces.
5. **Populated with hot-spot residues.** Trp, Tyr, Arg, and Phe disproportionately contribute to binding free energy at protein-protein interfaces. Patches that contain them are easier to engage.
Assessing these by eye is what a structural biologist would normally do with PyMOL. [Epitope Scout](/technology/epitope-scout) automates exactly this analysis — upload a PDB file, select a chain, and get a ranked list of candidate epitope patches scored on these five criteria. It is free, and the output is a CSV of residue selections that can be fed directly into RFdiffusion as a hotspot specification.
If there are existing antibodies or natural binding partners against your target with solved co-crystal structures, those contact surfaces are also valid targeting options — they are validated bindable by construction. The [SAbDab](https://opig.stats.ox.ac.uk/webapps/sabdab-sabpred/sabdab) and RCSB PDB are the databases to check.
In our practice, we strongly prefer running smaller design batches against three or four ranked patches rather than committing the full design budget to a single epitope. The chance of finding a working binder scales with epitope diversity, not with depth on any one surface, so it pays to keep the per-epitope batch lean and run several in parallel.
## Step 3: Pick a Design Algorithm
You have three mature open-source options for de novo binder design:
- **RFdiffusion.** Generates protein backbones conditioned on target hotspots. You provide the target structure and the hotspot residues; it generates backbones predicted to make contacts at those positions. Needs ProteinMPNN for sequence design after backbone generation.
- **BindCraft.** Jointly optimizes binder structure and sequence against an AlphaFold 2 confidence objective. Produces binders that are simultaneously optimized for fold and binding.
- **Boltzgen.** Boltzmann-weighted conformational sampling, useful when target flexibility matters or when you want design-time structure ensembles.
For a first campaign, RFdiffusion + ProteinMPNN is the best-characterized path. BindCraft tends to give higher filtered hit rates when the target is rigid and well-structured. Boltzgen is specialized. If budget is constrained, pick one algorithm and run it well, rather than running all three poorly.
## Step 4: Filter Before You Build
Generative models produce backbones that look plausible in silico. They do not all fold correctly in reality. Before committing any candidate to gene synthesis, run a self-consistency check:
1. Take the designed sequence.
2. Fold it with an independent structure predictor — ESMFold, ColabFold, or Boltz-2.
3. Align the predicted structure back to the intended backbone.
4. If the RMSD is above roughly 2 Å, the sequence does not encode the structure you designed. Discard it.
This filter typically removes 50-80% of raw RFdiffusion output and is the single most important computational quality gate. Campaigns that skip it synthesize candidates that were never going to fold and waste weeks of experimental time.
Other filters worth applying before synthesis:
- **Expression prediction.** Very hydrophobic, very charged, or very long candidates often fail to express.
- **Solubility prediction.** SolubleMPNN or simple net-charge / hydrophobic-patch heuristics.
- **Target binding prediction.** AlphaFold 2 co-folding of designed binder + target with pLDDT on the interface residues.
A 1,500-design RFdiffusion run typically filters down to 200-500 candidates worth synthesizing.
## Step 5: Decide How to Validate Experimentally
You have three broad options for experimental validation of designed binders:
- **Yeast surface display.** The workhorse for de novo binder validation. Well-validated for extracellular targets, high throughput (thousands to hundreds of thousands of variants), quantitative FACS readout, NGS-compatible.
- **Mammalian display.** Preferred when the displayed protein requires mammalian post-translational modifications. More expensive and slower than yeast.
- **One-by-one purification.** Express each top candidate as a purified protein and run SPR, BLI, or a direct binding assay. Low throughput but gives clean affinity measurements. Viable for 5-20 candidates from a filtered pool; infeasible for hundreds.
Most teams running a first campaign combine yeast display (for pool-level hit calling) with purification of the top 5-10 ranked hits (for orthogonal affinity validation).
In our work, raw mini-binder designs out of a single RFdiffusion plus ProteinMPNN round typically land in the 10 to 100 nM range before any affinity maturation. Sub-10 nM straight out of the gate happens, but it is exceptional and usually means the target was particularly cooperative. Plan the validation funnel for that band: if the downstream application needs single-digit nanomolar or better, budget for an affinity maturation round on the top hits rather than expecting it to come out of the first design pass.
## If Your Team Does Not Have Display Infrastructure
This is the situation most seed-stage biotechs are in. Setting up yeast display, FACS, and NGS analysis from scratch is a six-month project and requires a bench scientist with the relevant training. For a first campaign, it is usually faster to outsource the experimental side.
The [Binder Pilot](/binder-pilot) program at Ranomics is scoped exactly for this case — target structure in, ranked hit list out. We do not replace the strategic decisions (target choice, epitope selection, downstream follow-up); we handle the experimental throughput. The scoping call is the right place to discuss whether your starting AlphaFold model is good enough and which epitope patches to prioritize.
For multi-target pipelines or teams that need the milestone structure and 100% binder guarantee of a flagship program, the [AI Binder Sprint](/ai-binder-sprint) is the scope up from a Pilot.
## Common Pitfalls
A few failure modes to watch for when moving from an AlphaFold model to a first binder campaign:
- **Targeting a disordered loop that looks plausible in cartoon view.** Always check pLDDT or B-factors locally.
- **Designing against an interface that is only bindable in a PTM-dependent form.** If the native binding partner only engages a phosphorylated or glycosylated version of your target, a de novo binder on a non-modified recombinant protein will miss.
- **Skipping self-consistency filtering.** Trusting raw generative output directly into gene synthesis is the fastest way to burn a budget on non-folding candidates.
- **Choosing a hotspot far from the functional site.** A binder that binds the target is not necessarily a binder that inhibits or activates the target. If your downstream readout is a functional assay, the hotspot selection has to consider mechanism, not just bindability.
- **Validating only on SPR or BLI without a display step.** In our pipeline, direct purification followed by biolayer interferometry will report on-target affinity for a sticky, polyspecific binder without flagging the polyspecificity itself. We run every design candidate through a display screen first so the off-target signal is visible in the same experiment as the on-target signal, and polyspecific hits get filtered out before any candidate is purified.
## Summary
An AlphaFold model is a starting point, not a finished target. The work between "we have a model" and "we have a validated binder" is mostly about three decisions: is the model usable, where on the target should you bind, and how will you filter and validate. The tools to answer those questions are largely free and open-source. The experimental infrastructure is where most small teams run into bandwidth constraints — and that is usually the right place to outsource.
{/* ranomics:related-services */}
## Related Ranomics services
- **[Binder Pilot](/binder-pilot):** Scoped for academic / seed-biotech teams starting their first binder campaign.
- **[Epitope Scout](/technology/epitope-scout):** Free tool: identify exposed, designable epitopes on your AlphaFold model.
---
## Glycoprotein Engineering: Yeast Can't Do This, Mammalian Can
> Glycoprotein engineering requires mammalian PTM machinery. When to skip yeast display and go straight to CHO/HEK293.
Source: https://ranomics.com/glycoprotein-engineering-yeast-vs-mammalian/
Published: 2026-05-11
Glycoprotein engineering is the part of antibody discovery and protein engineering where the choice of expression host stops being a workflow preference and becomes a hard constraint. Yeast cannot install mammalian-complex N-glycans. *E. coli* cannot install N-glycans at all. For targets whose biology depends on those glycans — and for antibody Fc engineering where effector function is the deliverable — the campaign has to be built around mammalian cells from the start.
This article is the decision framework for when glycosylation forces you out of yeast display and into mammalian, plus a quick survey of the glycoengineering tools that the field uses to install specific glycan profiles.
## What N-glycosylation actually does
N-linked glycosylation isn't ornamental. The glycan chain on an Asn-X-Ser/Thr motif contributes to:
- **Protein folding and stability.** The high-mannose precursor (Man9-GlcNAc2) is a quality-control signal in the ER; glycoproteins are retained for refolding until the signal is processed correctly.
- **Half-life in circulation.** Sialylated complex glycans extend serum half-life by masking the asialoglycoprotein receptor on hepatocytes. Desialylated glycoproteins are cleared rapidly.
- **Receptor binding and signaling.** Many cell-surface receptors require their N-glycans for ligand interaction. The IgG Fc-FcγR interaction is the canonical example.
- **Effector function.** Antibody Fc fucosylation status dictates ADCC potency; afucosylated antibodies show 10–100× enhanced ADCC.
- **Immunogenicity.** Non-human glycans (yeast high-mannose, plant xylose, α-galactose) are immunogenic in humans. Therapeutic biologics must avoid them.
The glycan is part of the molecule's biology. Engineering the protein without engineering (or at minimum specifying) the glycan is incomplete.
## Yeast glycosylation — high-mannose only
*S. cerevisiae* and *P. pastoris* install glycans through the same ER pathway as mammalian cells up to the Man9-GlcNAc2 precursor. Then the pathways diverge. Yeast Golgi enzymes extend the mannose chain into hyperglycosylated high-mannose structures (Man20+ for *S. cerevisiae*, Man8-Man15 for *P. pastoris*). Mammalian Golgi enzymes trim mannose and add complex glycans.
For a yeast-displayed antibody discovery campaign, two consequences:
1. **The displayed antibody itself is hyperglycosylated.** If the antibody has N-glycosylation sites near the binding interface, those sites carry yeast-style mannose during display but will carry complex glycans in mammalian production. Apparent affinity on yeast and final affinity in production may differ.
2. **The target, if displayed against, is yeast-style if produced in yeast.** Antibodies selected against yeast-produced glycoprotein target may not bind the mammalian-produced version. This is the larger concern.
For pure protein-epitope targets (no glycan involvement in the binding mode), yeast display works fine. For glycan-involving targets, yeast is the wrong host.
## Mammalian glycosylation — CHO vs HEK profiles
Two production hosts dominate, with different glycan profiles.
**CHO** installs complex biantennary N-glycans with low sialylation (most CHO lines produce mostly α2,3-sialylation, less α2,6 than human cells). High core fucosylation is the default. Galactosylation is moderate. The glycan profile is well-characterized and regulatory authorities are familiar with it.
**HEK293** installs complex N-glycans closer to native human — higher sialylation, both α2,3 and α2,6 linkages present, lower core fucosylation than CHO. Useful for research-grade antibodies that need human-like glycosylation, but HEK is rarely the production host for therapeutic biologics.
Neither host installs identical glycans to those a human cell would. The differences are typically small and clinically tolerable, but glycoengineered variants of both hosts exist (see below) when specific glycan profiles are required.
## When the campaign needs mammalian display from the start
The decision criteria:
**Target is a glycoprotein with glycan-dependent epitopes.** Examples: complement components, mucin-domain proteins, glycan-specific therapeutic targets (sialyl-Lewis X for cancer, GD2 for neuroblastoma, blood group antigens). Yeast-displayed antibodies will miss these epitopes systematically.
**The antibody format is full-length IgG.** Yeast can't display full IgG efficiently. If the program is post-discovery and needs full IgG validation, mammalian is the right host.
**Effector function is part of the deliverable.** ADCC, CDC, or ADCP potency depends on Fc glycosylation. Engineering for these activities requires mammalian production from the start, ideally in glycoengineered host lines.
**Late-stage developability validation.** Even for non-glycoprotein targets, the mammalian production glycoform is what reaches the clinic. Validating developability in mammalian cells catches issues yeast can't predict.
If none of these conditions apply, yeast display is fine. Run yeast for discovery; validate in mammalian downstream as part of the [two-platform workflow](/the-two-platform-approach-using-yeast-display-for-affinity-and-mammalian-display-for-developability). If any apply, mammalian display from round one. We covered the platforms in [mammalian cell display: CHO and HEK293](/mammalian-cell-display-cho-hek293).
## Engineering tools — glycoengineering CHO lines
When a defined glycan profile is the deliverable, the host cell line itself is engineered:
**Afucosylated lines (FUT8 KO).** Knockout of α1,6-fucosyltransferase eliminates core fucose. Antibodies produced in these lines show 10–100× higher ADCC potency. Commercial lines: GlycArt (Roche/Glycart-style), Potelligent (BioWa), Glymaxx. Used for clinical antibodies targeting solid tumors.
**Bisected GlcNAc lines (GnTIII overexpression).** Adds a bisecting N-acetylglucosamine that also enhances ADCC. Less dramatic effect than fucosylation removal but pairs with it.
**Galactosylation engineering.** β1,4-galactosyltransferase overexpression increases terminal galactose content. Relevant for biosimilar matching to innovator glycan profiles.
**Sialylation engineering.** α2,6-sialyltransferase knock-in delivers more human-like sialylation. Used when the program targets the FcRn-mediated half-life or anti-inflammatory IgG (IVIG-like) activity that requires sialylated Fc.
**Mannose receptor avoidance.** Engineered to avoid high-mannose contamination from incomplete processing — relevant for biologics where mannose-receptor-mediated clearance shortens half-life.
For most clinical antibody programs, the host is either an off-the-shelf CHO line or one of the established glycoengineered variants (Potelligent for ADCC, BioWa for bisecting glycans). For research-grade work, HEK293 with its more human-like glycosylation is often sufficient.
## Practical workflow for glycoprotein engineering
For a target where glycosylation matters:
1. **Define the glycan profile target.** Specific glycoforms? Avoid certain glycans? Match an innovator? The downstream host choice depends on this.
2. **Source target antigen from mammalian cells.** Yeast-produced or *E. coli*-produced antigen for selection is a false economy if the production target is glycosylated.
3. **Build the antibody library in mammalian display from the start.** CHO or HEK293 depending on stage; engineered host lines if specific glycoforms are required.
4. **Counter-select against alternative glycoforms.** During sorting, include negative selection against the unwanted glycoform to enrich for glycan-specific binders.
5. **Validate in production-relevant host.** Confirm the final lead's binding mode and effector function match expectations in the production line.
The cost premium of mammalian-only vs yeast-led discovery is 2–4× in screening cost and 1.5–2× in wall-clock. For glycoprotein-relevant programs, the cost is the price of getting the right answer.
## Decision summary
If the target involves glycan-dependent epitopes, or the antibody format is full-length IgG, or effector function is a deliverable: skip yeast and go to mammalian display from round one.
If glycoengineering is required (afucosylation, defined sialylation, bisecting GlcNAc): use an engineered CHO line as the production host and run mammalian display in the same line.
If the target is a soluble protein with no glycan dependence and the antibody format is scFv or VHH: yeast display is the right discovery platform; mammalian display joins the workflow downstream for developability validation.
---
If you're scoping a glycoprotein engineering or PTM-dependent target program, see our [mammalian display services](/technology/mammalian-display) or reach out via the [contact page](/ranomics-contact). For combined yeast-and-mammalian workflows, see the [AI Binder Sprint](/ai-binder-sprint).
{/* ranomics:related-services */}
## Related Ranomics services
- **[Mammalian display](/technology/mammalian-display):** CHO and HEK293 display platforms for glycoprotein and PTM-dependent targets.
- **[Cell engineering](/cell-engineering):** Glycoengineered CHO line development.
- **[AI Binder Sprint](/ai-binder-sprint):** Multi-platform programs covering glycoprotein targets end-to-end.
---
## hotspot-guided-protein-binder-design.mdx
Source: https://ranomics.com/hotspot-guided-protein-binder-design/
import TryToolCallout from "../../components/TryToolCallout.astro";
import InOurWork from "../../components/InOurWork.astro";
When designing protein binders from scratch, one of the most consequential decisions is how you constrain the computational search. An unconstrained RFdiffusion campaign will generate diverse backbones that geometrically approach the target surface, but with no guidance about where on that surface the binder should actually make contact. The result is a large pool of candidates that engage varying epitopes with varying efficiency, most of which will not bind detectably.
Hotspot conditioning changes this. By specifying which residues on the target must be contacted by the designed scaffold, you focus the generative model on the region of the target where binding energy can actually be extracted.
## Not all interface residues contribute equally to binding energy
The principle is well established in protein-protein interaction biochemistry: at most natural protein interfaces, a small subset of residues (typically 2-8 out of 10-20 interface contacts) account for the majority of the binding free energy. These are the hotspot residues. The remainder of the interface contributes little to delta-G(binding) and is largely structural or solvent-exclusion.
Alanine scanning mutagenesis, the classic method for identifying hotspots, replaces each interface residue with alanine and measures the change in binding affinity. Residues where the Ala mutation causes a delta-delta-G of >1-2 kcal/mol are operationally defined as hotspots. This has been measured for hundreds of protein-protein interactions. The pattern is consistent: binding energy is concentrated.
For de novo design, this means: if you specify hotspot residues and constrain the design model to contact them, you are directing the scaffold to engage the part of the target surface where the energy reward for binding is highest. Unconstrained generation, by contrast, may produce scaffolds that contact peripheral residues with low contribution to binding energy, generating sequences that look structurally plausible but fail in the binding assay.
In our campaigns we strongly prefer running multiple smaller AI design experiments against several pre-selected epitopes over one deep run on a single epitope. Programs are happy to spend twenty thousand designs converging on whichever surface they latched onto first, and that surface is not always the right one. The chance of finding a working binder scales with epitope diversity at the start of a campaign, not with depth on any single surface.
## How to identify hotspot residues computationally
When experimental alanine scanning data exist, they are the preferred source. Where they do not:
**Interface energy decomposition.** If a complex structure is available (or can be modeled with AlphaFold-Multimer / Boltz-2), per-residue interface energy can be decomposed using Rosetta's InterfaceAnalyzer or FoldX. Residues with large negative contributions to the interface score are candidate hotspots.
**Evolutionary conservation at functional sites.** Residues conserved across orthologs at a binding interface are under selection pressure that correlates with functional importance. ConSurf or custom evolutionary trace analysis can identify these positions.
**Literature-derived binding epitopes.** For targets with published structural biology of natural protein-protein complexes, the natural binding interface defines the epitope. Structures of ligand-receptor or protein-inhibitor complexes in the PDB are a direct source.
**Cryo-EM and HDX-MS data.** Hydrogen-deuterium exchange mass spectrometry maps solvent exposure changes on binding, identifying protected (interfacial) regions even without an atomic-resolution structure.
---
**Identify epitopes on your own target.** [Epitope Scout](https://scout.ranomics.com) scores and ranks surface patches on any PDB structure. Free to use.
---
## Specifying hotspots for RFdiffusion
In practice, hotspots are specified as a list of residue identifiers from the target structure. RFdiffusion uses these as geometric constraints during backbone generation: the produced scaffolds are required to place atoms within contact distance of the specified hotspot residues.
Practical guidance from production campaigns:
**Specify 3-8 hotspot residues.** Fewer than 3 gives insufficient constraint; the model will produce diverse backbones but many won't contact the intended region. More than roughly 10 overconstrains the problem and reduces the scaffold diversity you need for a successful screen.
**Prefer hotspots in the core of the epitope, not the periphery.** Edge residues at the interface boundary are often partially solvent-exposed and contribute less to binding energy. Central, buried hotspot residues are better anchors for a designed scaffold.
**Check structural accessibility.** A computationally predicted hotspot that is in a crystal packing contact, involved in an allosteric site, or at the base of a very deep groove may not be practically accessible to an external binder. Visual inspection of the target structure before finalizing the hotspot list is worthwhile.
**Avoid specifying hotspots in disordered or flexible loop regions.** RFdiffusion's conditioning on hotspot positions assumes those positions are well-defined in 3D space. High B-factor residues or regions with significant structural heterogeneity across crystal forms are poor hotspot anchors.
## Hotspot specification vs. hit rate: what the data show
In Ranomics' campaigns, hotspot-conditioned runs consistently outperform unconstrained generation in confirmed hit rate per screened sequence. The improvement is most pronounced for targets with complex surface topology where the productive epitope is a small fraction of the total accessible surface.
For flat or extended surfaces with multiple equally accessible regions, the gain is smaller. In those cases, unconstrained generation covers the productive epitope by chance at reasonable frequency.
The decision to invest in hotspot definition (experimental alanine scanning, or rigorous computational decomposition) should be scaled to the target difficulty and campaign budget. For a straightforward extracellular domain, an AlphaFold-Multimer-derived hotspot estimate is usually sufficient. For a challenging target where previous campaigns have failed, investing in experimental epitope definition before running a de novo design campaign is often the highest-value use of resources.
The rule of thumb from our work: rather than picking the single best-scoring hotspot region and running it deep, we pick the top three to four ranked patches and run smaller design batches against each. Epitope diversity at the start of a campaign has consistently bought us more confirmed binders than depth on any one surface.
---
**Ranomics uses hotspot-guided design in every binder campaign:** [AI Protein Binder Design](/ai-protein-binder-design)
{/* ranomics:related-services */}
## Related Ranomics services
- **[Epitope Scout](/technology/epitope-scout):** Free surface epitope identification for hotspot selection.
- **[AI Binder Sprint](/ai-binder-sprint):** Hotspot-conditioned de novo design with experimental validation.
---
## how-not-to-build-dataset-ai-protein-engineering.mdx
Source: https://ranomics.com/how-not-to-build-dataset-ai-protein-engineering/
import InOurWork from "../../components/InOurWork.astro";
The fastest way to ensure your AI protein engineering project produces useless results is to sabotage it from the start with a poorly constructed dataset.
## Rule 1: Embrace the Noise. Accuracy is a Guideline, Not a Rule
Skip biological replicates. Rely on single measurements. Average replicates to hide variability. And whatever you do, withhold raw data from your models.
If you actually want results: run biological replicates, report variance, and give models access to the full distribution of measurements. Single-point data without error bars is not data. It's a guess.
## Rule 2: Stick to What You Know. Diversity is Overrated
Include only successful variants. Exclude negative data. Limit mutations to familiar sequences.
If you actually want results: negative data is as informative as positive data. A model that only sees winners cannot learn what failure looks like. Include the full spectrum of functional outcomes, and diversify your sequence space beyond the comfortable neighborhood of known hits.
We are increasingly running NGS on both the binding and non-binding pools out of our yeast display campaigns. The foundational training datasets for protein design models skew heavily positive, and the negative pool from a sorted library is some of the cleanest failure-labelled data anyone has access to. The cost of the second sequencing arm is small relative to the model-training value of having the negatives explicitly labelled.
## Rule 3: Keep It Interesting. The Virtue of Inconsistency
Alter protocols between batches. Mix incompatible assays without labeling sources.
If you actually want results: standardize protocols across all experiments. Label every data point with its source assay, batch, and conditions. Batch effects are real, and unlabeled inconsistencies become invisible confounders that corrupt model training.
## Rule 4: Trust Your Gut. Process Data Aggressively
Normalize everything. Pool diverse variants into single measurements.
If you actually want results: minimal, transparent processing. Document every transformation applied to raw data. Avoid collapsing distinct measurements into summary statistics unless the model explicitly requires it.
## Conclusion: From Misdirection to Meaningful Results
Building a high-quality dataset for AI protein engineering requires deep expertise and rigorous execution. The experimental design, data collection, and curation steps matter as much as the model architecture. In our experience, this is the difference between a dataset that trains a useful classifier and one that trains a confident but wrong one. Partnering with an experienced team that understands both the biology and the machine learning is the most reliable path to datasets that actually work.
{/* ranomics:related-services */}
## Related Ranomics services
- **[Deep mutational scanning](/technology/deep-mutational-scanning):** Quality-controlled variant datasets built for ML training.
- **[Protein engineering services](/protein-engineering):** Dataset generation run against the target and readout your model needs.
---
## RFdiffusion in Practice: What Works and What Fails
> Operational lessons from running RFdiffusion binder design campaigns. Scaffold topology biases, hotspot conditioning tradeoffs, partial diffusion for scaffold grafting, failure modes on flat targets and membrane proteins, and how to avoid redundant candidate pools.
Source: https://ranomics.com/resource-hub/how-rfdiffusion-works-protein-designers-guide/
Published: 2025-02-10
import TryToolCallout from "../../components/TryToolCallout.astro";
Most explanations of RFdiffusion describe how diffusion models work in general: add noise to protein structures, train a model to reverse the process, condition on a target surface. That is correct but not operationally useful. If you are planning or evaluating an RFdiffusion campaign, what you need to know is where the method performs well, where it struggles, and which parameter choices actually affect outcomes.
This post covers what we have learned running RFdiffusion campaigns at Ranomics across dozens of targets.
## Scaffold topology bias: what the model generates and what it avoids
RFdiffusion was trained on structures from the Protein Data Bank. This means the model has strong priors toward topologies that are well-represented in the PDB: compact globular folds, alpha-helical bundles, mixed alpha/beta structures, and immunoglobulin-like domains.
In practice, this creates predictable patterns in what the model generates. For hotspot-conditioned binder design, the most common output topologies are three-helix bundles and four-helix bundles with short connecting loops. These are compact, thermodynamically stable folds that present well-defined binding surfaces, and they work well for many targets.
The limitation appears when the target geometry demands something different. If the epitope sits in a narrow groove, the model needs to generate an extended loop or beta-hairpin to reach it. RFdiffusion can produce these topologies, but at lower frequency and with lower structural confidence scores. You will need to increase the number of designs sampled (from 10,000 to 50,000 or more) to find enough non-helical solutions in the output pool.
Membrane protein targets present a specific challenge. RFdiffusion has no explicit membrane model. If the target epitope is near the membrane-proximal region, generated scaffolds may clash with the lipid bilayer in a physiological context even though they look structurally valid in isolation. Filtering against a membrane plane model after generation is necessary for these targets.
## Hotspot conditioning: the highest-leverage decision in any campaign
The residues you specify as hotspots constrain the entire downstream output. Getting this wrong is the most common reason campaigns fail to produce validated binders.
**Too few hotspots (1 to 2 residues)** gives the model too much freedom. Generated scaffolds contact the specified residues but the rest of the interface is unconstrained, leading to weak and geometrically variable binding modes. Hit rates after experimental screening are consistently low.
**Too many hotspots (more than 8 to 10 residues)** overconstrains the problem. The model struggles to find scaffold geometries that simultaneously contact all specified positions, and output diversity collapses. You end up with many near-identical designs that all fail or all succeed together.
**The productive range is 3 to 8 hotspot residues** for most targets. Within this range, the choice of which residues to specify matters more than the count. Prioritize residues that are solvent-accessible, structurally rigid (low B-factor or high pLDDT), and chemically diverse (mix of hydrophobic and polar contacts). Avoid specifying hotspots on flexible loops or disordered regions, as the model will generate scaffolds that contact the modeled position of those residues, which may not reflect their actual conformation.
---
**Identify epitopes on your own target.** [Epitope Scout](https://scout.ranomics.com) scores and ranks surface patches on any PDB structure. Free to use.
---
One pattern we see repeatedly: researchers specify hotspots based solely on biological importance (e.g., residues known to be critical for receptor-ligand binding) without checking structural accessibility. A residue can be biologically critical and geometrically buried. RFdiffusion will attempt to reach it, producing scaffolds with strained geometries that fail during structural validation.
## Partial diffusion: when to start from an existing scaffold
Standard RFdiffusion starts from pure noise and generates backbones from scratch. Partial diffusion starts from an existing structure, adds a controlled amount of noise, and then runs the reverse diffusion process conditioned on a new target.
This is useful in two specific scenarios.
**Scaffold grafting.** You have a validated binder scaffold against one target and want to adapt it to a related target with a different epitope. Partial diffusion preserves the global fold topology while allowing the interface region to be rebuilt. The noise level controls how much of the original structure is retained: low noise (10 to 20% of full diffusion steps) makes minor adjustments, high noise (60 to 80%) essentially redesigns the scaffold while keeping a topological bias toward the original.
**Topology steering.** When you want the output to have a specific fold class (e.g., beta-sheet scaffold for a target that requires a flat binding surface), you can seed with an exemplar structure of that topology. This biases the output distribution toward the desired fold type without rigidly constraining it.
Partial diffusion does not guarantee the output will retain the input fold. At high noise levels, the model may diverge entirely from the seed structure. Always verify that the output topology matches your intent before advancing candidates.
## Target-dependent failure modes
**Flat, featureless surfaces.** Some targets (certain cytokine receptors, viral capsid proteins, designed repeat proteins) present large, convex surfaces with minimal topographic features. RFdiffusion generates scaffolds that sit on these surfaces but lack the geometric complementarity that drives high-affinity binding. These campaigns require larger sampling (30,000 to 50,000 designs minimum) and aggressive post-design filtering, and they still tend to produce lower hit rates than campaigns against concave or grooved epitopes.
**Glycosylated surfaces.** RFdiffusion does not model glycans. If the target surface near the epitope is glycosylated in vivo, generated scaffolds may clash with glycan chains that are not represented in the input structure. Check the target for known glycosylation sites and exclude or flag epitopes that are partially occluded by glycans.
**Very large interfaces.** Designing scaffolds that span more than approximately 1,500 square angstroms of buried surface area pushes the model toward large, multi-domain architectures that are harder to express and fold. For targets requiring extensive interfaces, consider splitting the design into smaller binders that engage adjacent but non-overlapping epitopes.
## Avoiding redundant candidate pools
A subtle failure mode is generating thousands of designs that look structurally diverse but converge to a small number of unique sequences after ProteinMPNN. This happens when multiple distinct backbones present similar local geometries at the interface: ProteinMPNN assigns similar amino acid identities at the contact positions, and the resulting sequences cluster tightly.
Monitor sequence identity across the candidate pool before committing to synthesis. If more than 30 to 40% of candidates share greater than 80% sequence identity, the effective diversity of the pool is lower than the structural diversity suggests. In this case, either increase the scaffold length range to force more topological variation, reduce the number of hotspot constraints, or run ProteinMPNN at higher sampling temperatures (0.3 to 0.5) to increase sequence diversity at the cost of average predicted stability.
## The bottom line
RFdiffusion is a powerful backbone generator, but it is one step in a multi-step pipeline. Its outputs are only as good as the hotspot specification, and its failures are often invisible until experimental screening. Running large campaigns, filtering aggressively, and combining RFdiffusion with BindCraft and Boltzgen to cover different regions of structure space is the approach that consistently produces validated binders.
[See the full pipeline](/technology/ai-design-engine) | [Start a project](/ranomics-contact)
{/* ranomics:related-services */}
## Related Ranomics services
- **[AI protein binder design](/ai-protein-binder-design):** De novo binder design services using RFdiffusion, BindCraft, and experimental validation.
- **[RFdiffusion](/technology/rfdiffusion):** Production RFdiffusion pipelines for hotspot-conditioned binder design.
- **[AI Binder Sprint](/ai-binder-sprint):** Flagship 6–8 week program combining RFdiffusion with experimental validation.
---
## In Vivo Mutagenesis for AI Training Data
> How in vivo DNA mutagenesis systems like CRISPR-guided base editors and error-prone polymerases generate the large, unbiased protein variant datasets that machine learning models need. Practical comparison with synthetic library approaches for AI-driven protein engineering.
Source: https://ranomics.com/in-vivo-dna-mutagenesis-as-a-data-strategy-for-ai-driven-protein-engineering/
Published: 2026-01-30
Machine learning models for protein fitness prediction are only as good as the data they train on. The standard approach to generating training data is [deep mutational scanning](/technology/deep-mutational-scanning) with synthetic libraries: design a set of variants, synthesize them, screen them, and use the functional measurements as labeled training data.
This works. But it has a structural limitation that becomes visible at scale: synthetic libraries encode the biases of the designer.
NNK saturation mutagenesis overrepresents certain amino acid substitutions (leucine, serine, arginine each appear at 3/32 frequency while methionine and tryptophan appear at 1/32). Site-directed combinatorial libraries only sample positions the designer chose to diversify. Even "comprehensive" single-site saturation libraries cover only 19 substitutions per position, missing insertions, deletions, and multi-site epistatic combinations entirely.
For a predictive model, these biases mean the training data systematically underrepresents exactly the regions of sequence space where the model's predictions are most uncertain. The model learns well in the neighborhood of the wild-type sequence and the designer's hypotheses, but extrapolates poorly to novel combinations.
In vivo mutagenesis offers a fundamentally different data generation strategy.
## How in vivo mutagenesis systems work
In vivo mutagenesis introduces mutations directly into DNA inside living cells, bypassing the synthesis-cloning-transformation bottleneck entirely. Several systems are now mature enough for production use.
**CRISPR-guided base editors** (cytidine deaminases or adenine deaminases fused to catalytically inactive Cas9) introduce C-to-T or A-to-G transitions at guide RNA-specified loci. The editing window is typically 4 to 8 nucleotides wide, and multiple guides can be deployed simultaneously to diversify several regions of a gene in parallel. Editing rates of 20% to 60% per cell division are achievable depending on the editor and genomic context.
**Error-prone DNA polymerases** (such as PolI variants in the EvolvR system or orthogonal DNA polymerase-plasmid pairs) increase the local mutation rate by 10,000 to 100,000 fold relative to background. These systems operate continuously: every cell division introduces new mutations, and the diversity of the population increases over time without any manual intervention.
**Retroelement-based systems** (T7-based retrotranscription mutagenesis) continuously mutate a target gene by repeatedly reverse-transcribing it through an error-prone intermediate. Mutation rates of 10^-3 to 10^-4 per base per generation are typical.
The practical consequence is that a single flask culture, grown for 20 to 50 generations, can accumulate millions of unique variants without a single cloning step.
## Why stochastic diversity is better for ML training data
The value proposition of in vivo mutagenesis for AI is specific: it generates data that is less biased than synthetic libraries across three dimensions that matter for model generalization.
**Substitution uniformity.** Error-prone polymerases and base editors do not follow the codon table biases of NNK. The mutation spectrum depends on the enzyme, not on codon degeneracy. This means amino acid substitutions that are rare in NNK libraries (tryptophan, methionine, cysteine) appear at frequencies closer to other residues, reducing the blind spots in the training data.
**Multi-site combinations.** After 30 to 50 generations of continuous mutagenesis, a significant fraction of the population carries two, three, or more mutations simultaneously. These multi-mutant combinations are the exact data points needed to train models that predict epistatic interactions, which are the interactions between mutations at different positions that cannot be predicted from single-mutant effects alone.
**Neutral and deleterious variant coverage.** Synthetic libraries are often designed around positions expected to be important. In vivo mutagenesis does not make this distinction. It mutates conserved and variable positions alike, generating a balanced representation of functional, neutral, and deleterious variants. Models trained on this distribution learn where the fitness boundaries are, not just where the peaks are.
## Practical considerations
In vivo mutagenesis is not a replacement for synthetic libraries in all contexts. The tradeoffs are real.
**Loss of positional control.** You cannot specify which positions to mutate with the same precision as site-directed mutagenesis. CRISPR base editors offer partial control (guide RNA targeting), but error-prone polymerase systems mutate the entire target region stochastically.
**Mutation spectrum bias.** Every mutagenesis system has its own spectrum. Base editors are restricted to transition mutations (C-to-T or A-to-G). Error-prone polymerases tend to favor transitions over transversions. No single system covers all 19 possible substitutions at every position uniformly. Combining multiple systems or correcting for known biases computationally is often necessary.
**Fitness selection during growth.** Because mutagenesis occurs in living cells, deleterious variants are depleted during growth. This is a form of unintentional selection that biases the variant distribution toward functional sequences. For some applications this is useful (it enriches for expressible variants). For others it introduces a confound that must be accounted for in the training data.
**Sequencing depth requirements.** The diversity generated by continuous mutagenesis can exceed 10^6 to 10^7 unique variants per experiment. Achieving adequate sequencing coverage requires deep NGS (typically 10x to 50x coverage per variant), which is feasible but adds cost relative to smaller synthetic libraries.
## When to use each approach
The decision is straightforward once you frame library construction as a data generation problem.
**Use synthetic libraries** when you need precise control over which positions are diversified, when you are testing specific hypotheses about structure-function relationships, or when the target region is small enough that saturation mutagenesis achieves complete coverage (fewer than 5 to 6 positions simultaneously).
**Use in vivo mutagenesis** when the goal is broad coverage of sequence space for model training, when epistatic interactions are important, when you need multi-mutant combinations, or when you want to run continuous evolution experiments that accumulate diversity over many generations.
**Use both** when the project requires an initial hypothesis-driven screen (synthetic) followed by a broad exploration phase (in vivo) to expand the training data for a second-generation model.
At Ranomics, we treat library generation as a data strategy decision. The molecular biology is a means to an end. The end is a dataset that trains a model capable of predicting variant fitness accurately enough to guide the next round of design.
[Start a project](/ranomics-contact)
{/* ranomics:related-services */}
## Related Ranomics services
- **[Directed evolution](/directed-evolution-protein-engineering):** In vivo mutagenesis campaigns generating unbiased ML training data.
- **[Protein engineering services](/protein-engineering):** Mutagenesis strategy matched to the downstream model and selection.
---
## Industrial Enzyme Engineering: A CRO Playbook for Biocatalysis
> How CROs engineer enzymes for industrial biocatalysis: substrate scope, organic solvent tolerance, regiospecificity, and process compatibility.
Source: https://ranomics.com/industrial-enzyme-engineering-cro/
Published: 2026-05-11
Industrial enzyme engineering sits at the intersection of protein engineering, process chemistry, and large-scale manufacturing. The customers are specialty chemical producers, pharmaceutical intermediates manufacturers, agrochemical companies, food and detergent ingredient makers. The objectives are concrete and quantitative — turnover number at the operating temperature, substrate scope at the relevant concentration, half-life in the process solvent, regiospecificity that meets product purity targets.
This article is a playbook for what an industrial enzyme engineering project actually looks like when delivered by a CRO. The methods are mature; the integration is the value.
## The industrial enzyme value chain
Industrial biocatalysis is now embedded across multiple sectors:
- **Pharmaceutical intermediates**: transaminases for chiral amine synthesis, ketoreductases for alcohol synthesis, lipases for kinetic resolution, monooxygenases for selective hydroxylation.
- **Specialty chemicals**: nitrilases, esterases, lyases for chiral building blocks. Polymerases and glycosyltransferases for bioplastics monomers.
- **Agrochemicals**: enzymes for crop protection compound synthesis. Glufosinate, phosphinothricin, and adjacent chiral chemistries.
- **Food and feed**: phytases, proteases, amylases, lipases for animal nutrition and food processing.
- **Detergents**: high-temperature stable proteases and lipases for cleaning formulations.
- **Biofuels and bioplastics**: cellulases, lipases, polyhydroxyalkanoate synthases.
Each sector has different performance bars and regulatory contexts, but the engineering work follows the same playbook.
## Common targets — what gets engineered
The repeat targets across industries:
**Transaminases** for chiral amine synthesis. Engineering objectives: substrate scope (accepting bulky α-keto acids), stability at the process pH, activity in organic cosolvents (DMSO, isopropanol).
**Lipases** for kinetic resolution. Engineering objectives: enantioselectivity (E > 100), regiospecificity for sn-1 vs sn-3 positions in triglycerides, activity in non-aqueous solvents.
**Ketoreductases** for alcohol synthesis. Engineering objectives: substrate scope, cofactor preference (NADH vs NADPH), thermostability at operating temperature.
**Oxidoreductases and monooxygenases** for selective oxidation. Engineering objectives: substrate scope, peroxide tolerance, regiospecificity, total turnover number before inactivation.
**Hydrolases and esterases** for ester synthesis and hydrolysis. Engineering objectives: substrate scope, pH stability, product inhibition tolerance.
**Glycosyltransferases** for complex sugar synthesis. Engineering objectives: donor and acceptor substrate scope, regiospecificity, expression yield.
For each enzyme class, a mature literature exists on which positions tolerate mutation and which engineering approaches deliver. The CRO advantage is integrating that literature with high-throughput experimental work.
## Engineering objectives
The dimensions an industrial enzyme campaign optimizes:
**Substrate scope.** The wild-type enzyme catalyzes a natural substrate. The industrial application needs a structurally different (often bulkier, sometimes charged) substrate. Engineering targets are usually the active-site loops that gate substrate access.
**Solvent tolerance.** Industrial processes often run in cosolvent (DMSO, methanol, isopropanol) or partial-aqueous systems to solubilize substrates. Wild-type enzymes denature in cosolvent; engineered variants tolerate 30–60% organic without significant activity loss.
**Regiospecificity and stereoselectivity.** The enzyme must produce the desired isomer at >99% selectivity for product purity. Engineering targets are typically the residues that orient the substrate in the active site.
**Thermostability.** Covered in detail in [enzyme thermostability engineering](/enzyme-thermostability-engineering). Tm shifts of +10 to +20 °C are routine; higher gains require sustained iteration.
**Product inhibition tolerance.** Many enzymes are inhibited by their own product, capping conversion at low concentrations. Engineering targets are the product-binding subsite, expanded or destabilized to lower product affinity.
**Half-life in the process.** The integrated measure that customers care about. Combines thermostability, solvent tolerance, oxidation resistance, and aggregation propensity into one number: how long the enzyme stays active in the actual process.
## Workflow — DMS, focused mutagenesis, AI design, scale-up
The CRO workflow chains four phases:
**Phase 1 — Discovery and characterization (4–8 weeks).** Express the wild-type. Validate activity on the customer's substrate. Establish baseline performance metrics. Map the structural and sequence landscape — homologs, structural model, predicted hot spots.
**Phase 2 — Library design and screening (8–16 weeks).** Build the focused library (DMS-scale or directed-evolution-scale, depending on screening capacity). Screen against the relevant performance dimensions. Generate the position-by-position fitness map. Identify the top single-mutant hits. We covered DMS methodology in [DMS fitness landscapes](/deep-mutational-scanning-a-high-throughput-approach-to-mapping-protein-fitness-landscapes) and AI integration in [AI and DMS for enzyme engineering](/leveraging-ai-and-deep-mutational-scanning-to-engineer-enzymes).
**Phase 3 — Combinatorial optimization (8–12 weeks).** Combine top single mutants. Test for positive epistasis. Iterate as needed. Output: a sequence-defined evolved variant meeting the performance specification.
**Phase 4 — Scale-up and delivery (4–8 weeks).** Express the evolved variant in the customer's preferred production host (typically *E. coli* or *Pichia pastoris* for industrial enzymes, sometimes *Bacillus* for proteases and amylases). Characterize at scale. Deliver the construct plus the engineering history documentation.
Total timeline: 6–11 months. Total cost is dominated by Phase 2 screening; Phases 1, 3, and 4 are predictable.
## The integrated approach
The CRO advantage is in the integration. Most academic and in-house programs can run one of the methods well — directed evolution, or DMS, or computational design. A CRO running industrial campaigns combines all of them in the right sequence:
- **Consensus design + Rosetta ddG** as the first-pass focused library, before any high-throughput experiments.
- **DMS** to map the tolerable mutations exhaustively, in one library.
- **Directed evolution** to iterate on the DMS-revealed top hits and find epistatic combinations the single-mutant scan misses.
- **AI tools** (ProteinMPNN-driven sequence sampling, ESM-derived fitness predictions) to focus combinatorial libraries.
- **Computational scaffold design** when none of the natural homologs is a good starting point — rare in industrial work but increasingly relevant for new chemistries.
The methods compound. A 16-month all-classical-directed-evolution campaign can be compressed to 9 months by starting with a consensus + DMS pass. The reverse is also true: AI-only approaches without wet-lab iteration rarely deliver industrial-grade variants.
We covered the rational design vs evolution comparison in [rational enzyme engineering](/rational-enzyme-engineering) and the directed evolution side in [the directed evolution technical guide](/a-technical-guide-to-directed-evolution-for-enhancing-protein-stability-and-function).
## When to use a CRO vs build in-house
The economic case for a CRO:
- **One-time program**: building enzyme engineering infrastructure for a single product doesn't amortize. Use a CRO.
- **Multiple programs but small team**: CROs handle the labor-intensive screening and engineering work; the customer keeps the strategic decisions and IP.
- **Specialized capability requirement**: DMS infrastructure, high-throughput automation, AI-design pipelines are expensive to build and maintain. CROs amortize them.
- **Speed**: dedicated CRO teams hit a 6–12 month timeline that in-house teams typically can't match without dedicated headcount.
The case against:
- **Strategic IP**: if the enzyme is the company's core IP and engineering trajectory must stay internal, in-house is the right answer.
- **Continuous iteration**: programs that need ongoing engineering across many years often justify in-house investment.
For most industrial enzyme programs — especially in mid-cap chemical and pharma companies — the CRO route delivers faster and cheaper than building.
## Decision summary
For a new industrial enzyme engineering program:
1. Define the performance specification precisely (numbers, conditions, units).
2. Characterize the wild-type. Confirm what already works and what doesn't.
3. Choose the engineering method based on screening infrastructure: DMS-led if high-throughput is available, directed evolution-led if not, AI-augmented in both cases.
4. Run the campaign in three phases (discovery → library → combinatorial), with checkpoint reviews at each phase boundary.
5. Plan scale-up early. The construct that wins the engineering campaign is not always the one that expresses well at scale; co-validate yield during the engineering work.
---
If you're scoping an industrial enzyme engineering program, see our [enzyme engineering services](/applications/enzyme-engineering) or reach out via the [contact page](/ranomics-contact). For multi-target or multi-objective programs, see [protein engineering services](/protein-engineering).
{/* ranomics:related-services */}
## Related Ranomics services
- **[Enzyme engineering](/applications/enzyme-engineering):** Industrial biocatalysis enzyme engineering campaigns.
- **[Directed evolution](/directed-evolution-protein-engineering):** The iterative method underlying most industrial enzyme campaigns.
- **[Protein engineering](/protein-engineering):** Full protein engineering portfolio including non-enzyme targets.
---
## Introduction to Protein Developability: What Makes a Good Biologic Drug?
> A biologic with high potency is only half the battle. Many promising candidates fail due to poor developability and manufacturability. The four pillars of protein developability explained.
Source: https://ranomics.com/introduction-to-protein-developability-what-makes-a-good-biologic-drug/
Published: 2025-09-29
High potency alone does not make a viable biologic drug. Many promising candidates fail in development due to poor manufacturability, instability, aggregation, or chemical liabilities. Integrating developability assessment from the earliest stages of discovery is essential.
## Pillar 1: Manufacturability and Expression
The candidate must be producible at viable commercial titers in standard expression systems like CHO cells. Target titers of >1 g/L are the benchmark for commercial feasibility.
## Pillar 2: Stability (Thermodynamic and Colloidal)
Thermodynamic stability is assessed via melting temperature (Tm) using differential scanning calorimetry (DSC) or differential scanning fluorimetry (DSF). Colloidal stability (the propensity for protein-protein interactions in solution) is measured by dynamic light scattering (DLS) or self-interaction nanoparticle spectroscopy (SINS).
## Pillar 3: High Solubility and Low Aggregation Propensity
Surface hydrophobicity analysis identifies patches that drive aggregation. Size-exclusion chromatography (SEC) testing quantifies the fraction of monomer vs. aggregate species. High monomer percentage at target formulation concentrations is essential.
## Pillar 4: Chemical Stability and Lack of Post-Translational "Hotspots"
Sequence liabilities that drive chemical degradation must be identified and eliminated early:
- **Deamidation:** Asparagine residues, particularly at N-G and N-S motifs
- **Isomerization:** Aspartate residues converting to isoaspartate
- **Oxidation:** Methionine and tryptophan residues
- **Unpaired cysteines:** Leading to disulfide scrambling and aggregation
- **Unwanted glycosylation sites:** N-X-S/T sequons in binding regions
## Conclusion: Integrating Developability from Day One
The most efficient path to a viable biologic is to assess developability in parallel with affinity and specificity screening, not as a late-stage gate. Surface display platforms enable simultaneous selection for function and expression, making early developability assessment practical at library scale.
{/* ranomics:related-services */}
## Related Ranomics services
- **[Antibody engineering](/applications/antibody-engineering):** Developability screening integrated into antibody discovery.
- **[Mammalian display](/technology/mammalian-display):** Developability-focused screening on a mammalian display platform.
---
## Leveraging AI and Deep Mutational Scanning to Engineer Novel Enzymes
> A comprehensive guide to combining deep mutational scanning with machine learning for enzyme engineering, covering variant library generation, functional selection, data analysis, and the iterative AI-DMS cycle.
Source: https://ranomics.com/leveraging-ai-and-deep-mutational-scanning-to-engineer-novel-enzymes/
Published: 2025-07-15
import InOurWork from "../../components/InOurWork.astro";
Deep Mutational Scanning (DMS) is a high-throughput functional genomics method that combines massively parallel mutagenesis with functional selection and deep sequencing to systematically measure the effects of thousands, or even millions, of protein variants in a single experiment.
## Generation of Variant Libraries
Three primary methods for generating variant libraries:
**Error-Prone PCR:** Introduces random mutations across the entire gene. Simple to implement but offers limited control over mutation type and position.
**Oligonucleotide-based Methods:** Using NNN, NNS, or NNK codon schemes, or more sophisticated approaches like the 22-codon group, these methods enable precise control over which positions are diversified and which amino acids are sampled.
**CRISPR-based Endogenous Mutagenesis:** Emerging methods that introduce mutations directly in vivo, enabling continuous diversification without ex vivo library construction.
## The Functional Selection Assay
Two major assay types dominate DMS experiments:
**Fitness/Survival-Based Assays:** Variants compete under growth selection, and functional variants outgrow non-functional ones. Simple but limited to phenotypes linked to growth.
**Binding and Stability Assays:** Using display technologies (yeast, phage, mammalian), variants are sorted based on binding or expression levels via FACS.
A persistent challenge in interpreting DMS data is "biophysical ambiguity": a single functional score often conflates a mutation's effect on protein stability and abundance with its effect on a specific activity like binding.
## Deep Sequencing and Variant Quantification
Robust quantification requires Unique Molecular Identifiers (UMIs), biological replicates, and sufficient sequencing depth on Illumina platforms. Emerging long-read technologies (PacBio CCS, Oxford Nanopore) are expanding the scope of DMS to full-length protein variants.
## Data Analysis: From Raw Counts to Functional Scores
The analytical pipeline involves calculating enrichment ratios, applying normalization strategies, and using statistical modeling tools (dms_tools, Enrich2) to generate sequence-function heatmaps that visualize the fitness landscape.
## Applications
1. **Protein Engineering and Optimization,** exemplified by anti-CRISPR protein engineering
2. **Unraveling Allosteric Regulation,** demonstrated in bacterial allosteric transcription factor studies
3. **Designing Novel Therapeutics: Antimicrobial Peptides,** the dmSLAY technique applied to Protegrin-1
4. **Plant Science,** engineering the AtSWEET13 transporter
## The Computational Symbiosis: Integrating Machine Learning
DMS datasets are natural training data for machine learning models:
- **Data Preprocessing:** Normalization and one-hot encoding of sequence data
- **Supervised Models:** CNNs and RNNs/LSTMs for predicting variant fitness from sequence
- **Generative Models:** VAEs and GANs for proposing novel sequences with desired properties
The language analogy is useful: amino acids are the alphabet, motifs are words, and full sequences are sentences. Protein language models learn this grammar from large corpora of natural and engineered sequences.
In our enzyme work, we have consistently found that custom models trained on bespoke deep-mutational-scanning fitness maps outperform off-the-shelf open-source generators. The training data in the public structural databases is too sparse on enzyme fitness for general-purpose models to do useful work. The path that has worked for us is to generate the DMS data first, train a small model on the resulting sequence-to-fitness map, then sample stabilizing or activity-improving combinations that pure random sampling would not have reached in the same library size.
## The Iterative Cycle in Practice
**Engineering AAV Gene Therapy Vectors:** In vivo fitness landscapes from DMS feed ML models that propose novel capsid variants, which are then validated experimentally, closing the design-build-test-learn loop.
**Designing Selective Antimicrobial Peptides:** The dmSLAY technique generates DMS data on antimicrobial activity, and ML models learn selectivity rules to propose safer peptide therapeutics.
## A Critical Perspective
**Advantages:** Unmatched scale and throughput, unbiased discovery of beneficial mutations, and high per-variant efficiency.
**Limitations:** The functional assay design remains the primary bottleneck. Scalability depends on biological context. Biophysical ambiguity conflates stability and function. Data noise requires careful experimental design and statistical treatment.
In our work, the biophysical-ambiguity failure mode (a single fitness score conflating stability with activity) is the single biggest source of false positives in enzyme campaigns. We design assays that decouple the two when the experimental setup allows, or we run paired assays so the deconvolution can happen at the analysis step. Skipping that step is the most common reason a "promising" DMS hit fails downstream characterization.
## The Future Promise
Multi-phenotype assays that deconvolve stability from function, whole-genome scanning approaches, and comprehensive functional atlases of protein variation will define the next generation of DMS-ML integration.
{/* ranomics:related-services */}
## Related Ranomics services
- **[Enzyme engineering](/applications/enzyme-engineering):** AI + DMS cycles for activity, stability, and substrate scope.
- **[Deep mutational scanning](/technology/deep-mutational-scanning):** Training-quality variant datasets for ML-guided enzyme design.
---
## Library Size Limitations in Yeast Display: Maximizing Diversity
> Discover proven strategies to maximize yeast display library diversity despite transformation limitations, including optimized protocols, Golden Gate cloning, and smart library design.
Source: https://ranomics.com/library-size-limitations-in-yeast-display-advanced-strategies-for-maximizing-diversity/
Published: 2025-08-11
## The Library Size Challenge
Yeast display libraries typically achieve diversities in the range of 10^7 to 10^9 unique variants, representing a significant constraint compared to phage display systems that routinely generate libraries containing 10^11 to 10^12 variants.
Even under optimal conditions, transformation efficiencies rarely exceed 10^6 to 10^7 transformants per microgram of DNA.
For antibody engineering applications, where researchers typically optimize both heavy and light chain variable regions simultaneously, the theoretical sequence space can exceed 10^20 possible combinations.
## Library Construction Strategies
### Optimized Transformation Protocols
Achieving maximum transformation efficiency requires mid-logarithmic phase cells (OD600 0.6-0.8) and specialized electroporation protocols exceeding 10^7 transformants per microgram. Voltage, capacitance, resistance, and pulse duration all require optimization for each strain and construct.
### Golden Gate Cloning for Enhanced Library Construction
Golden Gate assembly enables simultaneous multi-fragment assembly using type IIS restriction enzymes. Key requirements include equimolar fragment ratios and high-fidelity polymerases for amplicon generation.
### Sequential Enrichment Strategies
When a single library cannot span the required diversity, multiple smaller libraries can be screened independently. Computational design guides the diversification strategy for each sub-library. Hits from independent screens can be recombined via DNA shuffling or overlap extension PCR.
### Smart Library Design Approaches
Structure-based design, sequence analysis, and machine learning predictions focus diversification on key binding regions rather than uniformly randomizing all positions. This reduces the effective library size needed to cover functionally relevant sequence space.
## Quality Control and Library Validation
### Analytical Assessment of Library Composition
NGS analysis of the pre-selection library quantifies mutation distribution, identifies synthesis biases, and confirms that the designed diversity is actually present. Functional testing of random clones and inclusion of negative controls validate the selection assay.
### Expression and Display Monitoring
Flow cytometry quantification of display levels, dual-labeling for simultaneous expression and binding assessment, and temporal monitoring of display stability across growth phases ensure that library quality is maintained through the screening campaign.
{/* ranomics:related-services */}
## Related Ranomics services
- **[Variant library construction](/technology/variant-library-construction):** High-diversity library design, synthesis, and transformation at scale.
- **[Yeast surface display services](/yeast-display):** Screening campaigns against libraries sized for the right target.
---
## Mammalian Cell Display: When CHO and HEK293 Outperform Yeast
> CHO and HEK293 mammalian display platforms preserve PTMs and disulfide bonds that yeast display cannot. When to use which.
Source: https://ranomics.com/mammalian-cell-display-cho-hek293/
Published: 2026-05-11
Mammalian cell display is the platform of last resort and the platform of first choice. Last resort because it is slower, more expensive, and lower-throughput than yeast or phage. First choice when the target's biology requires it — when the epitope depends on complex glycans, when the antibody format is full-length IgG, when developability has to match production. This article is the decision framework for both cases.
## Why mammalian display exists
Yeast display can install eukaryotic disulfide bonds and chaperone-assisted folding. Phage display can present 10^11 variants at once. Neither can replicate the post-translational modification machinery of mammalian cells — and for therapeutic targets whose epitope or function depends on those PTMs, neither platform delivers leads that translate to production.
The specific gaps that mammalian display closes:
- **Complex N-glycosylation.** Yeast installs high-mannose-only glycans. Mammalian cells trim mannose and add complex sugars (galactose, sialic acid, fucose, N-acetylglucosamine). For antibody Fc engineering, ADCC-relevant constructs, glycan-specific binders, and glycoprotein targets, this matters.
- **Full-length IgG display.** Four-chain assembly with proper inter-chain disulfides and Fc-mediated effector function is a mammalian-only proposition at scale.
- **PTM-dependent epitopes.** Phosphorylation, sulfation, sialylation, O-linked glycosylation. Targets bearing these modifications are routinely missed by yeast-displayed libraries that don't see the same modifications.
- **Cell-surface-context epitopes.** Some membrane proteins fold and oligomerize correctly only in mammalian membranes. Antibody discovery against these targets requires presenting the antigen on mammalian cells and ideally selecting the antibody library on mammalian cells too.
## CHO vs HEK293 — biological differences
The two workhorse mammalian hosts differ in ways that matter for display.
**HEK293** (human embryonic kidney) transfects with very high efficiency, grows rapidly in suspension, and produces high yields per cell. It is the right host for early-stage library work — fast cycle times, less infrastructure burden. The glycosylation profile leans toward sialylated complex glycans typical of human cells. Downside: HEK293 is rarely the production host for therapeutic biologics, so leads validated in HEK293 still need a CHO confirmation step before scale-up.
**CHO** (Chinese hamster ovary) is the regulatory-favored production host. Glycosylation has its own profile — less sialylation than human, characteristic core fucosylation, slightly different α-galactose content. For late-stage developability and process-development-grade lead selection, CHO display is the closest analog to what production manufacturing will see. Trade-off: lower transfection efficiency, slower growth, more infrastructure.
The decision rule: HEK293 for fast discovery and library work, CHO for late-stage developability validation. Many programs use both — HEK293 for the initial sort rounds, CHO for the final lead-selection round.
## Library construction in mammalian cells
Three integration methods are in routine use, each with trade-offs.
**Lentiviral transduction.** Mature, broadly used, integrates randomly. Library complexity scales with viral titer and transduction efficiency. Random integration means clones differ in expression level due to position effects, which adds noise to the display-level normalization. Workaround: titrate carefully and gate on display level.
**PiggyBac transposase.** Higher copy numbers per cell, larger cargo capacity than lentivirus. Good for large constructs. The trade-off is multi-copy integration: each cell may carry 2–10 copies of the displayed variant, which complicates downstream NGS deconvolution.
**Sleeping Beauty transposase.** Site-preferred integration relative to PiggyBac, simpler safety profile. Library complexity caps somewhat lower than the alternatives but the integrations are cleaner. Good default for libraries under 10^7.
For most discovery campaigns we run lentiviral if the library is small (10^6–10^7) and Sleeping Beauty if cleaner integration is the priority. PiggyBac for special cases where cargo size or copy number is the driver.
## Throughput limits
Mammalian display caps at meaningfully lower throughput than yeast at every stage.
| Stage | Yeast | Mammalian |
|---|---|---|
| Library size | 10^7–10^9 | 10^6–10^8 |
| Cells per FACS hour | 10^7–10^8 | 10^6–10^7 |
| Sort rounds | 3–4 | 2–3 |
| Wall-clock per round | 3–5 days | 5–7 days |
| Cost per round | 1× | 3–5× |
The 3–5× cost differential per round drives the workflow design. We use mammalian display where its specific advantages pay off and yeast everywhere else — typically yeast for discovery breadth, mammalian for developability validation. We documented this in [the two-platform approach](/the-two-platform-approach-using-yeast-display-for-affinity-and-mammalian-display-for-developability).
## When to switch from yeast to mammalian mid-campaign
A common decision point: discovery on yeast yields a list of binders, but the program targets clinical development. Should the next step be mammalian display or direct biochemical characterization?
We switch to mammalian display when:
- The target is a glycoprotein and yeast-displayed binders systematically fail to recapture the relevant epitope. Membrane-bound complement proteins are a typical example.
- The candidate count is 10–50 and the program needs developability triage before committing to expression scale-up. Mammalian display under titrated stringency reveals which clones express, fold, and bind under production-relevant conditions; this catches issues that biochemical characterization on individual clones is too slow to find.
- The antibody format must be full-length IgG. Yeast can't do this efficiently; mammalian can.
We stay on yeast (and proceed direct to biochemical characterization) when:
- The format is scFv or VHH and the target is a soluble antigen without complex PTM dependencies.
- The candidate count is below 10 — biochemical characterization is faster than setting up a mammalian display round for a handful of clones.
## Practical workflow
For a typical mammalian display arm of a discovery campaign:
1. **Plasmid library construction**: gene synthesis or pool synthesis from yeast hits, cloned into PiggyBac/Sleeping Beauty/lenti destination vector with display tag.
2. **Transfection or transduction**: 10× excess plasmid per cell, 24-hour expression window.
3. **Antibiotic selection** (if needed): 5–7 days to remove non-integrated cells.
4. **Display level QC**: stain with anti-tag antibody, confirm >50% display-positive cells.
5. **FACS round 1**: stain with antigen at moderate concentration (1–10 nM); sort top 1% double-positive.
6. **Expansion and re-display**: 5–7 days.
7. **FACS round 2**: tighten antigen concentration; sort top 0.5%.
8. **NGS or single-clone isolation**: depending on downstream workflow.
Total wall-clock: 4–6 weeks. Total cost (reagents + sort time): typically $25K–$60K per arm for a small library.
## Decision summary
Use mammalian display when the target requires it (glycosylation, full-length IgG, PTM-dependent epitope) or when the program stage requires it (late-stage developability, production-relevant validation).
Use HEK293 for fast iterative library work; switch to CHO for late-stage lead selection that must match production.
Don't use mammalian display when the target is a soluble protein with no PTM dependence and the antibody format is scFv or VHH — yeast is 3–5× cheaper for the same answer.
---
If you're scoping a campaign and want a second opinion on whether mammalian display is necessary, see our [mammalian display services](/technology/mammalian-display) or [start a Binder Pilot](/binder-pilot). For multi-target programs with developability validation, see the [AI Binder Sprint](/ai-binder-sprint).
{/* ranomics:related-services */}
## Related Ranomics services
- **[Mammalian display](/technology/mammalian-display):** CHO and HEK293 display platforms for PTM-dependent targets.
- **[Yeast surface display](/yeast-display):** Faster, cheaper discovery platform that pairs with mammalian for validation.
- **[AI Binder Sprint](/ai-binder-sprint):** Flagship program using both yeast and mammalian display in sequence.
---
## Mammalian Cell Engineering: From Stable Line to Functional Assay
> Mammalian cell engineering services for stable cell lines, knockouts, knock-ins, reporter cells, and high-throughput functional assays.
Source: https://ranomics.com/mammalian-cell-engineering/
Published: 2026-05-11
Mammalian cell engineering is the substrate underneath many of the downstream biology services that pharma and biotech rely on: stable production lines, knockout disease models, reporter cells for high-throughput screening, glycoengineering for biologics manufacturing. The methods are mature, the throughput is the constraint, and the choice of host plus integration strategy determines whether the engineered line is fit for purpose or a costly distraction.
This article is an overview of what mammalian cell engineering delivers and how the workflow decisions cascade — host choice, integration method, validation strategy, downstream coupling.
## The cell engineering value chain
Mammalian cell engineering serves four broad use cases:
1. **Stable cell line development** — for biologics production, the engineered line produces the target biologic at scale-compatible yields with characterized glycosylation and clonal stability.
2. **Disease model development** — knockout or knock-in lines that recapitulate human disease genotypes, used for drug discovery and mechanism studies.
3. **Reporter cells for HTS** — engineered to express a reporter (luciferase, fluorescent protein, calcium indicator) in response to pathway activation, enabling high-throughput screening of compounds or biologics.
4. **Functional assay platforms** — engineered with the relevant target receptor, downstream signaling components, and reporter, enabling on-target functional characterization of antibody or small-molecule leads.
Each use case has different requirements, but the underlying methods overlap heavily. The same CHO or HEK293 chassis, same integration tools, same validation steps.
## Stable cell lines — CHO, HEK293, primary
**CHO** (Chinese hamster ovary, multiple sublines: CHO-K1, CHO-S, CHO-GS) is the dominant therapeutic biologics production host. Decades of regulatory familiarity, established large-scale fermentation, characterized glycosylation patterns. For any program heading to clinical manufacturing, CHO is the default. Stable line development from a plasmid takes 4–6 weeks for a pooled line, 4–6 months for a clonal high-producer.
**HEK293** (and its derivatives HEK293F, HEK293T, Expi293) is the workhorse for research-grade expression and rapid prototyping. Higher transfection efficiency than CHO, faster growth, simpler suspension culture. Sialylation patterns are closer to native human than CHO. For small-scale antibody expression, viral vector production, and early-stage research, HEK293 is faster and cheaper.
**Primary cells** (donor T cells, iPSC-derived cardiomyocytes, primary hepatocytes) are used when the biology requires native cellular context. Engineering primary cells is harder — limited proliferative capacity, donor variability, lower transfection efficiency — but for some functional assays (cytotoxicity, calcium handling, drug metabolism), nothing else suffices.
## Genome editing methods
The CRISPR toolset is the standard. Three classes of edits cover most programs:
**Indel-generating knockouts** use Cas9 with a single guide RNA to cut a target locus; non-homologous end joining repairs the break with insertion or deletion events that disrupt the open reading frame. Efficiency is high (50–95% of cells edited), but the indel population is heterogeneous. Clonal isolation and sequencing confirm the specific edit.
**Knock-ins** use Cas9 plus a donor template; homology-directed repair installs the donor at the cut site. Efficiency is much lower (typically 1–10% of cells edited), and HDR-competent cells are mostly in S/G2 phase. Knock-in efficiency depends on cell type, donor design, and cut-site choice; some loci are easier than others.
**Base editors and prime editors** install single-base changes or short insertions/deletions without double-strand breaks. Lower throughput and narrower target site requirements than standard CRISPR, but cleaner — no risk of inversions or large deletions. Useful for installing specific point mutations (disease alleles, codon-optimized variants).
For each method, off-target activity must be characterized — usually by GUIDE-seq, CIRCLE-seq, or amplicon sequencing of predicted off-target sites. For therapeutic-relevant lines, this is non-negotiable.
## Reporter cells for high-content assays
Reporter cell lines are engineered to convert a biological signal into a detectable readout. The standard readouts:
- **Luciferase** (firefly or Renilla) responding to transcription factor activation. Throughput-friendly via plate-reader luminescence. Dynamic range 100–1000× over baseline. Good for NF-κB, STAT, MAP kinase pathways.
- **Fluorescent protein** (GFP, mCherry, tdTomato) under inducible promoter or fused to the protein of interest. Lower dynamic range than luciferase but enables live-cell imaging and flow cytometry.
- **Calcium indicators** (GCaMP variants, Fluo-4) for GPCR or ion channel signaling. Sub-second temporal resolution, single-cell quantitation.
- **Bioluminescent or fluorescent caspase reporters** for apoptosis-readout assays.
Reporter design depends on the pathway, the desired temporal resolution, and the assay throughput. For high-throughput screening (1,000s of compounds per day), luciferase is the default; for mechanistic studies, calcium or fluorescence imaging.
## Functional assay design
A functional assay couples a target receptor (or pathway) in an engineered cell to a reporter that quantifies the biology of interest. Examples:
- **Antibody-mediated receptor activation/inhibition**: engineered reporter cell expressing target receptor + downstream pathway reporter (NF-κB-luciferase, calcium-GCaMP). Add antibody, read out activation or inhibition.
- **ADCC and ADCP**: engineered effector cells (NK or macrophage) with knock-in FcγR variants, plus target cells expressing the antibody's antigen. Co-culture; readout: target cell lysis or phagocytosis.
- **Pathway crosstalk**: engineered cell with dual reporters for two pathways. Reveals on- and off-target activity of candidate biologics.
Designing the functional assay is half the engineering challenge. The reporter must be on-target, the dynamic range must be sufficient, the time course must capture the relevant biology, and the assay must be robust across replicates and batches.
## Pairing with display screening
Engineered host cells aren't only deliverables — they're also chassis for downstream library work. A common pattern: engineer a CHO line with knockouts of competing receptors and knock-ins of the target receptor, then build a mammalian display library in that engineered line. The display campaign now reports binding under conditions where the target is the dominant receptor presented, reducing off-target enrichment.
We covered the display side in [mammalian cell display: CHO and HEK293](/mammalian-cell-display-cho-hek293). The engineering side feeds in upstream — host chassis preparation typically takes 2–4 months before the first display library can be screened.
## Validation requirements
Engineered cell lines need defined validation before downstream use:
- **Genotype confirmation**: PCR, Sanger sequencing, or amplicon NGS at the edited locus. For clonal lines, off-target site characterization (top 10–50 predicted off-targets) by amplicon sequencing.
- **Phenotype confirmation**: expression of inserted gene (Western blot, flow cytometry, qPCR), absence of disrupted gene, functional readout of the engineered pathway.
- **Clonal stability**: serial passage characterization (10–20 passages) to confirm the edit doesn't drift or revert.
- **Mycoplasma and sterility**: standard release testing.
- **Karyotype** (for therapeutic-relevant lines): confirm no major chromosomal changes.
For research-grade lines, basic genotype and phenotype confirmation suffice. For therapeutic-relevant lines, the full release-testing package is required.
## Decision summary
For stable line development for biologics production: CHO host, lentiviral or transposase integration, 4–6 month timeline for a clonal high-producer.
For research-grade reporter cells: HEK293 host, transient or stable integration depending on assay duration, 1–3 month timeline.
For disease models: cell type matches the disease biology (often iPSC-derived), CRISPR knockout or knock-in, 3–6 months for clonal validated lines.
For functional assays paired with display screening: engineered host as a chassis, 2–4 month upstream investment, then reusable for multiple campaigns.
---
If you're scoping a cell engineering project, see our [cell engineering services](/cell-engineering) or reach out via the [contact page](/ranomics-contact). For mammalian display campaigns built on engineered host chassis, see [mammalian display](/technology/mammalian-display).
{/* ranomics:related-services */}
## Related Ranomics services
- **[Cell engineering](/cell-engineering):** Stable lines, CRISPR engineering, reporter cells, and functional assay development.
- **[Mammalian display](/technology/mammalian-display):** Display platforms built on engineered host cells.
- **[Protein engineering](/protein-engineering):** Protein-side engineering services paired with cell-side capabilities.
---
## Natural, Synthetic, AI-Designed Libraries for Antibody Discovery
> A successful antibody discovery campaign begins with choosing the right source of diversity. Comparing natural, synthetic, and AI-designed libraries for different therapeutic goals.
Source: https://ranomics.com/natural-synthetic-and-ai-designed-libraries-choosing-a-strategy-for-antibody-discovery/
Published: 2025-10-09
import KeyTakeaway from "../../components/KeyTakeaway.astro";
A successful antibody discovery campaign begins with choosing the right source of diversity. The three main library types each carry distinct advantages and trade-offs that must be matched to your target, therapeutic goals, and technical capabilities.
There is no universally best antibody library. Use a natural library when developability is the priority and immunization is viable. Use a synthetic library to bypass immune tolerance or reach unusual geometries. Use an AI-designed library when computational search can reach candidates the other two cannot. The strongest campaigns often combine them.
## Natural Diversity Libraries
Natural diversity libraries are harvested from immune repertoires, typically from immunized animals or human B-cell populations. These libraries contain antibody sequences that have already passed biological quality control: they fold, they express, and they have been selected for function in vivo.
**Advantages:** Biological validation and structural integrity. Sequences are drawn from a functional immune response, providing a baseline of expressibility and stability.
**Disadvantages:** Immunological tolerance constrains the diversity available. Self-reactive clones are eliminated by the immune system, creating blind spots for targets that resemble self-antigens. Library scope is limited to the donor's immune history.
## Synthetic Diversity Libraries
Synthetic libraries are built using synthesized oligonucleotides, with designed diversity introduced at specific CDR positions on validated framework scaffolds. This approach offers complete control over the sequence space sampled.
**Advantages:** Unlimited design control and massive scale. No dependence on immunization or donor availability. Can target sequence space that natural libraries cannot access.
**Disadvantages:** Sequences lack biological pre-validation. Not all designed variants will fold or express correctly. Synthetic CDR loops can carry immunogenicity risks if they diverge significantly from human germline sequences.
## AI-Designed Libraries
AI-designed libraries use machine learning trained on large antibody datasets to generate sequence diversity that balances novelty with predicted foldability and function. These approaches learn patterns from natural antibodies and use that knowledge to propose sequences that extend beyond the natural repertoire while respecting structural constraints.
**Advantages:** Combines the learning from natural antibody data with computational prediction of fitness. Can explore regions of sequence space that are difficult to reach by either natural or purely synthetic approaches.
**Disadvantages:** Dependent on training data quality and breadth. Requires specialized computational expertise. Model predictions require experimental validation.
## Choosing the Right Strategy
The choice depends on the specific campaign:
- **Natural libraries** are recommended when developability is the priority and the target is amenable to standard immunization.
- **Synthetic libraries** are the right choice when the target requires bypassing immunological tolerance or when the desired binding geometry is unusual.
- **AI-designed libraries** are best suited for cutting-edge applications where computational exploration of sequence space can identify candidates that neither natural nor synthetic approaches would find.
In practice, the strongest campaigns often combine approaches, using computational design to guide synthetic library construction, or using AI to filter and prioritize candidates from natural repertoire screens.
{/* ranomics:related-services */}
## Related Ranomics services
- **[Antibody engineering](/applications/antibody-engineering):** Custom antibody discovery across natural, synthetic, and designed libraries.
- **[Variant library construction](/technology/variant-library-construction):** Library strategy + build matched to your target class.
---
## Library Design Decisions That Determine Screening Campaign Success
> How diversity calculations, codon strategy, transformation efficiency, and NGS-based QC directly determine whether your screening campaign produces leads or wastes months of work.
Source: https://ranomics.com/navigating-the-hurdles-a-researchers-guide-to-optimizing-high-throughput-screens-library-design-and-protein-engineering/
Published: 2025-04-16
Most screening campaigns that fail to produce leads were doomed before a single colony was picked. The library was the problem. Specifically, the gap between the library that was designed on paper and the library that actually existed in the screening pool. Closing that gap requires precise decisions about diversity, codon strategy, transformation, and quality control.
## Theoretical Diversity Is Not Achievable Diversity
A six-position NNK library encodes 32⁶ = 1.07 × 10⁹ theoretical variants. That number is meaningless if your transformation yields 10⁷ transformants. You are sampling less than 1% of the designed space, and the variants you do recover are not a random draw.
The distinction between theoretical and achievable diversity is the first decision point in any library design. Achievable diversity is bounded by three constraints: the number of independent transformants, the uniformity of variant representation, and the fraction of the library that is functional (in-frame, no stop codons, no frameshifts).
For a saturation mutagenesis library, the standard coverage formula applies. To observe every variant at least once with 95% probability, you need approximately 3× oversampling of the library size. For 99% coverage, you need roughly 5×. For quantitative measurements (deep mutational scanning), 10× or higher is required to achieve adequate read depth per variant.
This means a library of 10⁵ unique variants requires 3 × 10⁵ to 10⁶ transformants depending on the application. Plan backward from the transformation bottleneck, not forward from the sequence design.
## NNK Is a Default, Not an Optimized Choice
NNK (N = A/T/G/C, K = G/T) at degenerate positions encodes all 20 amino acids in 32 codons. It is the most common codon strategy for saturation mutagenesis because it is simple and commercially available. It is also suboptimal for most applications.
The problem: NNK produces uneven amino acid representation. Leucine, serine, and arginine are each encoded by three codons. Tryptophan and methionine get one each. This 3:1 bias means rare amino acids are systematically undersampled unless you compensate with additional oversampling, which costs transformants you may not have.
Custom codon mixes (e.g., "22c trick," Tang et al.) reduce redundancy to one codon per amino acid using defined nucleotide mixtures at each position. This drops the codon count from 32 to 22 per position, compresses the library by (22/32)ⁿ for n randomized positions, and produces near-uniform amino acid representation. For a six-position library, that compression is 3.3-fold: 22⁶ ≈ 1.13 × 10⁸ versus 32⁶ ≈ 1.07 × 10⁹.
That compression directly translates to feasibility. A library that requires 10⁹ transformants for 3× coverage is out of reach for most labs. The same library at 10⁸ is achievable with standard electrocompetent cells.
For focused libraries where you want to restrict amino acid identity at specific positions (e.g., hydrophobic only, charged only), trimer phosphoramidite synthesis offers exact codon control with zero redundancy and zero stop codons. The cost per oligo is higher, but the savings in screening throughput and downstream validation make it the better investment for libraries under 10⁴ variants.
## Transformation Efficiency Is the Bottleneck
Electrocompetent *E. coli* typically yield 10⁸ to 10⁹ transformants per microgram of DNA for small plasmids (< 6 kb). For larger constructs, yeast display vectors, or lentiviral transfer plasmids, expect 10⁶ to 10⁷.
These numbers set a hard ceiling on library complexity. Every library design should include an explicit calculation: given the expected transformation efficiency and the amount of DNA available, what is the maximum achievable library size at the target coverage?
Common failure mode: designing a library on paper that requires 10⁸ unique transformants, then transforming into cells that yield 10⁷. The result is a library with 10% coverage, heavy sampling bias, and missing variants that may include the best hits.
Scale transformations appropriately. For libraries exceeding 10⁷ variants, multiple independent electroporations pooled together are standard practice. Track the total colony count rigorously. Estimating transformation efficiency from a dilution plate is not optional.
## QC the Library Before You Screen It
The cheapest experiment you can run is sequencing your library before committing to a screen. NGS-based library QC answers three questions that determine whether to proceed or rebuild.
**Is the variant distribution uniform?** Plot the frequency of each expected variant. A well-constructed library shows a tight distribution (coefficient of variation < 0.5). A skewed library means some variants are overrepresented by 100× or more while others are absent. Skewed libraries produce biased screening results because you are not sampling the designed space uniformly.
**What fraction of sequences are functional?** Frameshifts from oligonucleotide synthesis errors, deletions at ligation junctions, and premature stop codons from degenerate codon schemes all reduce the functional fraction of the library. A library with 30% frameshift contamination requires 3× more screening throughput to achieve the same effective coverage. For NNK libraries, stop codons appear at a rate of 1/32 per randomized position, compounding across multiple positions.
**Are all designed variants present?** Missing variants are invisible in the screen. If your library is missing 20% of the designed variants, you cannot find hits in that 20% regardless of screening throughput. NGS at 100× read depth per expected variant is sufficient to confirm presence and measure representation.
The cost of Illumina sequencing for library QC is a few hundred dollars. The cost of screening a defective library is months of work and consumables. This is not a tradeoff.
## Common Failure Modes
**Designing for theoretical rather than achievable diversity.** The library exists on paper but not in the flask. Always calculate backward from transformation efficiency.
**Using NNK by default when custom codons would compress the library into a feasible range.** A 10-minute codon optimization calculation can be the difference between a screenable and unscreenable library.
**Skipping library QC.** Proceeding to screen without confirming variant representation is the single most common and most expensive mistake in directed evolution campaigns.
**Insufficient oversampling.** 1× coverage means roughly 63% of variants are represented (Poisson sampling). 3× gets you to 95%. Anything below 3× is undersampled for hit identification.
## The Takeaway
Library design is not a preliminary step. It is the experiment. The choices made during library construction, codon strategy, and QC directly determine the ceiling of what a screening campaign can discover. A perfect assay screening a defective library produces nothing.
Get the library right first. Everything downstream depends on it.
At Ranomics, we design and validate variant libraries using computational protein design, custom codon strategies, and NGS-based QC to ensure campaigns start with libraries that can actually produce leads. [Start a project](/ranomics-contact).
{/* ranomics:related-services */}
## Related Ranomics services
- **[Variant library construction](/technology/variant-library-construction):** Library design, synthesis, assembly, and QC at scale.
- **[NGS analysis](/technology/ngs-analysis):** Library validation and post-sort enrichment analysis.
---
## NNK vs NNS vs Trimer: Picking a Codon Scheme for a VHH Library
> A practical comparison of NNK, NNS, NNN, and trimer codon schemes for VHH and scFv library design, covering stop codon frequency, amino acid coverage, S. cerevisiae codon bias, and when each scheme is the right choice for a yeast display campaign.
Source: https://ranomics.com/nnk-vs-nns-vs-trimer-codons-vhh-library/
Published: 2026-04-21
Codon scheme selection is one of the first cost versus coverage tradeoffs in any synthetic library. Choose wrong and you either over-spend on trimer synthesis or waste a measurable fraction of your displayed library on stop codons and disproportionately represented amino acids. For a VHH library diversifying CDR1, CDR2, and CDR3, the codon scheme sets the upper bound on what the downstream FACS sort can find. This post walks through how NNK, NNS, NNN, and trimer schemes differ in practice, how they interact with S. cerevisiae codon bias, and when each is the right choice for a yeast display campaign.
## How Each Codon Scheme Works
Four codon schemes dominate VHH and scFv library design. Each makes a different tradeoff between coverage, redundancy, stop codon rate, and synthesis cost.
**NNK** uses N1 N2 K3, where K = G or T at the third position. That produces 32 codons encoding all 20 amino acids plus one stop (TAG). Stop codon frequency is roughly 3% per position. Amino acid coverage is complete but non-uniform: Leu, Ser, and Arg each get three codons, while Met, Trp, and several others get only one. This degeneracy maps directly onto library bias.
**NNS** uses N1 N2 S3, where S = G or C. Same 32 codons as NNK at the count level, same one stop (TAG), same ~3% stop frequency. The amino acid distribution is slightly different because the codon subset is not identical. In practice, NNK and NNS produce libraries with comparable information content and comparable hit rates. Groups that have optimized one tend to stay with it.
**NNN** is full randomization: all 64 codons, 20 amino acids, 3 stops (TAA, TAG, TGA). Stop codon frequency per position is roughly 5%. The higher stop rate is the main reason NNN is rarely used for yeast display libraries. At 6 diversified positions, the fraction of stop-free clones drops from ~83% (NNK) to ~73% (NNN). That is real library burned on truncated product.
**Trimer mixes** use 20 pre-synthesized trinucleotide codons, one per amino acid, pooled at defined ratios. No stop codons. No redundancy. Amino acid frequency is fully user-controlled. The cost is the main constraint: trimer synthesis typically runs two to three times the price of standard randomized oligo synthesis, and vendor turnaround is longer.
## When to Pick NNK
NNK is the workhorse of first-round VHH and scFv library construction. Pick NNK when:
- The library is naive or semi-naive and the goal is broad exploration of CDR sequence space.
- You are diversifying fewer than about 10 positions and the stop codon tax is tolerable. For a 6-position NNK block the stop-free fraction is (31/32)^6 ~= 82.8%; for 9 positions it drops to ~75%; for 12 positions to ~68%. Past that point trimer starts to pay for itself.
- Cost is a real constraint and the campaign has budget for one round of library construction, not three.
- You are screening against a soluble target where FACS enrichment will sort through the stop codon noise in the first round anyway.
The practical ceiling for NNK in a standard yeast display campaign is around 8 to 10 fully randomized positions. Beyond that, the redundant codons for Leu, Ser, and Arg start to dominate and under-sample the smaller amino acids (Trp, Met, Cys).
## When to Pick NNS
NNS is close enough to NNK in behavior that the decision between them is usually historical rather than principled. The amino acid distributions differ slightly at the third-position wobble: NNK over-represents a handful of amino acids that NNS under-represents, and vice versa.
Pick NNS when:
- Your oligo vendor supplies NNS as standard and you want a drop-in substitution.
- You have existing NNS protocols, NNS-designed primers, or NNS reference libraries and the downstream bioinformatics already accounts for that bias.
- You are running a paired-library comparison where you want two slightly different bias profiles against the same target.
For most first-time VHH campaigns, NNK and NNS are interchangeable. The differences are visible in a deep NGS readout after sort, not in the first-round hit rate.
## When to Pick Trimer
Trimer is the right choice when amino acid representation matters more than cost. Specifically:
- High-diversity positions where uniform coverage is required. If you are randomizing 12 or more positions in a single CDR, the compounding effect of stop codons and codon redundancy in NNK or NNS degrades your effective library size enough that the extra synthesis cost is worth it.
- Affinity maturation of a validated lead where you want every position to sample the same 20 amino acids at the same frequency. Controlled trimer mixes can exclude undesired amino acids entirely (for example, Cys for stability, Met for oxidation, Pro for conformational rigidity).
- Focused alphabet sub-libraries. Trimer can encode any user-defined amino acid subset, which is hard to replicate with a single wobble position.
- Campaigns where you have already validated the design concept with NNK and are scaling into a larger, more expensive maturation library.
Trimer requires earlier vendor coordination and adds one to three weeks of lead time compared to standard oligo synthesis. For campaigns on a tight timeline, that alone can be the deciding factor.
## Codon Bias in S. cerevisiae
Picking a codon scheme at the DNA level does not fully determine what actually displays on the yeast surface. *S. cerevisiae* has measurable codon preferences. Rare codons translate more slowly and can reduce expression or induce ribosomal stalling, which reduces the effective display level of variants carrying them. Arg CGG, for example, is used at roughly 4% frequency in highly expressed yeast genes, compared to more than 20% for Arg AGA or AGG. A variant whose Arg is encoded by CGG displays at a lower level than a variant whose Arg is encoded by AGA, even if both cells carry functionally equivalent protein.
For NNK and NNS libraries, this matters because the redundant codons for Arg, Leu, and Ser are not all equally well-expressed in yeast. Some variants drop out of the sort not because they fold poorly or bind weakly, but because the yeast ribosome does not produce them efficiently. This is one source of the well-known bias against Arg and Leu variants in yeast display NGS readouts.
Trimer mixes sidestep this entirely. Each amino acid is encoded by a single yeast-preferred codon, chosen at synthesis time. The library that arrives is codon-optimized by construction. For a maturation library where every position matters, this is a real advantage.
NNK and NNS libraries can be partially corrected by codon-usage-aware oligo design, but not fully: the wobble position fundamentally constrains which codons are possible. The practical path is to accept the bias at the first round and use NGS to quantify which amino acids are under-represented in the post-sort population.
## The Math for a VHH CDR3
A representative calculation makes the tradeoffs concrete. Consider a 6-position NNK randomization in VHH CDR3:
- Theoretical DNA diversity: 32^6 = ~10^9 unique oligo sequences.
- Theoretical amino acid diversity: 20^6 = ~6.4 * 10^7 distinct protein sequences.
- Stop-free clones: (31/32)^6 ~= 83% of transformants encode full-length protein.
- Yeast transformation ceiling: ~10^8 unique transformants with optimized electroporation.
At this scale, a well-transformed NNK library can theoretically sample the full 6.4 * 10^7 protein space with coverage, and FACS plus NGS can resolve enrichment across it. Push to 8 NNK positions and theoretical protein space is 20^8 = ~2.6 * 10^10, which exceeds the yeast ceiling by two to three orders of magnitude. At that point the library undersamples by construction and codon scheme choice can either reduce or amplify the undersampling.
The [Yeast Display Library Planner](/technology/library-planner) runs this math against your scaffold and diversification pattern before you commit to synthesis. It flags cases where the theoretical diversity exceeds the yeast ceiling, where stop codon load reduces the effective library size below your screening depth, or where codon redundancy clusters the library around a small set of amino acids.
## Putting It Together
For a first VHH campaign on a single target, NNK at 6 to 8 diversified positions per CDR, screened against a well-chosen hotspot, is a defensible starting point. For a maturation campaign on a validated lead, trimer at the positions that need uniform coverage is worth the cost difference. For paired libraries or method-comparison studies, NNS is a reasonable alternative to NNK with a slightly different bias profile. NNN is rarely the right choice for yeast display given the stop codon tax.
Codon scheme is downstream of three upstream decisions: the target affinity goal (naive discovery vs affinity maturation), the diversification footprint (how many positions, across how many CDRs), and the available budget for synthesis and screening. Get those right first and the codon scheme falls out of the specification.
## Scoping a Campaign
If you want the math done for your specific scaffold and diversification pattern, the [Yeast Display Library Planner](/technology/library-planner) takes a parent sequence, diversified positions, and target library size and returns a scoping-ready plan with codon scheme recommendations, stop codon load, theoretical and achievable diversity, and NGS read depth for each sort round.
For a full campaign scoped end to end, including display construct design, library construction, FACS or MACS selection, and NGS hit calling, the [Yeast Display service](/yeast-display) page covers platform capabilities. To scope a specific project, use the [contact form with the yeast display prefill](/ranomics-contact?service=yeast-display) and include the scaffold, target, and approximate diversification footprint.
{/* ranomics:related-services */}
## Related Ranomics services
- **[Yeast Display Library Planner](/technology/library-planner):** Free planning tool for codon scheme, library size, and NGS depth.
- **[Yeast Display](/yeast-display):** End-to-end yeast surface display CRO service for scFv, VHH, and scaffold libraries.
- **[AI Binder Sprint](/ai-binder-sprint):** Computational binder design paired with yeast display validation.
---
## Engineering pH-Dependent Antibodies on Yeast Surface Display: A 640-Clone Case Study
> A technical walkthrough of a real pH-dependent antibody engineering campaign — 640-clone yeast display library, six FACS sorts, convergent hotspot residues, and quantitative enrichment-score ranking.
Source: https://ranomics.com/ph-dependent-antibody-engineering-yeast-display-case-study/
Published: 2026-04-20
## Why pH-Dependent Binding Matters
A binder that grips its target with equal force at pH 7.4 and pH 5.5 is, for many antibody engineering problems, the wrong answer. The useful answer is a binder that switches.
Several of the most commercially relevant antibody formats depend on exactly this behavior. FcRn-mediated recycling requires antibodies that bind tightly at the acidic endosomal pH and release into circulation at neutral pH — the basis of half-life extension and recycling antibody technology. Tumor microenvironments run measurably more acidic than healthy tissue, making pH-selective binding a mechanism for improving therapeutic index. Antibody-drug conjugate linker chemistry relies on pH-dependent stability to hold payload in circulation and release it in the lysosome.
Engineering this behavior is not the same problem as engineering tight binding. Selection pressure has to discriminate between two binding states — not just present-versus-absent, but present-at-one-pH-versus-absent-at-the-other. Library design, selection format, and hit-ranking methodology all have to accommodate that two-state requirement from the start.
This case study walks through a real campaign we ran for a client (details anonymized, convergent hotspot residues designated H-1 and H-2). The goal was to engineer pH-dependent binding into an existing antibody lead using yeast surface display. Six cycles of FACS selection on a 640-clone targeted mutagenesis library converged on a small panel of ranked candidates with quantitatively validated pH-switching behavior.
## The Library
640 clones is small by naive-library standards and deliberately so. pH-dependent behavior is usually controlled by a handful of ionizable residues — typically histidines positioned within the binding interface — and the high-yield approach is targeted mutagenesis rather than random diversification. We diversified a defined set of interface positions with chemistry-guided substitutions focused on ionizable and neutral alternates. Every variant was designed, not random.
Library diversity was confirmed by NGS before selection began. Coverage and uniformity are the two things that matter here: every designed variant should be present, and no single variant should dominate early reads. Both checks passed.
## Six-Cycle FACS Selection
Selection was structured as paired sort arms — one at pH 7.4, one at pH 5.5 — with cycle-to-cycle stringency tightened on both sides. The goal in each round was not simply to enrich binders, but to enrich the difference between binding states.
Two-color FACS labeling (surface expression marker plus fluorescent antigen) enabled gating on binding normalized to expression level. Normalization is non-negotiable for yeast display campaigns: raw antigen signal conflates genuine affinity with display level, and without correction the enrichment score picks up noise from expression variation rather than the property you care about. (For a deeper treatment of this, see our article on [correctly titrating display levels for reliable affinity data](/correctly-titrating-display-levels-for-reliable-affinity-data-in-yeast-and-mammalian-systems).)
Six cycles is more than a typical affinity maturation campaign. The reason is signal-to-noise on the enrichment trajectory. A variant that happens to sort well in one round from statistical chance will not sort well across six consecutive rounds. Cumulative enrichment across the full sort series is the discriminator that separates real hits from library noise.
## Enrichment Scores, Not Read Counts
A variant ranked by raw read count at cycle six is an artifact of starting frequency as much as selection. A variant ranked by per-cycle enrichment — log-fold change in frequency from one sort to the next, summed across all cycles — is a much cleaner readout of how selection actually behaved. Enrichment-score methodology turns each variant's full NGS trajectory into a single number that is comparable across the library.
The full case study PDF includes the per-cycle enrichment distributions and the cumulative scatter that generated the final candidate ranking. The tails are where the interesting biology lives.
## Convergent Hotspots
Two positions emerged as dominant across enriched lineages. We designate them H-1 and H-2 to preserve client anonymity. Convergence is the single strongest piece of evidence that an engineering campaign has identified real biology rather than library drift. When independent variant lineages — built around different combinations of substitutions at surrounding positions — all terminate at enriched variants that share a substitution at H-1 or H-2, the interpretation is that those two positions genuinely control the pH-switching behavior. Variants that lack either hotspot substitution do not enrich, regardless of what else is in their sequence.
This has a practical consequence for the client's next round of engineering: library design for any follow-on campaign fixes H-1 and H-2 and diversifies elsewhere. Knowing what to hold constant is as valuable as knowing what to vary.
## What the Full Case Study Covers
The PDF walks through library design and QC, FACS gating strategy per cycle, the full enrichment-score methodology with per-cycle distributions, the convergent hotspot analysis with annotated scatter plots, the cumulative ranking that produced the final candidate panel, and a decision framework for scoping pH-dependent yeast display campaigns on other targets.
It is written for protein engineers and antibody discovery leads who want to understand how the pieces fit together on a real campaign, not a synthetic example.
[**Read the full case study (PDF, 14 pages) →**](/ph-dependent-antibody-engineering)
{/* ranomics:related-services */}
## Related Ranomics services
- **[Yeast surface display services](/yeast-display):** FACS and MACS-based screening of protein libraries, including pH-dependent and condition-dependent selection schemes.
- **[Affinity maturation](/applications/affinity-maturation):** Systematic optimization of lead binders using targeted mutagenesis and iterative display selection.
- **[Antibody engineering](/applications/antibody-engineering):** End-to-end antibody discovery and optimization, including pH-dependent and recycling antibody campaigns.
---
## Phage Display vs Yeast Display: When to Choose Which Platform
> Phage display, yeast display, and mammalian display each fit different campaigns. A practical decision framework for choosing the right platform.
Source: https://ranomics.com/phage-display-vs-yeast-display/
Published: 2026-04-30
import KeyTakeaway from "../../components/KeyTakeaway.astro";
If your binder discovery campaign is between phage and yeast display, the choice usually comes down to library size, post-translational modifications, and how you plan to read out affinity. There is no universal "best" platform. Choosing wrong wastes a quarter; choosing right compounds. This is how we decide.
Choose phage display when you need a library larger than 10^9 variants and cost is the main constraint. Choose yeast display when the library fits within 10^9 and you need quantitative per-clone affinity ranking from the screen itself. For AI-designed libraries, which are smaller, yeast display is usually the right call.
## The three platforms in one paragraph each
**Phage display** uses M13 or T7 phage to present scFv, Fab, or peptide libraries on the phage coat. Library sizes routinely reach 10^11 to 10^12 unique variants. Selection happens through panning rounds against immobilized antigen. The displayed protein is folded in the bacterial periplasm, which limits post-translational modifications to bacterial chaperone-assisted folding. Phage displays at 1-5 copies per particle, so avidity effects are present but tractable.
**[Yeast surface display services](/yeast-display)** anchor proteins to the yeast cell wall via the Aga2p mating protein or similar fusions. Library sizes are 10^7 to 10^9 unique variants, set by transformation efficiency. Display levels of 10,000-100,000 copies per cell create strong avidity, which is both an advantage (rare binders pull out) and a liability (apparent affinity can be inflated 100-1000x; see [the avidity artifacts article](/eliminating-false-positives-mastering-avidity-effects-in-yeast-display-screening)). Selection uses FACS or [MACS](/beyond-facs-an-introduction-to-magnetic-activated-cell-sorting-macs-for-library-pre-enrichment), and quantitative readout via flow cytometry is the standout feature.
**Mammalian cell display** (CHO or HEK293) presents proteins through transmembrane fusion or PDGFRβ-anchored constructs. Library sizes are 10^6 to 10^8 — smaller than yeast — but the cells handle full glycosylation, complex disulfide networks, and the secretory pathway machinery that yeast and bacteria cannot replicate. Used when target binding depends on PTMs the other platforms cannot install.
## The decision framework
| Dimension | Phage | Yeast | Mammalian |
|---|---|---|---|
| Library size ceiling | **10^11-10^12** | 10^7-10^9 | 10^6-10^8 |
| PTM fidelity (glycosylation, disulfides) | Limited | Moderate (yeast-style high-mannose) | **Native human** |
| Quantitative affinity readout | Indirect (ELISA after panning) | **Direct (flow cytometry)** | Direct (flow cytometry) |
| Avidity inflation | Low (1-5 copies/particle) | High (10K-100K copies/cell) | Moderate (titratable) |
| Throughput per round | High (10^11 input) | Moderate (10^9 input) | Lower (10^7 input) |
| Cost per campaign | Lowest | Mid | Highest |
| Time per round | 1-3 days | 3-5 days | 5-7 days |
| Suitable for full-length IgG | No | No (Fab/scFv only practical) | **Yes** |
## When to use which
**Choose phage display when:**
- You need to interrogate a sequence space larger than 10^9. The library size advantage is real and not replaceable.
- Your scaffold is small (peptides, scFv, single-domain antibodies) and doesn't require complex PTMs.
- You're at the discovery stage and willing to triage false positives in subsequent rounds.
- Cost is the primary constraint. Phage panning is the cheapest of the three to run.
**Choose yeast display when:**
- You need quantitative affinity ranking from the screen itself (flow cytometry gives a Kd estimate per clone in one experiment).
- Your library fits in 10^9 — usually true for focused or AI-designed libraries.
- You need to characterize multiple binders simultaneously rather than just enrich for the strongest. The avidity-corrected, single-clone Kd readout is a serious advantage when ranking multiple leads.
- You plan to do directed evolution rounds — the [titratable display level](/correctly-titrating-display-levels-for-reliable-affinity-data-in-yeast-and-mammalian-systems) lets you tune selection stringency precisely.
**Choose mammalian display when:**
- The target requires native glycosylation, complex disulfide networks, or PTM-dependent epitopes (Fc-mediated effector function, Fc-FcRn binding, glycan-specific antibodies).
- You're optimizing a full-length IgG, where yeast display's truncation to scFv/Fab loses developability information.
- You're past the discovery stage and need [developability triage](/the-impact-of-post-translational-modifications-in-mammalian-protein-production-and-antibody-discovery) under conditions that match production cells.
## Edge cases — when you actually need more than one
Most of our flagship campaigns use **two platforms sequentially**, not one. The most common pattern: yeast display for affinity discovery, mammalian display for developability validation. The yeast round generates and enriches binders fast; the mammalian round filters out variants that fail PTM compatibility before commitment to expression scale-up. We've documented this pattern at length in [the two-platform approach article](/the-two-platform-approach-using-yeast-display-for-affinity-and-mammalian-display-for-developability).
A second pattern: phage display for the initial diversity sweep (10^11 library), followed by yeast display for the quantitative rank-order on the top 10^4-10^5 hits. This works well when the target is hard and a focused library would miss the right epitope.
A third pattern, less common: mammalian display from round one when the target itself is a glycoprotein and yeast-displayed binders consistently fail to recapture the relevant epitope. Membrane-bound complement proteins are a typical example.
## How AI design changes the calculus
When the input library is AI-designed rather than randomly diversified, the library-size advantage of phage display matters less. Ranomics' AI design pipeline typically produces 10,000 to 60,000 candidate binders per campaign — well within yeast display's working range. In this regime, the **quantitative affinity readout** becomes the dominant criterion, which favors yeast or mammalian display over phage.
The integrated picture: AI design narrows the search space, display screening validates and ranks. Phage display's "search" advantage is replaced by computational sampling. Yeast display's "rank" advantage remains essential.
## Decision summary
If you're between phage and yeast and you have an AI-designed library: choose yeast.
If you have a randomly diversified library that needs to be larger than 10^9: choose phage, with a yeast follow-up.
If your target's biology depends on PTMs: skip both and go to mammalian display.
If you're not sure: get into a 30-minute conversation about your target structure, library design, and downstream validation plan. The right platform is downstream of those choices, not upstream.
---
Ranomics designs and screens binders on all three platforms. If you're scoping a campaign and want a second opinion on platform fit, [start a Binder Pilot](/binder-pilot) or reach out via the [contact page](/ranomics-contact).
---
## Post-Translational Modifications in Mammalian Protein Production
> Post-translational modifications are a fundamental layer of biological regulation that dictates a protein's function and viability as a drug. Understanding glycosylation, disulfide bonds, and chemical liabilities.
Source: https://ranomics.com/the-impact-of-post-translational-modifications-in-mammalian-protein-production-and-antibody-discovery/
Published: 2025-10-07
After a protein is synthesized by the ribosome, the job is not done. The protein still has to undergo a series of chemical alterations known as post-translational modifications (PTMs).
Nowhere is this more critical than in antibody development and complex biologic drug development.
Understanding PTMs is essential for any protein engineer. These modifications can be the difference between a potent, stable therapeutic and a heterogeneous, ineffective, or even immunogenic product.
## Glycosylation: A Regulator of Function and Stability
**N-linked Glycosylation** occurs on the nitrogen atom of an asparagine (N) residue, typically within the sequon N-X-S/T (where X is any amino acid except proline). For antibody development, the conserved N-linked glycan in the Fc region is a critical regulator of effector function.
**O-linked Glycosylation** occurs on the oxygen atom of serine (S) or threonine (T) residues.
Unintended glycosylation at a newly created N-X-S/T site (e.g., in a CDR loop) is a major developability risk that can impair antigen binding.
## Disulfide Bond Formation: The Architect of Structure
For a typical IgG antibody, multiple cysteine residues must be correctly paired to create the iconic Y-shape.
Incorrect disulfide bond pairing leads to misfolding, loss of function, and often triggers the cell's quality control machinery to degrade the protein. The presence of unpaired cysteines on a final product is a major developability red flag.
## Proteolytic Cleavage: Essential Maturation and a Source of Heterogeneity
**Signal Peptide Removal:** Incomplete cleavage can lead to a heterogeneous product with a different N-terminus than intended.
**C-terminal Clipping:** For antibodies, clipping of the C-terminal lysine residue on the heavy chain is a common modification that must be characterized and controlled.
## "Problematic" PTMs: The Chemical Liability Hotspots
- **Deamidation:** Spontaneous conversion of asparagine (N) into aspartic acid or isoaspartic acid. Particularly common at N-G and N-S motifs.
- **Oxidation:** Most commonly affects methionine (M) and tryptophan (W). Oxidation of a methionine residue within a CDR can dramatically reduce an antibody's binding affinity.
- **Isomerization:** Conversion of aspartic acid (D) into isoaspartate, altering the backbone geometry.
## Conclusion: PTMs as a Critical Quality Attribute
Post-translational modifications are not an afterthought in antibody development and protein production; they are a central determinant of a biologic's safety, efficacy, and viability as a drug. Thinking about PTMs from day one is key to success.
{/* ranomics:related-services */}
## Related Ranomics services
- **[Mammalian display](/technology/mammalian-display):** Proper post-translational modifications on a display platform.
- **[Antibody engineering](/applications/antibody-engineering):** PTM-aware antibody development including glycoform control.
---
## Protein Engineering Design in the Age of Machine Learning
> Modern protein engineering design increasingly relies on machine learning, but experimental data and workflow integration remain the true bottlenecks. A guide to the six-stage design cycle.
Source: https://ranomics.com/protein-engineering-design-in-the-age-of-machine-learning/
Published: 2026-02-10
Protein engineering is entering a new phase. Machine learning has dramatically expanded our ability to generate novel protein sequences and structures, but success in protein engineering design is no longer limited by model capability alone.
## The Modern Protein Engineering Design Cycle
### 1. Backbone & scaffold generation
**Tools:** RFdiffusion, Protpardelle-1c, Chroma, BoltzDesign 1
Bad backbones dominate late-stage failures. The quality of the initial scaffold determines the ceiling for everything downstream.
### 2. Sequence generation & initial binder design
**Tools:** BindCraft, BoltzGen, PXDesign, Protein Hunter, ColabDesign, Germinal
Some tools bias toward **hit rate**, others toward **exploration**. That choice directly shapes what your experimental screens will see.
### 3. Multi-objective optimization
**Tools:** Mosaic, ProteinMPNN, Rosetta FastDesign/Relax
Most experimental attrition is not due to lack of binding. It is due to expression, aggregation, or instability. This stage addresses the developability gap.
### 4. Diversity expansion & hypothesis coverage
**Tools:** PXDesign, Protein Hunter, RFdiffusion (noise/temperature tuning), Neighborhood sampling
This stage is about **coverage**, not convergence. Generating enough structural and sequence diversity to hedge against prediction failures.
### 5. Filtering, scoring & triage
**Tools:** AlphaFold2 metrics (pLDDT, PAE), Rosetta InterfaceAnalyzer, FoldX, Aggregation/solubility predictors
Most failures are filtered out here. Computational triage reduces the experimental burden by orders of magnitude.
### 6. Experimental data -> learning loop
**Key screens:** Display-based selections, deep mutational scanning, expression and stability screens, cell-based functional assays
This is the part most design discussions skip, and where differentiation now lives. The quality of your experimental data determines whether the next design cycle improves or plateaus.
## Conclusion
What is becoming clear is that generative protein design is no longer about finding the best model. It is about how different tools shape the hypotheses you generate and, ultimately, the experimental data you collect.
Protein engineering design is no longer defined by any single model or algorithm. As protein engineering machine learning continues to improve hit rates, competitive advantage is shifting toward experimental strategy, design diversity, and high-quality data generation.
## FAQ
**What is protein engineering design?**
The process of modifying or creating proteins with desired functions using computational and experimental methods.
**How is machine learning used in protein engineering?**
Models generate, score, and optimize protein sequences, but experimental validation remains essential.
**What limits protein engineering today?**
The primary limitation is no longer sequence generation, but experimental throughput and high-quality functional data.
{/* ranomics:related-services */}
## Related Ranomics services
- **[Protein engineering services](/protein-engineering):** ML-integrated protein engineering from target to validated variant.
- **[AI design engine](/technology/ai-design-engine):** Production ML stack for design-build-test-learn cycles.
---
## Protein Folding Optimization in Yeast Display: Engineering Better Expression Systems
> Master protein folding optimization in yeast display systems with this guide covering signal peptide engineering, chaperone co-expression, ER retention strategies, and promoter optimization.
Source: https://ranomics.com/protein-folding-optimization-in-yeast-display-engineering-better-expression-systems/
Published: 2025-08-19
## The Protein Folding Challenge in Yeast Display
Heterologous proteins expressed in [yeast display](/yeast-display) systems face a gauntlet of quality control checkpoints in the secretory pathway. Co-translational translocation into the ER, interaction with chaperones (BiP, PDI, calnexin), disulfide bond formation, and passage through the ERAD (ER-Associated Degradation) pathway all determine whether a protein reaches the cell surface or is degraded.
For non-native proteins, particularly those with complex disulfide architectures or large domains, the yeast ER environment is often insufficient. Optimizing the folding environment is essential for achieving display levels that support meaningful screening.
## Signal Peptide Optimization: The Gateway to Proper Folding
The alpha-factor signal peptide is the default choice for yeast secretion, but it is not universally optimal. Alternative signal peptides can dramatically improve translocation efficiency for specific protein classes.
Optimization approaches include rational design of the hydrophobic core and charge distribution, as well as empirical screening of signal peptide libraries. Cleavage site efficiency directly affects the homogeneity of the displayed protein.
## Chaperone Co-expression: Enhancing the Folding Environment
**BiP (Kar2p) overexpression** increases the capacity of the ER to handle misfolded intermediates, reducing aggregation and improving the yield of correctly folded protein.
**PDI1 co-expression** is particularly beneficial for disulfide-containing proteins, accelerating the formation of correct disulfide bonds and reducing the accumulation of misfolded intermediates.
Multi-chaperone strategies can provide synergistic improvements, but the timing and level of chaperone expression relative to the target protein must be optimized to avoid overwhelming the secretory pathway.
## ER Retention Strategies: Extending Folding Time
**KDEL retention signals** slow the transit of proteins through the secretory pathway, providing additional time for chaperone-assisted folding. This is particularly useful for slow-folding proteins.
**Cleavable retention signals** offer a compromise: the protein is retained long enough to fold correctly, then released for surface display.
**Temperature control** in conjunction with retention strategies further slows folding kinetics, favoring the native state over kinetically trapped misfolded conformations.
## Promoter Engineering and Expression Optimization
The choice of promoter directly affects the rate and level of protein expression:
- **GAL1:** Inducible, strong expression. Standard for yeast display.
- **PGK1:** Constitutive, moderate expression.
- **TDH3:** Constitutive, high expression.
Expression timing relative to growth phase, translation rate effects on co-translational folding, and the use of tandem upstream activation sequences all provide additional levers for optimization.
The key principle: more expression is not always better. For complex proteins, slower translation rates can improve the fraction of correctly folded product by giving chaperones time to act on each nascent chain.
{/* ranomics:related-services */}
## Related Ranomics services
- **[Yeast surface display services](/yeast-display):** Expression-optimized constructs for difficult-to-display proteins.
- **[Protein engineering services](/protein-engineering):** Stability and expression engineering beyond display screening.
---
## proteinmpnn-sequence-design-explained.mdx
Source: https://ranomics.com/proteinmpnn-sequence-design-explained/
import TryToolCallout from "../../components/TryToolCallout.astro";
import InOurWork from "../../components/InOurWork.astro";
Generating a protein backbone and designing a protein sequence are two separate computational problems. RFdiffusion solves the first: given a target surface, generate a backbone geometry that could engage it. ProteinMPNN solves the second: given that backbone, which amino acid sequences will actually fold into it?
This distinction matters because a backbone without a sequence is not a protein. Many plausible-looking backbone geometries from diffusion models cannot be realized by any sequence that would actually fold stably. ProteinMPNN's role is to identify sequences that are geometrically and thermodynamically compatible with the backbone, and to do so with enough diversity that you get a pool of candidates to screen, not a single prediction.
## The inverse folding problem
Structure prediction (AlphaFold, Boltz-2) takes a sequence and predicts its structure. Inverse folding takes a structure and asks: what sequences are consistent with this geometry? This is fundamentally a harder and less determined problem. For any given backbone, many sequences are at least partially compatible, but most will not fold stably.
ProteinMPNN addresses this using a message-passing neural network trained on structure-sequence pairs from the PDB. Given backbone coordinates (C-alpha, C-beta, C, N, O positions), it learns to predict which amino acids at each position are consistent with the local and global structural context. The model predicts a probability distribution over the 20 amino acids at each position, conditioned on the full backbone geometry.
At inference time, sequences are sampled from this distribution. The sampling temperature controls the sharpness of the distribution: low temperature selects the highest-probability residue at each position (more conservative, higher predicted stability), high temperature samples more diversely (lower average confidence, higher sequence diversity).
## How ProteinMPNN fits into a de novo binder design pipeline
After RFdiffusion generates backbone structures, ProteinMPNN is run on each backbone to generate candidate sequences. A typical production workflow:
1. RFdiffusion produces 10,000-50,000 backbone structures, each geometrically conditioned on the target hotspot.
2. An initial quality filter removes backbones with low diffusion confidence scores or implausible geometry.
3. ProteinMPNN generates 8-16 sequences per remaining backbone, sampled at multiple temperatures.
4. The sequence pool (tens of thousands of candidates) is passed to structure prediction (Boltz-2, ESMFold, ColabFold) for validation.
5. Candidates are filtered by predicted complex quality (ipTM, PAE, interface pLDDT).
6. The filtered pool (typically 500-2,000 sequences) advances to synthesis.
ProteinMPNN is fast relative to structure prediction: running it on thousands of backbones takes minutes. The computational bottleneck in the pipeline is the complex structure prediction step, not sequence design.
---
**Identify epitopes on your own target.** [Epitope Scout](https://scout.ranomics.com) scores and ranks surface patches on any PDB structure. Free to use.
---
## Fixed-position residues: locking interface contacts
One of the most useful features of ProteinMPNN is the ability to fix specific residue identities and allow the model to design only the non-fixed positions. In binder design, this is used to preserve contacts at predicted hotspot positions.
If the backbone places a residue in direct contact with a hotspot on the target, and structural analysis suggests that a specific amino acid (e.g., an Arg or Tyr forming a hydrogen bond network with the target interface) should be conserved, that position can be fixed. ProteinMPNN will design the rest of the sequence freely while respecting the constraint.
This is particularly useful when:
- A pilot screen has identified a weak binder and you want to preserve the key contact while redesigning the scaffold
- Structural modeling suggests a specific interaction motif (aromatic stacking, charged-charged, etc.) that should be maintained
- You are doing affinity maturation of a confirmed hit and want to hold the known binding contacts while diversifying the surrounding scaffold
In our work, we have seen raw ProteinMPNN output produce sequence motifs that no naturally occurring protein would tolerate: stretches of poly-glutamate or poly-lysine, runs of aliphatic residues, motifs like EYEVFYEY that simply do not appear in the natural sequence record. The reason is that hallucinated backbones often have geometry that does not match what natural proteins actually fold to, so the model fills those positions with whatever its training distribution allows at that local context. Post-processing with amino-acid biases that pull the composition toward natural reference frequencies is non-negotiable for us before any sequence advances to synthesis.
## Temperature sampling: managing the diversity-stability trade-off
The sampling temperature parameter is the main lever for controlling sequence diversity in ProteinMPNN output.
At **low temperature (T = 0.1-0.2)**: sequences are close to the model's maximum-likelihood prediction. High average sequence confidence, high predicted stability, low diversity. Suitable when you want a small number of high-confidence predictions per backbone.
At **standard temperature (T = 0.3-0.5)**: the default for most production runs. Reasonable diversity with acceptable predicted quality. Generates sequences that are distinct from each other without excessive deviation from the backbone's design constraints.
At **high temperature (T = 0.7-1.0)**: high sequence diversity across the output pool. More exploration of sequence space, but a larger fraction of sequences will fail the downstream structural validation step. Useful when you want to characterize the sequence landscape around a scaffold topology or when you suspect the low-temperature output is too conservative.
In practice, running ProteinMPNN at two temperatures (e.g., T = 0.2 and T = 0.5) per backbone and pooling the outputs provides both confident predictions and a diversity buffer without substantially increasing compute time.
## What ProteinMPNN does not do
ProteinMPNN does not predict binding affinity. It predicts sequence-backbone compatibility (whether a sequence will fold into the provided structure), not whether the resulting protein will bind the intended target. A high-confidence ProteinMPNN score means the sequence is likely to fold; it says nothing about the quality of the binding interface.
This is why structural validation against the target (Boltz-2 / ESMFold / ColabFold complex prediction) is a required subsequent step, not optional. Sequences that ProteinMPNN scores favorably for backbone stability may still have weak, misaligned, or absent binding interfaces when the full complex is predicted.
The pipeline (RFdiffusion, ProteinMPNN, complex prediction, experimental screen) works because each step filters on a different property. No single computational step is sufficient. The experimental display screen is the only step that measures actual binding. In our pipeline, no sequence advances to gene synthesis without first passing a structural validation step and an amino-acid composition check against natural reference frequencies.
---
**See how ProteinMPNN fits into the Ranomics pipeline:** [AI Design Engine](/technology/ai-design-engine)
{/* ranomics:related-services */}
## Related Ranomics services
- **[ProteinMPNN](/technology/proteinmpnn):** Sequence design across designed backbones with developability-aware filters.
- **[AI design engine](/technology/ai-design-engine):** Orchestrated RFdiffusion + ProteinMPNN + BindCraft + validation pipeline.
---
## proteinmpnn-web-tool-sequence-design-no-install.mdx
Source: https://ranomics.com/proteinmpnn-web-tool-sequence-design-no-install/
ProteinMPNN is now available as a one-click web tool at [tools.ranomics.com/tools/mpnn](https://tools.ranomics.com/tools/mpnn). Upload a backbone structure, configure sampling temperature and the number of sequences, and download a FASTA. No GPU, no conda environment, no command line required.
This post covers what the tool does, when to use it, how to use it, and what it costs.
## What ProteinMPNN does (two sentences)
ProteinMPNN solves the inverse folding problem: given a backbone structure, it predicts which amino acid sequences are geometrically and thermodynamically compatible with that geometry. It outputs a probability distribution over the 20 amino acids at each position conditioned on the full backbone, from which sequences are sampled at a user-specified temperature.
For a full treatment of the model architecture, the temperature parameter, and how fixed-position constraints work, see the companion article: [ProteinMPNN and the sequence design problem](/resource-hub/proteinmpnn-sequence-design-explained).
## When to use this tool vs. structure prediction tools
ProteinMPNN and AlphaFold/ColabFold/Boltz-2 operate on opposite directions of the sequence-structure map. They are not alternatives — they are sequential steps in the same pipeline.
**Use ProteinMPNN** immediately after a backbone generative step (RFdiffusion, BindCraft backbone output, or any designed PDB) to produce candidate sequences. ProteinMPNN takes backbone coordinates as input and returns sequences. It does not evaluate binding.
**Use structure prediction** (AlphaFold, ESMFold, ColabFold, Boltz-2) after ProteinMPNN to validate that the designed sequences actually fold into the intended backbone. The standard self-consistency check is: fold the designed sequence with ESMFold or ColabFold, align the result to the input backbone, and filter on RMSD. Candidates above ~2 Å are discarded.
**Use complex structure prediction** (AlphaFold 3, Boltz-2 in complex mode) as a separate downstream step to evaluate whether the designed binder actually engages the target. ProteinMPNN does not model the binding partner. A sequence that passes self-consistency may still produce a poor binding interface — complex prediction is the filter for that.
The practical pipeline order: RFdiffusion → ProteinMPNN → self-consistency filter → complex prediction → synthesis.
## Tool walkthrough
The interface at [tools.ranomics.com/tools/mpnn](https://tools.ranomics.com/tools/mpnn) requires a free account.
**Step 1: Upload a backbone.** Accepted formats are PDB and mmCIF. The backbone must contain at least one protein chain with full backbone atom records (N, CA, C, O). Side-chain coordinates are not used. If your structure has multiple chains, the tool will design sequences for all chains by default — specify which chains to design if you want to hold some fixed.
**Step 2: Set sampling temperature.** The default is 0.3. For most binder design workflows, running at two temperatures (0.2 and 0.5) and pooling the outputs provides conservative high-confidence predictions alongside a diversity buffer. The tool accepts any value between 0.05 and 1.0.
**Step 3: Set the number of sequences.** The default is 8 per backbone. For production runs feeding into a downstream filter step, 8–16 sequences per backbone is standard. For exploratory work on a single scaffold, 50–100 sequences is reasonable.
**Step 4: Download output.** Results are returned as a FASTA file with one sequence per record, annotated with the backbone filename, temperature, and ProteinMPNN's per-sequence score. The score is a log-probability averaged over all positions — lower (more negative) is better. Sort by score before passing sequences to the next filter step.
## A concrete example from binder design
In a typical RFdiffusion binder design run, 10,000–30,000 backbone structures are generated against a target hotspot. After initial geometry filtering (diffusion confidence, clash score), a subset of 5,000–10,000 backbones advances to sequence design. Running ProteinMPNN at T = 0.2 and T = 0.5 on each backbone, with 8 sequences per backbone per temperature, produces a pool of 80,000–160,000 candidate sequences. That pool is then filtered by self-consistency (ESMFold RMSD), pLDDT on the designed chain, and interface quality from complex prediction, converging to a synthesis set of 500–2,000 sequences.
The ProteinMPNN step in this pipeline takes minutes on CPU, not hours. It is not the computational bottleneck. Running more sequences per backbone is cheap relative to the structure prediction steps that follow.
## Pricing
Each ProteinMPNN run costs around $0.05 of compute on the Ranomics tools hub, depending on backbone size. New accounts get $5 of compute credit on signup, no credit card required. That covers roughly 100 MPNN runs.
After the trial credit, top up your wallet in any amount from $20 and pay per run. See [tools.ranomics.com](https://tools.ranomics.com) for current pricing.
---
**Run ProteinMPNN now:** [tools.ranomics.com/tools/mpnn](https://tools.ranomics.com/tools/mpnn) — $5 free on signup, no credit card.
---
{/* ranomics:related-services */}
## Related Ranomics services
- **[ProteinMPNN tool](/tools/mpnn):** One-click sequence design on the Ranomics tools hub. Around $0.05 per run, $5 free on signup.
- **[Binder Pilot](/binder-pilot):** Full binder design campaign — RFdiffusion + ProteinMPNN + experimental validation. Target structure in, ranked hit list out.
- **[Epitope Scout](https://tools.ranomics.com/scout):** Free surface epitope identification — identify which patches on your target are worth designing against before running any sequence design.
---
## Rational Enzyme Engineering: Structure-Guided Strategies That Work
> Rational enzyme engineering uses structural insight to make targeted mutations. When it beats directed evolution, when it doesn't, and how computational tools sharpen the approach.
Source: https://ranomics.com/rational-enzyme-engineering-strategies-and-methods/
Published: 2026-05-05
import InOurWork from "../../components/InOurWork.astro";
Rational enzyme engineering is the practice of using structural and mechanistic knowledge to select specific residues for mutation, with the goal of altering an enzyme's activity, stability, selectivity, or substrate scope. It contrasts with directed evolution, which explores sequence space by random or semi-random diversification followed by selection — a powerful approach, but one that treats the protein as a black box. Rational design treats it as a mechanism: a three-dimensional machine where function can be tuned if you know which parts to adjust and in which direction.
The approach has a long history and a mixed record. Early rational engineering efforts in the 1980s and 1990s — predating the protein data bank's current scale and the availability of computational design software — produced modest results because structural models were scarce and energetic calculations were unreliable. The field improved as crystal structure deposition accelerated, as Rosetta and FoldX matured, and as AlphaFold changed the availability of high-quality structural models. Today, rational design is a standard first step in any well-scoped [enzyme engineering](/applications/enzyme-engineering) campaign.
## When Rational Design Wins
Rational design is most effective when a clear mechanistic hypothesis connects a specific residue or region to the desired functional change. Three situations consistently favor the rational approach.
### Thermostability via Well-Understood Mechanisms
Thermostability improvements are rational design's strongest track record. The mechanisms are understood: introduction of additional disulfide bonds, proline substitutions in flexible loops (reducing conformational entropy in the unfolded state), surface charge optimization to reduce electrostatic repulsion, and burial of hydrophobic surface area all produce predictable stability gains. Consensus design — identifying the most common residue at each position across a natural sequence family aligned by multiple sequence alignment — is a particularly reliable rational approach. Positions that deviate from consensus in a mesophilic enzyme are candidates for substitution toward the consensus residue. The logic is that natural selection has already sampled these positions across evolutionary timescales, and consensus residues reflect fitness under a broad range of thermal conditions.
The phosphite dehydrogenase case is illustrative: rational introduction of a consensus-guided set of mutations elevated Tm by approximately 20°C while preserving catalytic turnover. That kind of gain in a single design round is difficult to match with random mutagenesis.
### Cofactor Specificity Switching
Many industrial enzymes use NADPH as a hydride donor, but NADH is cheaper and more abundant in fermentation contexts. The cofactor specificity of oxidoreductases is determined largely by a small number of residues in the Rossmann fold that contact the 2'-phosphate of the adenosine ribose. Substituting these positions — typically 2 to 4 residues — switches NADPH-preference to NADH-preference. This is textbook rational engineering: a well-characterized structural motif, predictable geometry, and a simple activity assay to confirm the switch. Ketoreductases (KREDs) in pharmaceutical intermediate synthesis have been engineered this way repeatedly, enabling cost-effective cofactor recycling at manufacturing scale.
### Substrate Scope Adjustment in a Known Active Site
When a crystal structure reveals what the active-site cavity looks like and which residues contact the substrate, substitutions that enlarge or reshape the cavity to accommodate a different substrate are tractable rational targets. Cytochrome P450cam (CYP101A1) is the paradigm case: substitution of Tyr96 and Phe87 — both lining the camphor-binding pocket — altered regioselectivity and expanded the substrate range to include polycyclic aromatic compounds. The Arnold laboratory later extended this logic systematically across the P450 family to enable enantioselective cyclopropanation of olefins by swapping only a handful of active-site residues.
The LovD acyltransferase redesign for simvastatin synthesis is a landmark rational-plus-directed-evolution campaign. Starting from structural knowledge of the natural substrate, rational active-site remodeling enabled LovD to accept a synthetic thioester donor in place of the natural acyl carrier protein. That initial rational hypothesis, confirmed in a few hundred variants, reduced the engineering problem to one that directed evolution could refine efficiently.
## When Rational Design Loses
Rational design fails predictably in two situations.
The first is multi-residue epistasis: when the optimal substitution at position A depends on what amino acid is present at position B, single-residue reasoning breaks down. Epistasis is common in enzyme active sites because residues are geometrically coupled — a cavity-expanding substitution at one position shifts the substrate binding geometry in a way that requires compensatory adjustment at a neighboring residue. Rosetta and FoldX model these interactions imperfectly. In practice, if you need a 5°C thermostability gain, rational design will usually find it in 1 to 3 rounds. If you need a 30°C gain, you almost certainly need to sample epistatic combinations, which requires a screening method with sufficient throughput.
The second failure mode is genuinely novel function — engineering an enzyme to catalyze a reaction it has never performed. The active-site geometry for a new reaction type is not derivable from the existing structure without prohibitive computational effort, and the design success rate is low enough that high-throughput screening becomes essential regardless. In these cases, directed evolution or AI-guided generative design (RFdiffusion, BindCraft, or ProteinMPNN applied to enzyme scaffolds) offers a more tractable path to the starting point, after which rational refinement of the initial hit is productive.
## Computational Tools
### FoldX
FoldX estimates the free energy change (ddG) of a point mutation from a crystal structure using an empirical force field. Computation takes seconds per mutation, making it practical to screen every single amino acid substitution across an entire protein surface in an afternoon. The accuracy is sufficient to distinguish clearly destabilizing substitutions (ddG > +2 kcal/mol) from potentially stabilizing ones (ddG < -0.5 kcal/mol). FoldX is best used as a pre-screening filter: eliminate variants predicted to be strongly destabilizing, prioritize a shortlist of predicted stabilizers for synthesis and characterization.
### Rosetta
Rosetta's enzyme design suite (RosettaDesign, RosettaMatch, and the Enzyme Design application) models mutations in full sidechain rotamer space and evaluates binding geometry for enzyme-substrate complexes. It is slower than FoldX but captures sidechain repacking effects that FoldX's fixed-backbone approximation misses. For active-site redesign, where sidechain geometry determines whether a substrate binds productively, Rosetta provides more reliable predictions. The combination — FoldX for stability screening across the full sequence, Rosetta for active-site geometry refinement on a shortlist — is a standard computational workflow.
### AlphaFold for Structure Provision
AlphaFold2 and its successors have effectively solved the protein structure prediction problem for single-chain enzymes with homologs in PDB. For enzymes without experimental structures, an AlphaFold model is now the standard starting point for rational design. The quality of AlphaFold models in well-predicted regions (pLDDT > 80) is sufficient for FoldX ddG calculations and for visual inspection of active-site geometry. The caveat is that loop regions with low confidence (pLDDT < 70) and complexes involving large conformational changes are still problematic — these should be treated as low-confidence regions for rational design purposes.
## Integrating Rational Design with DMS Validation
The most productive current workflow couples rational design with high-throughput experimental validation rather than treating the two as alternatives. Rational design generates a focused hypothesis — a shortlist of 20 to 200 substitutions predicted to improve the target property. [Deep mutational scanning](/leveraging-ai-deep-mutational-scanning-engineer-enzymes) or a focused site-saturation experiment then tests that hypothesis comprehensively: all 20 amino acids at each targeted position, simultaneously, with NGS readout. The result is a fitness landscape that confirms which computational predictions were correct, reveals unexpected beneficial mutations at the targeted positions, and identifies positions where the computational model was wrong.
This feedback loop serves two purposes. First, it validates hits faster than sequential site-directed mutagenesis — a 50-position site-saturation experiment characterizes 1,000 variants in a single NGS run. Second, it generates training data: the discrepancies between FoldX/Rosetta predictions and experimental outcomes reveal where the computational models fail for a specific enzyme scaffold, improving subsequent predictions.
At Ranomics, this coupling of rational hypothesis generation with DMS-scale experimental validation is standard for enzyme engineering campaigns where a clear structural model is available. Rational design reduces the library size to a tractable scope; experimental validation replaces the uncertainty of computational scoring with measured fitness values. The result is faster iteration and higher confidence in selected variants before moving to scale-up and process validation.
In our enzyme work, once a rational shortlist has been validated experimentally with DMS, the most useful next step is training a small model on the resulting sequence-to-fitness map and using it to propose combinations the original rational hypothesis would not have reached. Off-the-shelf open-source generative models do not work for this because public structural data is too sparse on enzyme fitness. Bespoke per-campaign models trained on the DMS readout do. The rational design narrows the search space, DMS measures fitness across that space, and the custom model extrapolates within the measured manifold to surface combinations a single round of point mutagenesis would have missed.
## Limitations to State Explicitly
No computational tool reliably predicts the effect of mutations in the absence of a high-resolution structure. Homology models built on templates with less than ~40% sequence identity introduce backbone errors that propagate into incorrect ddG estimates. AlphaFold models of disordered regions or multi-domain enzymes with flexible linkers may misrepresent the ground-state geometry. For these cases, the rational design results should be weighted accordingly — use them to generate hypotheses rather than to rank variants with precision.
Solvent effects are also imperfectly modeled. Industrial enzyme engineering frequently targets activity in aqueous-organic co-solvent systems (DMSO, methanol, ethylene glycol), where the implicit-solvent approximations in Rosetta and FoldX are least accurate. Frances Arnold's original subtilisin work in DMF required iterative experimental refinement that no computational model of that era could have guided — and the gap between computational prediction and experimental outcome remains wider in non-aqueous conditions than in standard aqueous buffer.
In our practice, this is why we treat rational computational predictions as triage rather than ranking. The predictions tell us where to look; the experimental DMS readout decides which substitutions actually work. Skipping the experimental validation step on the assumption that a clean ddG ranking is "good enough" is one of the most common reasons rational campaigns underdeliver.
## Rational Design as a Campaign Starting Point
Rational enzyme engineering is not a complete substitute for experimental screening, but it is an efficient first step that reduces the search problem. A well-executed rational design round turns an open-ended mutagenesis campaign into a focused experiment with a testable hypothesis. When the mechanistic logic is sound and a high-quality structure is available, the hit rate on predicted beneficial mutations is meaningfully higher than random mutagenesis — typically 20 to 40% of computationally prioritized variants show improvement versus 1 to 5% for random mutagenesis at equivalent library sizes.
For clients running [enzyme engineering services](/applications/enzyme-engineering) campaigns, the practical implication is reduced iteration time. Rather than committing to 5 to 8 rounds of directed evolution from a starting point with no mechanistic insight, a rational design round can establish which positions matter, which computational predictions hold up experimentally, and where directed evolution or DMS-scale validation is worth the additional investment.
{/* ranomics:related-services */}
## Related Ranomics services
- **[Enzyme engineering](/applications/enzyme-engineering):** Structure-guided rational design coupled with DMS-scale experimental validation for industrial and pharmaceutical enzyme campaigns.
- **[Deep mutational scanning](/technology/deep-mutational-scanning):** High-throughput fitness landscape mapping to validate rational design hypotheses and guide iterative optimization.
---
## RFdiffusion Outputs Need a Developability Check Before Wet Lab
> RFdiffusion, BindCraft, and ProteinMPNN optimize for structure and binding, not for developability. A concrete triage workflow for filtering generative outputs before synthesis or yeast display.
Source: https://ranomics.com/rfdiffusion-developability-check-before-wet-lab/
Published: 2026-04-21
RFdiffusion and BindCraft are very good at the tasks they were trained for. RFdiffusion produces protein backbones that fold, that respect the geometry of a specified hotspot or motif, and that are recognized as realistic by independent structure predictors. BindCraft produces sequence-structure pairs that AlphaFold is confident will fold into a binder geometry with reasonable interface metrics. ProteinMPNN, conditioned on a target backbone, produces sequences that match the fold distribution it learned from the PDB.
None of those three loss functions penalize developability liabilities. That is the problem this post is about.
## What Their Loss Functions Actually Optimize
A short honest read of what the generative stack is doing:
- RFdiffusion is a denoising diffusion model trained on backbone coordinates. Its objective rewards generating backbones that score well under RoseTTAFold and that sit in the distribution of real protein folds. It does not see sidechain identity. It does not see hydrophobicity. It does not see chemical stability.
- ProteinMPNN is an inverse-folding model trained with a sequence-recovery objective on the PDB. Given a backbone, it predicts sequences that the training set's folds tended to carry. It optimizes the distribution of residues that fit a given geometry. It does not optimize for aggregation, deamidation, isomerization, or glycan placement.
- BindCraft, in the common configuration, wraps ProteinMPNN and AlphaFold2 in a binder-biased loop, scoring designs on predicted AlphaFold interface metrics (ipTM, pAE of the interface, designed-chain pLDDT). These are binding-relevant signals, not developability signals.
The result, when you take raw outputs to wet lab, is predictable. Most designs fold. A meaningful fraction bind. A separate and often overlapping fraction carry developability liabilities that will cause problems at expression, purification, concentration, or long-term storage. If the campaign is a first-pass research-grade effort and you are not carrying these binders forward, this may not matter. If the campaign is going to feed anything downstream (affinity maturation, humanization, preclinical scale-up), filtering the raw outputs is non-negotiable.
## What Goes Wrong in Raw Outputs
Specific patterns we see repeatedly in raw RFdiffusion plus ProteinMPNN and raw BindCraft outputs when they are scanned for developability:
**Hydrophobic interface residues bleeding into exterior surface.** The generative stack is trained to build paratopes that make strong contacts with the target. Large hydrophobics (Phe, Trp, Leu, Ile) at the interface are a feature, not a bug. The issue is that these residues often extend past the interface footprint, leaving part of the hydrophobic sidechain exposed on the solvent-facing surface of the design. This produces exactly the HIC retention and PSR polyspecificity signatures that predict downstream CMC trouble, as measured in the Jain et al. 2017 clinical panel (Proc. Natl. Acad. Sci. USA, 114:944 to 949).
**Free cysteines.** ProteinMPNN assigns cysteine at a frequency close to the PDB background rate. In real proteins, most cysteines are in structural disulfides or buried. In raw de novo designs, cysteines often appear on the surface or in unpaired positions, where they drive oxidation, mispaired disulfide scrambling between molecules, and aggregation during purification. This is one of the most common reasons a promising de novo design fails to express cleanly.
**N-glycosylation sequons.** The N-X-S/T sequon occurs purely by chance at some frequency in any ProteinMPNN sequence. If the sequon falls on the surface of the design and the design is expressed in a eukaryotic host, you get variable-occupancy glycosylation that produces a heterogeneous product. This is a particular risk for teams that go directly from a ProteinMPNN output to a mammalian expression test, or that display the design on yeast (yeast N-glycosylation differs from human and can cause additional problems downstream).
**AlphaFold-confident but aggregation-prone designs.** High AlphaFold pLDDT and high ipTM do not imply the design is soluble. pLDDT measures the predictor's confidence in the fold. Aggregation is a thermodynamic property of the ensemble in solution, often driven by short aggregation-prone regions that the structure predictor confidently places in a buried or semi-buried position in the static model, but that become accessible during the folding pathway or in partially unfolded states. This is orthogonal to what AlphaFold measures, and the Sormanni and Vendruscolo CamSol work has explicitly separated these two signals.
**Deamidation and isomerization motifs in exposed loops.** NG, NS, DG appear at their background frequency in raw designs. When they land in a flexible exposed loop (common in BindCraft paratope regions), they are on the fast end of deamidation and isomerization kinetics and will accumulate damage in storage.
None of these failure modes are model bugs. They are consequences of optimizing the wrong objective relative to the shipping criterion.
## Why This Matters More for De Novo Than for Humanized Antibodies
Humanized antibodies carry the evolutionary pressure of the original immune repertoire plus the human germline frameworks they were grafted onto. That pressure is not a developability filter per se, but the sequences involved have been through selection for expression, solubility, and chemical stability at some level (in vivo selection against catastrophically bad sequences, germline conservation across evolutionary time, iterative humanization engineering in industry practice). The result is that humanized antibodies tend to arrive at discovery with a baseline developability profile that is imperfect but not catastrophic.
De novo designs carry no such pressure. The training signal is structural and binding-related, not evolutionary. The generative distribution contains sequences that look nothing like anything that has ever had to be made in a cell and handled at high concentration. Some of those sequences are fine. Some of them are much worse than anything you would find in a naive mammalian immune repertoire. You cannot tell the difference by looking at the design, the AlphaFold ipTM, or the BindCraft score.
The practical conclusion is that the developability filter is a harder requirement for de novo campaigns than it is for humanized programs. The de novo space has no inbuilt floor. Filtering is the floor.
## A Concrete Triage Workflow
What we recommend, for any RFdiffusion or BindCraft campaign that is going to feed anything downstream of first-pass research screening:
1. Generate your N backbone candidates with RFdiffusion, constrained against your target and your chosen hotspot. Epitope selection upstream of this step matters more than any downstream filter, which is a separate post.
2. Sequence the backbones with ProteinMPNN (for RFdiffusion outputs) or use the BindCraft sequence-plus-structure pipeline directly.
3. Run self-consistency filtering first. Fold each design with an independent predictor (ESMFold, ColabFold, Boltz-2) and keep only designs whose predicted structure matches the intended backbone to within a reasonable RMSD. This single step removes 50 to 80 percent of raw output and is the cheapest quality gate you can run. It is a binding-relevance filter, not a developability one, but it has to come first.
4. Run a [developability scan](/technology/developability-scout) on the self-consistency survivors. Specifically, check: exposed hydrophobic surface area, aggregation-prone regions (APR scores from Aggrescan, CamSol profile, or SAP), N-X-S/T glycosylation sequons, free cysteines, NG/NS/NT/DG/DS/DH motifs in exposed loops, and overall charge distribution. Flag or remove the worst offenders.
5. Rank the clean survivors by your actual binding-relevance metric (interface score, ipTM, binder pLDDT, whichever is primary for your campaign). Take the top K forward to synthesis and to display screening.
6. At the display step, the wet-lab selection adds another filter layer. Display screening on [yeast display](/yeast-display) naturally selects against the worst expression and folding failures, so developability problems that a sequence scan missed often still get filtered out experimentally. That is useful, but it is a much more expensive filter than the in silico step. Use the in silico filter first, then let display screening confirm.
The specific thresholds at each step depend on the campaign. For a research-grade tool binder, loose thresholds are fine. For anything that is going to be a lead for a therapeutic program, the thresholds tighten substantially. The Ranomics [AI Binder Sprint](/ai-binder-sprint) pipeline runs this filter by default with thresholds tuned for the therapeutic track.
## What You Get for This Filter
Concretely, running this in silico developability filter before synthesis tends to improve three numbers in a campaign.
Wet-lab hit rate goes up. Self-consistency plus developability filtering raises the fraction of synthesized designs that actually express and display. This is not because developability liabilities cause zero expression directly, but because the filter correlates with "designs that look more like real proteins" and those designs more reliably fold, express, and reach the display surface.
Synthesis budget goes down. If the filter removes 40 percent of an already self-consistency-filtered pool, you save 40 percent of the gene synthesis cost for the next synthesis order. For academic and seed-stage campaigns, this is a material saving. Gene synthesis is usually the largest line item in a binder campaign, larger than GPU compute.
Downstream campaigns inherit cleaner leads. The hits that come back from screening have fewer sequence liabilities, which means fewer red flags during the subsequent affinity maturation or humanization step. Less re-engineering, fewer points of potential regression.
## A Note on What This Does Not Do
The filter does not replace experimental developability confirmation. HIC, DSF, DLS, SEC, thermal stability, and accelerated stability studies are real measurements, and the correlations from in silico scores to wet-lab behavior are strong but imperfect. What the filter does is triage. It removes the designs that are visibly likely to fail before you commit synthesis budget, so that your experimental budget goes to the designs that have a reasonable chance of clearing a full developability panel.
The filter also does not tell you whether the binder will work against your target. Binding is a separate question and is tested experimentally by display screening, SPR, BLI, or a cell-based assay. The filter is a necessary filter, not a sufficient one.
## Where to Start
For DIY campaigns running RFdiffusion, ProteinMPNN, BindCraft, or any hybrid of the three, the fastest way to add a developability filter to your existing pipeline is to pipe the ProteinMPNN or BindCraft sequence output through [Developability Scout](/technology/developability-scout). It is free, browser-based, and produces a scorecard per sequence. It slots in between step 3 (self-consistency) and step 5 (synthesis order) in the workflow above.
For teams that want the full campaign run end to end with developability filtering integrated into the pipeline by default, that is what the [AI Binder Sprint](/ai-binder-sprint) is scoped to deliver. The Sprint includes RFdiffusion plus BindCraft plus ProteinMPNN in parallel, self-consistency filtering, developability filtering, synthesis of the filtered pool, yeast display selection, NGS hit calling, and validation of the top hits. Developability is a gate in that pipeline, not an afterthought.
For grant-scale single-target campaigns where the output is a ranked NGS hit list and the team is running synthesis and display in-house, the [Binder Pilot](/binder-pilot) program is the scoped-down version. Developability Scout is the free upstream piece for either path.
The one recommendation that applies to all three paths: do not take raw RFdiffusion or BindCraft outputs to wet lab. The filter is cheap, and skipping it is the most expensive way to learn what their loss functions do not cover.
{/* ranomics:related-services */}
## Related Ranomics services
- **[Developability Scout](/technology/developability-scout):** Free developability scan for de novo designs and antibody sequences before synthesis.
- **[AI Binder Sprint](/ai-binder-sprint):** Full de novo binder campaign with developability filtering and yeast display selection integrated.
---
## Directed Evolution: A Technical Guide for Protein Stability and Function
> Directed evolution uses iterative cycles of mutagenesis and selection to engineer proteins for stability, activity, and specificity — without requiring structural knowledge. A technical guide to the diversity generation cycle, screening methods, and integration with AI-guided design.
Source: https://ranomics.com/a-technical-guide-to-directed-evolution-for-enhancing-protein-stability-and-function/
Published: 2025-06-30
Directed evolution has matured from a novel academic concept into a transformative [protein engineering](/protein-engineering) technology, representing a paradigm shift in how new biological functions are created and optimized. It is a powerful, forward-engineering process that harnesses the principles of Darwinian evolution — iterative cycles of genetic diversification and selection — within a laboratory setting to tailor proteins for specific, human-defined applications. The profound impact of this approach was formally recognized with the 2018 Nobel Prize in Chemistry, awarded to Frances H. Arnold for her pioneering work that established [directed evolution](/directed-evolution-protein-engineering) as a cornerstone of modern biotechnology and industrial biocatalysis.
The primary strategic advantage of directed evolution lies in its capacity to deliver robust solutions — enhanced stability, novel catalytic activity, or altered substrate specificity — without requiring detailed a priori knowledge of a protein's three-dimensional structure or its catalytic mechanism.
## The Engine of Innovation: The Directed Evolution Cycle
### Principles of Laboratory-Accelerated Evolution
Directed evolution compresses millions of years of natural selection into weeks or months by controlling three variables: the rate and type of mutation, the size of the population, and the stringency of selection.
### Step 1: Generating Genetic Diversity
**Error-Prone PCR (epPCR):** Introduces random point mutations across the entire gene by using a low-fidelity polymerase or altered reaction conditions. The mutation rate is tunable, typically targeting 1-5 mutations per gene per round.
**DNA Shuffling:** Recombines fragments from homologous genes to create chimeric variants that combine beneficial mutations from different parents. This is particularly powerful for exploring epistatic interactions.
**Family Shuffling:** Extends DNA shuffling to multiple parent genes from different species, dramatically expanding the accessible sequence space.
**Site-Saturation Mutagenesis:** Targets specific positions identified by structural or sequence analysis, creating all 20 amino acid substitutions at each site. Combines the benefits of focused diversification with comprehensive coverage.
### Step 2: Linking Genotype to Phenotype
**Plate-Based Screening:** Individual colonies are assayed in microtiter plates. Low throughput (~10^4 variants per round) but compatible with any assay format.
**FACS-Based Screening:** Display technologies link each variant to the cell that encodes it. Throughput of 10^7-10^8 variants per round with quantitative, multi-parameter selection.
**Droplet Microfluidics:** Encapsulates single cells in picoliter droplets, enabling screening of secreted enzymes or intracellular activities at throughputs approaching 10^7 per hour.
## Engineering for Robustness: Enhancing Protein Stability
### The Value Proposition of Stability
A more stable protein tolerates a wider range of conditions (temperature, pH, solvent), has a longer shelf life, and, critically, provides a more robust starting point for further functional engineering.
### The Activity-Stability Trade-off
Mutations that increase stability often reduce activity, and vice versa. This trade-off is not absolute but requires multi-objective selection strategies that maintain function while improving biophysical properties.
### Case Studies in Stability Engineering
**Subtilisin E:** 256-fold improvement in activity in 60% dimethylformamide (DMF) through iterative rounds of epPCR and screening.
**p-nitrobenzyl Esterase:** +14C increase in melting temperature (Tm) while maintaining wild-type catalytic activity, demonstrating that the stability-activity trade-off can be navigated.
**Phosphite Dehydrogenase:** 7000-fold increase in half-life at elevated temperature, transforming a marginally stable enzyme into a robust industrial biocatalyst.
## Engineering for Performance: Improving Protein Function
### Catalytic Efficiency
Directed evolution can push enzymes toward their theoretical catalytic limits. The evolved horseradish peroxidase (HRP) variant achieved 10-fold faster turnover, approaching the diffusion-limited rate.
### Substrate Specificity
The aspartate aminotransferase case study demonstrates the power of directed evolution: a 2.1x10^6-fold switch in substrate specificity from aspartate to valine, effectively converting one enzyme into another.
### Stereoselectivity
Lipase enantioselectivity was improved from E = 1.1 (essentially non-selective) to E = 25.8 through four rounds of epPCR and screening, enabling production of enantiopure products for pharmaceutical synthesis.
## Strategic Implementation
### Directed Evolution vs. Rational Design
Directed evolution requires no structural knowledge but demands high-throughput screening. Rational design requires structural knowledge but can be applied with minimal screening. The approaches are complementary. When high-throughput data from a [deep mutational scanning](/deep-mutational-scanning-a-high-throughput-approach-to-mapping-protein-fitness-landscapes) experiment is available, it can directly inform which positions to target in the next directed evolution round — combining the coverage of DMS with the iterative refinement of directed evolution.
### Semi-Rational Design and "Smart" Libraries
The most effective modern campaigns combine structural insight with evolutionary selection. Computationally designed "smart" libraries focus diversity at positions most likely to yield improvements, reducing library sizes while maintaining functional coverage.
### Challenges
**The Screening Bottleneck:** The diversity of achievable libraries far exceeds the throughput of most screening methods. Library design must be tailored to screening capacity.
**Sequence Space Immensity:** Even a small protein of 100 residues has 20^100 possible sequences. No library can sample more than a vanishing fraction of this space.
**Evolutionary Dead Ends:** Greedy selection for the best variant at each round can trap the search in local optima. Strategies like neutral drift, back-crossing, and family shuffling help escape these traps.
{/* ranomics:related-services */}
## Related Ranomics services
- **[Directed evolution](/directed-evolution-protein-engineering):** Iterative mutagenesis + selection for stability and function.
- **[Enzyme engineering services](/applications/enzyme-engineering):** Directed evolution and DMS for enzyme activity, thermostability, and substrate specificity.
- **[Protein engineering services](/protein-engineering):** Full protein engineering portfolio beyond directed evolution alone.
---
## A Technical Guide to Sorting Strategies in Surface Display
> In any yeast or mammalian surface display campaign, the flow cytometer is your primary selection tool. A guide to gating strategies, antigen titration, off-rate ranking, and counter-screening.
Source: https://ranomics.com/a-technical-guide-to-sorting-strategies-in-surface-display/
Published: 2025-09-04
## The Foundation: Normalizing Binding to Expression
Every [yeast surface display](/yeast-display) or mammalian display sorting experiment begins with a two-color staining approach:
1. **Binding Signal (Y-axis):** Fluorescently-labeled antigen or target molecule
2. **Expression Signal (X-axis):** Fluorescently-labeled detection of the displayed protein using expression tags (c-myc, HA, Flag)
The goal is to target cells in the upper-left quadrant showing the highest binding-to-expression ratio. This normalization is critical. Without it, you are selecting for high-expressing clones, not high-affinity binders.
## Phase 1: Early Rounds (Rounds 1-2), Casting a Wide Net
The objective in early rounds is eliminating non-binders, not identifying the single optimal clone.
**Antigen/Ligand Concentration:** High, often saturating (100 nM or higher). At this stage, you want to capture everything that binds.
**Gating Strategy:** A generous trapezoid or diagonal gate collecting 5-15% of the population. Overly stringent gating at this stage risks losing rare high-affinity clones that happen to be underrepresented.
## Phase 2: Mid-to-Late Rounds (Rounds 3+), Applying the Pressure
**Antigen Concentration:** Decrease below the bulk EC50, forcing competition among binders. Only clones with sufficient affinity to capture the limited antigen will generate signal.
**Gating Strategy:** Progressive tightening to the top "tip" of the binding population, collecting 0.5-2%. This is where affinity-driven selection happens.
## Advanced Strategies
### 1. Off-Rate Ranking (Kinetic Selection)
1. Incubate the library with saturating labeled antigen and wash
2. Add 100x molar excess of unlabeled competitor
3. Incubate 30 minutes to 24 hours
4. Sort the remaining fluorescent cells (those with the slowest off-rate)
This approach selects for kinetic stability of the complex, which often correlates better with in vivo efficacy than equilibrium affinity measurements.
### 2. Specificity and Counter-Screening
1. Label the desired target with one fluorophore (APC/PE, red channel)
2. Label the counter-target with a different fluorophore (FITC, green channel)
3. Incubate the library with both targets simultaneously
4. Collect Red-Positive / Green-Negative cells
This dual-color approach eliminates polyreactive and cross-reactive clones in a single sort, enriching for specificity alongside affinity.
## Closing
Effective sorting is a deliberate process, not a simple hunt for the brightest cells. Each round should have a defined objective, and the gating strategy should be designed to achieve that objective while preserving the diversity needed for subsequent rounds.
For a worked example of a six-cycle sort series with paired pH-7.4 and pH-5.5 selection arms, see our case study on [pH-dependent antibody engineering via yeast surface display](/ph-dependent-antibody-engineering) — round-by-round stringency tuning from a real 640-clone library.
{/* ranomics:related-services */}
## Related Ranomics services
- **[Yeast surface display services](/yeast-display):** FACS and MACS sort strategies tailored to target and library size.
- **[Cell engineering services](/cell-engineering):** Full sort-to-hit pipeline with NGS-based hit calling.
- **[Case study: pH-dependent antibody engineering](/ph-dependent-antibody-engineering):** 14-page technical walkthrough of a real client campaign.
---
## The Numbers Game: Calculating Library Diversity with NGS
> The success of any surface display campaign depends on library quality. A practical framework for NGS-based library validation covering diversity metrics, uniformity assessment, and sequencing workflows.
Source: https://ranomics.com/the-numbers-game-a-practical-guide-to-calculating-and-validate-library-diversity-with-ngs/
Published: 2025-09-08
import InOurWork from "../../components/InOurWork.astro";
In any [yeast display](/yeast-display) or directed evolution campaign, the quality of your starting library is the single most important predictor of success. A diverse, accurately synthesized library contains the raw material for discovery; a biased or poorly constructed one guarantees failure. While traditionally assessed by Sanger sequencing of a few dozen clones, this method offers a dangerously incomplete picture.
Next-Generation Sequencing (NGS) provides a deep, quantitative, and actionable assessment of library quality before you invest weeks of time and resources into a screening campaign. This guide provides a framework for the key metrics and practical steps to validate your library's diversity using NGS.
## Why Sanger Sequencing is No Longer Enough
Sanger sequencing of 20, 40, or even 96 clones provides a qualitative snapshot at best. For a library with a theoretical diversity of millions, this is statistically insignificant. You might confirm that some diversity exists, but you have no quantitative insight into the two most critical parameters:
1. **True Diversity:** How many of your designed variants are actually present?
2. **Distribution:** Are the variants present at relatively uniform frequencies, or is your library dominated by a small number of "jackpot" sequences?
Relying on Sanger alone is like trying to survey a forest by looking at a single tree. You miss the big picture, and critical biases can go completely undetected until it is too late.
## The Core Metrics of an NGS Library QC Analysis
A robust NGS analysis goes beyond simple read counting. It involves calculating specific metrics that together paint a comprehensive picture of library quality.
### 1. Valid Read Percentage (In-Frame and No-Stop)
This is the most fundamental measure of synthesis quality. After sequencing, your bioinformatics pipeline should translate each DNA read into its corresponding amino acid sequence and calculate the percentage of reads that are:
- In the correct reading frame
- Free of premature stop codons
A high-quality library should have a valid read percentage of >70-80%. A low value suggests systemic issues in oligo synthesis or library construction that will severely hamper any downstream selection efforts.
### 2. Observed vs. Theoretical Diversity
This metric directly assesses the complexity of your library.
- **Theoretical Diversity:** The total number of unique variants you intended to create (e.g., for an NNK library at 5 positions, this is 32^5 = ~33.5 million DNA sequences, or 20^5 = 3.2 million amino acid sequences).
- **Observed Diversity:** The number of unique valid sequences detected by NGS.
The ratio of Observed / Theoretical Diversity is a key performance indicator. While achieving 100% is unlikely for very large libraries, a high-quality library should cover a significant fraction of its designed sequence space.
### 3. Library Uniformity: The Impact of Distribution Bias
This is arguably the most powerful, and most overlooked, metric. It measures the evenness of the distribution of variants. A library is not truly diverse if 90% of the population consists of only 10% of the unique variants.
Consider a library with a theoretical diversity of 1,000,000 unique variants:
- **A Good, Uniform Library:** All 1,000,000 variants are present at a similar frequency. If you sequence 10 million total molecules, each unique variant would be represented by approximately 10 reads. This gives every variant a fair chance to be selected based on its fitness.
- **Sub-optimal:** If 90% of your total molecules represent only 50% of your unique variants (500,000), your screen is heavily biased.
- **Poor:** If 90% of your molecules represent only 30% of your unique variants (300,000), the bias becomes severe.
- **Critical Failure:** If 90% of your molecules represent a mere 10% of your unique variants (100,000), the library is functionally useless.
**How This Bias Destroys a Screen:**
1. **Wasted Effort:** In the critical failure scenario, 90% of your screening effort is spent re-evaluating the same over-represented 100,000 clones.
2. **Loss of High-Potential Hits:** The remaining 900,000 unique variants are severely underrepresented or completely absent. The best potential binder may be in this suppressed group.
3. **False Convergence:** Your screen will quickly appear to "converge" on the over-represented clones, not because they are the best binders, but because they dominated the starting population.
In our recent campaigns, we are running NGS on both the binding and non-binding pools out of yeast display sorts. The negative pool is some of the cleanest failure-labelled data anyone has access to, and the cost of the second sequencing arm is small relative to the value of those labelled negatives for any downstream model work. The other readout that matters most in our hands is lineage convergence: clones sharing CDR-H3 or CDR3 sequences across rounds two through four are the campaign signal we trust. Isolated singletons rarely survive secondary validation.
## A Practical Workflow for NGS Validation
1. **Sample Preparation:** Take a representative sample of your plasmid DNA library before transformation. This ensures you are measuring the quality of the library itself, not any growth biases from biological amplification.
2. **Amplicon Generation:** Use a high-fidelity DNA polymerase to amplify the variable region. Use the minimum number of PCR cycles required (typically 10-15 cycles) to add sequencing adapters. Over-amplification is a primary source of introducing bias.
3. **Sequencing:** Use an Illumina platform (MiSeq or NextSeq) with paired-end reads long enough to fully cover your diversified region. Aim for a read depth that is at least 100x your theoretical diversity to ensure you capture low-frequency variants.
4. **Bioinformatic Analysis:** Process raw sequencing data to calculate the key metrics:
- Merging paired-end reads
- Filtering out low-quality reads
- Identifying and trimming constant flanking regions
- Translating DNA sequences to amino acid sequences
- Counting unique variants and their frequencies
## Advanced Workflow: Handling Diversified Regions Longer Than a Single Read
The workflow above is ideal for libraries where the diversified region is less than ~500 bp, easily spanned by a 2x300 bp Illumina run. For longer regions (e.g., a full-length scFv at ~750 bp), two approaches are available:
### 1. Tiling Amplicons with Short-Read Sequencing
Design multiple primer sets to generate overlapping amplicons that "tile" across the entire diversified region. For an scFv, you might have one amplicon for the VH domain and another for the VL domain.
**Pros:** Cost-effective, high-throughput, and uses well-established bioinformatic pipelines. Provides excellent quantitative data on diversity and uniformity within each sub-region.
**Cons:** You lose linkage information. You cannot determine which specific VH is paired with which specific VL in the original, full-length molecule.
### 2. Full-Length Analysis with Long-Read Sequencing
Use a long-read sequencing technology like PacBio HiFi or Oxford Nanopore. These platforms generate reads thousands of base pairs long, easily covering full-length constructs.
**Pros:** Preserves complete linkage information. You get the exact, full-length sequence for every variant molecule, the gold standard for library QC. PacBio HiFi accuracy is now on par with short-read methods.
**Cons:** Higher cost-per-read and lower throughput than Illumina. Technically feasible primarily with lower diversity libraries.
## Conclusion: An Essential Investment
NGS-based library validation is an essential quality control step for any serious screening campaign. It provides a quantitative baseline of your starting population, allowing you to diagnose problems early and interpret downstream enrichment data with confidence. Investing in a deep sequencing analysis before you begin screening can save you months of troubleshooting and is the first, and most important, step towards a successful discovery.
In our work, the metric that has saved us the most rounds of re-screening is uniformity rather than headline diversity. A library with 60% observed-to-theoretical at high uniformity outperforms a library with 95% observed-to-theoretical concentrated in 10% of clones, every time. The former gives every variant a fair shot; the latter just amplifies whatever the synthesis vendor happened to be good at.
{/* ranomics:related-services */}
## Related Ranomics services
- **[NGS analysis](/technology/ngs-analysis):** Library diversity QC, uniformity, and enrichment tracking.
- **[Variant library construction](/technology/variant-library-construction):** NGS-validated libraries built to the diversity and uniformity you spec.
---
## Troubleshooting Low Display Levels in Yeast and Mammalian Cells
> Low or non-existent display levels are a common roadblock in yeast and mammalian display campaigns. A systematic diagnostic checklist for identifying and resolving the root cause.
Source: https://ranomics.com/troubleshooting-low-display-levels-in-yeast-and-mammalian-cells-a-step-by-step-checklist/
Published: 2025-10-02
No fluorescence signal on your flow cytometer is never a good start after weeks of preparation. Low or non-existent display levels are a common roadblock that can bring a project to a screeching halt.
## The First Diagnostic: Is the Problem Global or Clone-Specific?
**Global Problem:** When unselected libraries and controls show low display, the issue likely lies with your core system: the vector, the host cells, or the experimental protocol.
**Clone-Specific Problem:** When positive controls display normally but certain clones show declining levels after selection rounds, the issue is almost certainly with the protein variants themselves.
## Path 1: Troubleshooting Global Low Display (System-Wide Issues)
### Step 1: Interrogate Your Plasmid Construct
- Is the correct promoter present? (GAL1 for yeast, CMV for mammalian)
- Is the signal peptide appropriate and in-frame?
- Has library cloning been confirmed in the correct reading frame?
- Is the surface anchor present (Aga2p for yeast, transmembrane domains for mammalian) and in-frame?
- Has codon optimization been performed for the host organism?
### Step 2: Scrutinize Your Host Cells
**Mammalian considerations:**
- Test for mycoplasma contamination, a silent killer of mammalian cell experiments
- Verify cell viability and confirm cells are in exponential growth phase
**Yeast considerations:**
- Confirm strain compatibility with selection markers
- Assess cell health microscopically
### Step 3: Audit Your Protocol
**Induction (Yeast):** Are you inducing in a galactose-containing medium (e.g., SG-CAA) and ensuring there is absolutely no glucose, which represses the GAL1 promoter? Temperature optimization (often 20C is better than 30C) and duration (16-24 hours is typical) are critical variables.
**Transfection (Mammalian):**
- Quantify transfection efficiency using GFP co-transfection
- Optimize DNA quantity and reagent-to-DNA ratio
- Adjust cell density at transfection
- Determine optimal harvest timing (typically 24-48 hours)
## Path 2: Troubleshooting Clone-Specific Low Display (A Developability Problem)
### Step 4: Analyze the Protein Variant Itself
Low display for a specific clone is often a result of the cell's own quality control machinery. Root causes include:
- **Intrinsic Instability:** The variant may have a low melting temperature (Tm) and be inherently unstable.
- **Exposed Hydrophobic Patches:** Mutations can expose hydrophobic regions, leading to aggregation within the secretory pathway.
- **Unpaired Cysteines:** The introduction of an odd number of cysteine residues can lead to improper disulfide bonding, misfolding, and aggregation.
- **Toxicity:** The protein variant itself might be toxic to the host cell, leading to reduced growth and protein synthesis.
## Conclusion: Low Display is a Feature, Not Just a Bug
While frustrating, low display levels are a critical source of information.
For clone-specific problems, low display is an early and invaluable filter for poor developability, allowing you to eliminate problematic candidates long before you invest significant time and resources in downstream characterization.
{/* ranomics:related-services */}
## Related Ranomics services
- **[Yeast surface display services](/yeast-display):** Full diagnostic + rebuild support for failed display campaigns.
- **[Mammalian display](/technology/mammalian-display):** Alternative platform for targets that refuse to display on yeast.
---
## The Two-Platform Approach: Using Yeast Display for Affinity and Mammalian Display for Developability
> Don't choose between yeast display and mammalian display. This guide details a two-platform biologics discovery workflow, using yeast for affinity and mammalian display to screen for developability.
Source: https://ranomics.com/the-two-platform-approach-using-yeast-display-for-affinity-and-mammalian-display-for-developability/
Published: 2025-11-04
You've spent months on your yeast display campaign. You identified a panel of high-affinity binders. But when you moved your top candidate into mammalian expression for characterization, the results were disappointing: poor protein expression and 50% aggregation.
For years, the field has debated **yeast display** vs. **mammalian display** as if they were competitors. The truth is, they are specialized partners.
Use yeast for what it is best at: raw discovery. Then, use mammalian for what it is best at: validation and developability.
## Yeast Display as the Discovery Engine
**Massive Numbers:** Yeast display libraries routinely achieve 10^7-10^9 diversity, providing exceptional statistical power for identifying rare binding events.
**Robustness and Speed:** Fast growth, easy handling, stable and uniform display make yeast the workhorse for iterative rounds of selection.
**Pure Affinity Readout:** The yeast system provides a clean selection focused on binding affinity and specificity, without the confounding effects of mammalian-specific post-translational modifications.
The recommended approach: 3-4 rounds of high-stringency sorting, then use NGS to identify the entire enriched polyclonal pool rather than cherry-picking individual clones.
## Mammalian Display as the "Developability Matrix"
The question shifts from "Can it bind?" to "Of these 10,000 great binders, which ones **can actually be manufactured?**"
Yeast cannot determine: correct folding in CHO cells, human-like post-translational modifications and glycosylation patterns, or stability at high concentrations required for therapeutic formulation.
### Four-Step Workflow
1. **Create a Focused "Hits Library".** Synthesize 1,000-10,000 hits identified from yeast display NGS data as a focused sub-library.
2. **Screen in Mammalian Display.** Transfect the focused library into HEK or CHO cells using CRISPR integration or lentiviral delivery for single-copy, uniform expression.
3. **Run the "Developability Sort".** Stain cells with antibodies against expression tags (c-myc or HA) to quantify surface display levels in the mammalian context.
4. **Sort for the Brightest Cells.** Collect the top 5-10% with the highest expression signal.
This simple sort is what I call "Developability Matrixing." You are running your panel of high-affinity hits through a matrix that filters for one thing: **recombinant protein expression** and stability in a manufacturing-relevant host.
## The Payoff
A candidate with high affinity in yeast and high expression in mammalian display represents prospective selection for:
- **High function** (from yeast display screening)
- **High developability** (from mammalian display filtering)
This approach "shifts left" on developability, moving critical late-stage failure points to early discovery phases where they cost orders of magnitude less to address.
{/* ranomics:related-services */}
## Related Ranomics services
- **[Yeast surface display services](/yeast-display):** Affinity-focused primary screening on yeast.
- **[Mammalian display](/technology/mammalian-display):** Developability-focused secondary screening with proper post-translational modifications.
---
## what-is-de-novo-protein-design.mdx
Source: https://ranomics.com/what-is-de-novo-protein-design/
import InOurWork from "../../components/InOurWork.astro";
import KeyTakeaway from "../../components/KeyTakeaway.astro";
We run de novo binder design campaigns every week at Ranomics. The pipeline has matured to the point where generating thousands of structurally plausible binder candidates is routine. The hard part is no longer computational. The hard part is confirming that any of them actually bind.
This post describes how a modern de novo design campaign works end to end, from target structure through experimental screening, and where the real failure modes live.
De novo protein design generates binders computationally from a target structure rather than from a natural scaffold. Modern pipelines can produce thousands of plausible candidates routinely; the hard part is confirming they bind. The bottleneck is experimental validation, not computation, so a campaign succeeds or fails on its display-screening infrastructure.
## The design layer: three generators running in parallel
A de novo campaign starts with a target structure and a defined binding surface. We specify hotspot residues on the target, the residues we want the designed binder to engage, and run three generators in parallel.
**RFdiffusion** generates protein backbones by reversing a learned noise process, producing novel scaffolds that are geometrically complementary to the specified hotspot. It does not start from any known protein. The outputs are backbone coordinate sets, no sequences yet.
**BindCraft** takes a different approach: it jointly optimizes backbone geometry and sequence in a single pass, using AlphaFold2 as an internal scoring function. BindCraft generates complete binder candidates (structure and sequence together) and applies its own filters for predicted binding affinity and structural confidence.
**Boltzgen** samples from a Boltzmann distribution over protein conformations, producing diverse backbone geometries that satisfy the target interface constraints through a generative flow matching approach.
---
**Identify epitopes on your own target.** [Epitope Scout](https://scout.ranomics.com) scores and ranks surface patches on any PDB structure. Free to use.
---
Running all three in parallel is deliberate. Each generator explores structure space differently. RFdiffusion excels at producing compact, high confidence scaffolds. BindCraft tends to find solutions that score well on AlphaFold metrics but may explore different topology space. Boltzgen adds conformational diversity that neither of the other two reliably captures. The union of their outputs gives broader coverage of viable binding modes than any single method alone.
## Sequence design: ProteinMPNN for backbone-only generators
RFdiffusion and Boltzgen produce backbones without sequences. Those backbones need sequences that will fold into the intended geometry and present the correct interface residues. ProteinMPNN handles this step: given a backbone, it predicts amino acid sequences optimized for structural stability and foldability.
We typically generate 8 to 16 sequences per backbone, then filter on ProteinMPNN confidence scores before advancing candidates. BindCraft skips this step entirely because it designs sequence and structure jointly.
## Structural validation: filtering before synthesis
Before any candidate reaches a gene synthesis order, it passes through computational structural validation. We run designed sequences through Boltz-2, ESMFold, and ColabFold to predict their folded structures independently of the design model that generated them.
The key metric is whether the predicted structure matches the designed structure. If a sequence was designed to fold into a four-helix bundle that engages the target through a specific interface, the predicted structure should reproduce that geometry. We measure this as backbone RMSD between the design model and the predicted structure, and we check that the predicted interface contacts match the designed ones.
Candidates that fail structural validation, those where the predicted fold deviates significantly from the design, are discarded before synthesis. This step eliminates 60% to 80% of initial candidates depending on the target. It is the single most cost-effective filter in the pipeline.
## Experimental screening: where campaigns succeed or fail
The survivors of computational filtering are synthesized and screened experimentally. At Ranomics, this means yeast surface display as the primary screening platform, with mammalian display for candidates that require post-translational modifications or where yeast expression is problematic.
Yeast display lets us screen thousands of candidates in parallel against labeled target protein, sorting for binding by FACS. The throughput is high enough to test the full computationally filtered set in a single campaign. First-round hit rates for well-designed campaigns typically fall in the 5% to 15% range, meaning 5% to 15% of synthesized candidates show detectable binding to the target.
For context, the best published hit rates from major industry labs running large-scale binder campaigns sit around 2%. The reason our range comes in higher in our work is aggressive computational filtering before any gene is ordered, and screening on yeast display rather than relying on ELISA or SPR alone, which rejects polyspecific and non-expressing designs early instead of paying for them downstream.
That number is the real metric of a de novo design campaign. Not how many backbones were generated, not how many passed computational filters, but how many confirmed binders came out of experimental screening.
## What determines success vs. failure
After running dozens of these campaigns, the patterns are clear.
**Target surface quality matters more than generator choice.** Flat, featureless surfaces are harder to design binders against than concave pockets or surfaces with prominent loops. No amount of computational sampling compensates for a target that presents minimal geometric features for a binder to grip.
**Hotspot selection is the highest leverage decision.** The residues you designate as the binding interface on the target constrain everything downstream. Choosing hotspots that are solvent-accessible, structurally rigid, and chemically diverse produces better candidates than choosing hotspots based solely on biological relevance.
**Computational filtering must be aggressive.** The temptation is to advance marginal candidates because synthesis is "only" a few hundred dollars per gene. In practice, every candidate that enters experimental screening consumes assay capacity. Tight structural validation cutoffs save weeks of wet-lab time.
**The team running computation must also run the experiments.** When design and validation are separated across organizations, the feedback loop breaks. If a campaign yields zero hits, understanding whether the failure was in hotspot selection, backbone sampling, sequence design, or expression requires access to both the computational parameters and the experimental data. We keep both under one roof for this reason.
## The bottleneck is experimental, and that changes the economics
Five years ago, the bottleneck in protein design was computation: generating enough plausible candidates was slow and expensive. Today, a single cloud GPU run produces more candidates than any lab can screen in a quarter. The constraint has flipped.
This means the economic value in de novo design is not in running the algorithms. It is in the experimental infrastructure to validate what the algorithms produce, and in the expertise to close the loop between computational predictions and biochemical reality.
De novo protein design works. It produces binders for targets where no natural scaffold exists and no library would reach. The question is no longer whether the approach is viable. The question is whether your validation pipeline can keep pace with what the generators produce.
AI design has matured beyond therapeutic antibodies. In our practice, we see strong fit for custom lab reagents, biomarker capture, diagnostic antibody pairings, ag-bio binders, and even cosmetics applications. The lower-regulation environments are some of the highest-leverage early adopters because the same pipeline that scopes a therapeutic program can deliver a reagent-grade binder in a fraction of the time and budget. The constraint shifts from "is this technology ready" to "what problem do you want to point it at."
[Start a project](/ranomics-contact)
{/* ranomics:related-services */}
## Related Ranomics services
- **[AI protein binder design](/ai-protein-binder-design):** De novo binders designed and validated end-to-end.
- **[Binder Pilot](/binder-pilot):** Starter program for first-time de novo design campaigns.
---
## when-to-use-epitope-scout-vs-a-structural-biologist.mdx
Source: https://ranomics.com/when-to-use-epitope-scout-vs-a-structural-biologist/
import InOurWork from "../../components/InOurWork.astro";
If you are running a de novo binder design campaign, the single most consequential decision is which surface on your target to direct the design against. In a well-resourced biopharma pipeline, this decision is made by a structural biologist looking at the target in PyMOL, cross-referencing the PPI literature, and integrating experimental context. In a small team without that expertise, it is often made by looking at the structure in a browser viewer and picking a spot that "looks nice." The gap between those two processes is where most first-time binder campaigns lose their budget.
[Epitope Scout](/technology/epitope-scout) is designed to automate the reproducible, literature-backed parts of that decision. It is not a replacement for a structural biologist on every project — but it is enough for most projects, and it is strictly better than eyeballing.
This post is about when it is enough and when it is not.
## What Epitope Scout Does
Epitope Scout takes a target structure (PDB or mmCIF, experimental or AlphaFold) and returns a ranked list of candidate surface epitope patches, each scored on a composite of five structural properties:
1. Hydrophobic exposure (weighted 30%)
2. Structural order — beta-strands and helices vs loops (20%)
3. Rigidity — B-factor or pLDDT (20%)
4. Surface accessibility (15%)
5. Hot-spot residue density — Trp, Tyr, Arg, Phe (15%)
These are the five criteria most consistently identified in the protein-protein interaction literature as predictive of bindable surface. The scoring is fully automated, runs in minutes on a standard PDB file, and produces an output ready to pass directly into RFdiffusion as a hotspot specification.
What Epitope Scout also adds beyond the composite score:
- **PPI interface detection.** Flags patches that overlap with natural protein-protein interaction interfaces present in the structure.
- **Known binder lookup.** Queries SAbDab and RCSB for existing antibody and nanobody structures against the target, showing which residues those binders contact.
- **Quality flags.** Non-scoring annotations for potential concerns: loop-only anchors, all-polar patches, glycan proximity, high local flexibility, electrostatic asymmetry.
This is the kind of triage a structural biologist would do in the first 30-60 minutes of looking at a target. Automating it means it happens consistently, reproducibly, and before any compute budget is spent.
In our work, the single biggest lift Epitope Scout gives us is forcing the campaign to sample multiple epitopes rather than committing to one. Even with strong domain expertise, picking a single epitope and running tens of thousands of designs against it loses against running smaller campaigns against three or four ranked patches. The chance of finding a binder scales with epitope diversity, not with depth on any single surface, and an automated ranking with quality flags is the cheapest way to get to that diversity.
## When Epitope Scout Is Enough
For most first-time binder design campaigns, Epitope Scout alone is sufficient for hotspot selection. Specifically:
- **Soluble extracellular targets with a known or well-predicted structure.** The five scoring criteria are well-validated for this class of target.
- **Targets where the downstream binder just needs to bind, not to compete with a specific partner.** If you are generating an affinity reagent, a tool binder, or a research-grade ligand for a detection assay, a high-scoring epitope patch is usually a valid choice regardless of its biological role.
- **Targets with existing structural data and a literature base.** The PPI interface detection and known binder lookup surface the context that would take a human several hours to compile manually.
- **Teams without in-house structural biology expertise.** In this case the alternative is not a structural biologist — it is an untrained picker. Epitope Scout is strictly better.
- **Early-stage scoping.** Before committing to a target, Epitope Scout can tell you in five minutes whether any part of the target is even plausibly bindable. That is useful triage.
## When You Need a Structural Biologist in the Loop
There are several classes of target and project where automated scoring alone is not sufficient, and a structural biologist (or a protein engineer with deep target-area expertise) needs to make the call.
**Allosteric or mechanism-dependent binding.** If the goal is not just to bind the target but to inhibit it, activate it, or modulate a specific conformation, the epitope choice is constrained by mechanism. Epitope Scout will rank patches by bindability; it does not know that a specific surface is the catalytic site, the dimerization interface, or the allosteric switch. A structural biologist familiar with the target family needs to translate "I want to block substrate entry" into a residue-level selection.
**Conformational ensembles and cryptic pockets.** If the target has multiple functionally relevant conformations (active/inactive states, open/closed forms, induced-fit pockets), picking an epitope from a single static structure can be misleading. A structural biologist familiar with the target's conformational landscape is needed to decide which state to design against.
**Membrane proteins and complex topologies.** GPCRs, transporters, and multi-pass membrane proteins have accessibility constraints that Epitope Scout does not fully model. A human with domain expertise needs to decide whether a patch is exposed in the membrane context, which detergent or nanodisc conditions will preserve it, and how the display system will present it.
**Targets with known failed campaigns.** If the target has a history of failed binder attempts in the literature, the specific reasons for failure matter. Sometimes the target is fundamentally hard (glycan shield, fast conformational dynamics, low surface hydrophobicity); sometimes prior attempts targeted the wrong surface. A structural biologist can read that literature and reconcile it against a new design strategy.
**Therapeutic epitope selection.** For clinical or preclinical antibody programs, epitope choice has to account for developability, immunogenicity risk, cross-reactivity with homologous human proteins, and patent freedom-to-operate. None of that is structural — it is a strategic decision that an automated tool will not make for you.
**Post-translational modification dependency.** If the native binding partners of the target recognize a phosphorylated, glycosylated, acetylated, or otherwise modified version of the surface, the PTM state of your recombinant target during display matters. A structural biologist or target-area expert needs to confirm that the displayed target carries (or lacks) the relevant modifications.
## A Checklist for Your Target
To decide which side you are on, ask:
- Is the target soluble, extracellular, and structurally well-defined? → Epitope Scout is likely enough.
- Is the target a GPCR, ion channel, transporter, or multi-pass membrane protein? → Get a structural biologist involved.
- Is mechanism (inhibition, activation, allostery) a required property of the binder? → Get a structural biologist involved.
- Is the target heavily glycosylated or PTM-dependent? → Get a structural biologist involved.
- Is this a first-in-class binder against a target with no prior structural validation? → Epitope Scout for initial triage, then human review before synthesis.
- Is this an affinity reagent, detection tool, or research-grade binder where bindability alone is the criterion? → Epitope Scout is almost certainly enough.
- Are you preparing a campaign for a therapeutic program with IP, immunogenicity, or clinical considerations? → Get a structural biologist, a regulatory consultant, and an IP attorney involved. None of those decisions are automatable.
## The Honest Answer
For most first binder campaigns that small biotech and academic teams run, Epitope Scout is the right tool and a structural biologist is not available anyway. The practical upgrade path is:
1. Run Epitope Scout on your target. Get the ranked patches and the quality flags.
2. If everything looks green — high-scoring patches, no red-flag quality issues, clear separation between top and bottom patches — move into a binder design campaign.
3. If the quality flags raise real concerns (loop-only anchors, glycan proximity, PTM dependency suspected, no clear winning patch), that is the signal to bring in a structural biologist for a one-off consultation before committing compute and synthesis budget.
The goal is not to automate the structural biology. It is to automate the parts that should be automated, so that when you bring in a human expert, their time goes to the decisions that actually need human judgment. In our practice, that means using the ranked patches to plan a multi-epitope campaign by default, and reserving the structural biologist's time for the design and mechanistic decisions that the ranking output cannot resolve.
If you want to try this on your own target: Epitope Scout is [free to use](https://scout.ranomics.com), no credit card, no NDA. If you decide to move into a campaign afterward, the [Binder Pilot](/binder-pilot) and [AI Binder Sprint](/ai-binder-sprint) programs both accept Epitope Scout output directly as the hotspot specification.
{/* ranomics:related-services */}
## Related Ranomics services
- **[Epitope Scout](/technology/epitope-scout):** Free surface epitope identification at scout.ranomics.com.
- **[AI Binder Sprint](/ai-binder-sprint):** Full structural review + de novo binder design when Epitope Scout alone isn't enough.
---
## yeast-display-antibody-discovery.mdx
Source: https://ranomics.com/yeast-display-antibody-discovery/
import TryToolCallout from "../../components/TryToolCallout.astro";
import InOurWork from "../../components/InOurWork.astro";
Yeast display for antibody discovery has become the default discovery platform for campaigns where the input library is smaller than 10^9 variants and the readout needs to be quantitative on a per-clone basis. Phage display still wins for raw diversity; mammalian display still wins for full-length IgG and complex PTM dependence. Between them, yeast display occupies the practical center of the discovery workflow — and for AI-designed libraries especially, it is the platform that pulls the most useful information out of every clone screened.
This article is a methods primer for the choices that matter: which scaffold to display, how to build the library, how to sort, how to call hits from NGS, and how to triage candidates before committing to expression scale-up.
## Why yeast display for antibody discovery
The structural argument is direct. Yeast surface display anchors antibody fragments to the cell wall via Aga2p fusion, presents the fragment at 10,000–100,000 copies per cell, and labels each cell with two independent fluorophores — one for display level (anti-myc or anti-HA on the C-terminal tag), one for antigen binding. Flow cytometry then reports both signals per cell, which means every clone passing through the sorter delivers a normalized affinity estimate, not just a binary "binds / doesn't bind" call.
That normalization is the part that matters. Phage display enrichment reports population-level changes in clone abundance after panning; you learn which sequences enriched, but not their relative affinities without follow-up titrations. Yeast display reports each clone's mean fluorescence intensity ratio, which is a quantitative proxy for Kd within the linear range of the assay. For ranking the top 100 hits from a 10^4-hit pool, yeast saves a quarter of validation work.
A second argument is folding fidelity. Yeast — like mammalian cells but unlike *E. coli* — has the secretory pathway, BiP and PDI for disulfide chaperoning, and quality control that aborts misfolded proteins before surface display. The displayed fraction of a yeast library is the fraction that actually folded. Phage doesn't enforce this filter; bacterial periplasmic chaperoning is a weaker quality gate, and aggregation-prone clones can still display.
It is easy to inflate apparent expression on yeast by fusing a solubility tag like MBP or SUMO to a struggling construct. The display signal looks great. Then the same sequence is reformatted for standalone production downstream, the tag is gone, and the protein no longer expresses. In our work, we run display constructs without rescue tags so that the signal in the sort actually predicts what we can produce in the next stage of the campaign.
## Scaffold choices: scFv, Fab, IgG, VHH
The four common antibody-derived scaffolds present trade-offs you commit to at the library-design stage.
**scFv** (single-chain Fv) is the workhorse for yeast display. The VH and VL chains are joined by a flexible linker (typically (Gly₄Ser)₃), expressed as a single polypeptide, and display efficiently. The downside: scFvs are prone to thermal instability and aggregation, and the linker can interfere with the binding mode when the antibody must engage a specific epitope geometry. For early discovery, scFv format is the cheapest and most informative. For lead optimization beyond the discovery hit, conversion to Fab or IgG is required.
**Fab** (antigen-binding fragment) displays both chains independently. One chain is fused to Aga2p; the other expresses cytoplasmically and assembles in the secretory pathway. Fab displays correlate better with full IgG binding mode than scFv, and the format is closer to a developable lead. The trade-off is library-construction complexity: you have two chains to diversify, which means combinatorial libraries grow combinatorially.
**IgG** in yeast is not practical at library scale. The molecule is too large, and assembly of the four chains in the yeast secretory pathway is inefficient. For full IgG discovery, mammalian display is the right platform.
**VHH** (single-domain antibodies from camelids) is the cleanest yeast display scaffold by far. One domain, one chain, ~110 residues, no light chain to coordinate, no linker artifact, intrinsically stable. Library sizes of 10^9 unique VHHs are routine, and the format delivers fully developable leads with minimal reformatting. For programs where the IgG-format constraint isn't load-bearing (intracellular targets, multivalent constructs, bispecifics built on a VHH scaffold), VHH-on-yeast is the platform of choice.
We default to scFv or VHH for discovery depending on the downstream format. Fab when full-length IgG developability is a near-term concern.
## Library construction: natural, synthetic, AI-designed
The three library-construction strategies trade off diversity, design control, and cost. We covered the comparison in detail in [natural vs synthetic vs AI-designed libraries](/natural-synthetic-ai-designed-libraries-antibody-discovery); the summary applicable to yeast display:
- **Natural libraries** (cDNA amplified from immunized animals or naïve donor B cells) deliver evolved diversity at low cost. Library sizes reach 10^9 in yeast. The trade-off is no design control: you get what the immune repertoire produced, including liabilities (deamidation hotspots, oxidation-prone methionines, aggregation-prone CDR-H3 sequences).
- **Synthetic libraries** (germline frameworks plus combinatorial CDR diversification using NNK, NNS, or trimer codons) deliver designed diversity with controllable framework choice and CDR length distribution. Library sizes are also 10^9. Costs are higher but liabilities can be engineered out at the design stage.
- **AI-designed libraries** (ProteinMPNN-resampled, RFdiffusion-conditioned, BindCraft-generated) deliver pre-filtered diversity concentrated near the target's binding mode. Library sizes are 10^4–10^6 — orders of magnitude smaller — but hit rates per clone are correspondingly higher.
For naïve discovery against a new target, we run synthetic-on-yeast as the default. For targets where AI-designed candidates exist, we run AI-on-yeast as a second arm and union the hits.
## Sorting: MACS pre-enrichment, then FACS rounds
A 10^9 yeast library cannot be FACS-sorted in one pass. Flow rates cap at ~10^7–10^8 cells per hour on a high-end sorter, and the volume of buffer to handle 10^9 cells is impractical. The solution is [magnetic-activated cell sorting (MACS) pre-enrichment](/beyond-facs-an-introduction-to-magnetic-activated-cell-sorting-macs-for-library-pre-enrichment) before the first FACS round.
The pattern:
1. **MACS round 1**: incubate 10^9 yeast with biotinylated antigen at ~100 nM; capture on streptavidin microbeads. Yield: ~10^7 enriched cells. This single step removes the 99% of the library that is non-displaying or non-binding.
2. **FACS round 1**: stain with biotinylated antigen at ~10 nM plus anti-myc-PE for display normalization. Sort the top ~1% of double-positive cells. Yield: ~10^5 enriched cells.
3. **FACS rounds 2–4**: decrease antigen concentration each round (1 nM, 100 pM, 10 pM as the campaign proceeds). Each round tightens selection stringency. Yield: ~10^3–10^4 enriched cells per round.
Total wall-clock: 3–4 weeks for a complete campaign. Total reagent burden: ~50 μg biotinylated antigen if titrations are conservative. Across rounds, the population shifts from 0.1% antigen-positive to >50% antigen-positive — a clear sorting signature.
The trick that matters: **titrate display levels** before staining, especially in late rounds. As selection narrows the population, the surviving clones may differ in display level by 5–10×, which inflates apparent affinity differences. Gating on a fixed display-level window normalizes this. We documented the protocol in [correctly titrating display levels](/correctly-titrating-display-levels-for-reliable-affinity-data-in-yeast-and-mammalian-systems).
A second practical note from our campaigns: for a target or scaffold we have not screened before, our first pass is to incubate overnight. Most published protocols specify one to two hours, but the time a new yeast display assay takes to reach equilibrium is not always the time the protocol guide assumes. Overnight is a safe upper bound, and we titrate the incubation time back from there once we have a reliable signal.
## NGS hit calling
After each FACS round, plasmid is recovered from the sorted population, the antibody coding region is PCR-amplified with NGS-compatible primers, and the population is sequenced on Illumina (typically 1–5 million reads per round per population). The output is a per-clone abundance trace across rounds.
What you look for:
- **Enrichment factor**: clones that increase ≥10× from round 2 to round 4 are the lineage of interest. Background clones (carry-over non-binders, contamination) typically stay flat or decrease.
- **Lineage convergence**: clones with shared CDR-H3 sequences (or VHH CDR3) are evidence the campaign is converging on a productive solution rather than enriching noise.
- **Read-depth sufficiency**: clones below 100 reads in the final round are statistically unreliable. Either re-sort or skip them.
For the full math on what library diversity and read depth deliver, see [the numbers game on NGS library diversity](/the-numbers-game-a-practical-guide-to-calculating-and-validate-library-diversity-with-ngs).
The output is typically 50–200 unique CDR sequences that meet enrichment and read-depth criteria. From there, you triage to the top 20–50 for biochemical follow-up.
In our recent campaigns, we are increasingly running the same NGS step on the non-binding pool too. The training datasets behind the public AI design models skew heavily positive, and the negative pool from a sorted yeast library is some of the cleanest failure-labelled data anyone has access to. The marginal sequencing cost is small relative to the value of those labelled negatives for any downstream model work.
## From hit to lead
The top NGS hits are reformatted to soluble Fab or IgG, expressed at small scale (typically 1–10 mL in HEK293 or CHO transient expression), and characterized for:
- **Binding by BLI or SPR**: confirm Kd is in the expected range (typically 0.1–10 nM for a successful campaign).
- **Specificity**: counter-screen against off-target proteins (related family members, serum albumin) and against PSR (polyspecificity reagent).
- **Thermal stability**: nanoDSF Tm. Below 65 °C is a yellow flag for downstream developability.
- **Hydrodynamic behavior**: SEC profile. Aggregation, dimers, fragments — all visible in this single readout.
- **Production yield**: mg/L from the small-scale prep. Below 10 mg/L predicts difficult scale-up.
Hits passing all five criteria advance to scaled expression and full in vitro characterization. Hits failing one criterion enter optimization (engineering, [DMS-guided affinity maturation](/deep-mutational-scanning-a-high-throughput-approach-to-mapping-protein-fitness-landscapes), or framework re-grafting). Hits failing multiple criteria are deprioritized.
The conversion rate from "NGS-enriched clone" to "developable lead" in a well-run campaign is typically 30–60%. Below 20%, something is wrong with sorting stringency, library design, or target preparation.
## Two-platform handoff
Yeast display is a discovery platform, not a developability platform. The high-mannose glycans yeast installs are not the complex glycans mammalian production cells will install, and a clone that displays well on yeast can still fail on mammalian expression for reasons yeast cannot model.
For programs heading to therapeutic development, the right workflow runs [the two-platform sequence](/the-two-platform-approach-using-yeast-display-for-affinity-and-mammalian-display-for-developability): yeast display for affinity discovery, mammalian display for developability validation. The yeast arm delivers the hit list; the mammalian arm filters out clones that fail PTM compatibility before committing to expression scale-up.
## Decision summary
If your target is a soluble antigen, your library is smaller than 10^9, and your downstream format is scFv, Fab, or VHH: yeast display is the right discovery platform.
If your library must be larger than 10^9 (truly naïve diversity, no design control): start with phage display, then triage the top 10^4–10^5 hits on yeast for quantitative ranking.
If your target depends on PTMs yeast cannot install (complex glycans, sulfation, specific O-linked structures): skip yeast and run mammalian display from the start.
---
If you're scoping a yeast display campaign and want a second opinion on scaffold choice, library design, or sorting strategy, see our [yeast surface display services](/yeast-display) or [start a Binder Pilot](/binder-pilot). For multi-target programs and full IgG development, see the [AI Binder Sprint](/ai-binder-sprint).
{/* ranomics:related-services */}
## Related Ranomics services
- **[Yeast surface display services](/yeast-display):** Full-stack yeast display campaigns for antibody discovery and engineering.
- **[AI protein binder design](/ai-protein-binder-design):** De novo binder design feeding into yeast display validation.
- **[AI Binder Sprint](/ai-binder-sprint):** 6–8 week flagship program combining AI design with yeast/mammalian display.
---
## How Big a Yeast Display Library Do You Need for a 10 nM Binder?
> A practical walkthrough of library size math for a 10 nM affinity target on yeast display: starting material, sort gate stringency, Poisson coverage for NGS, and the KD ladder across multiple rounds.
Source: https://ranomics.com/yeast-display-library-size-for-10nm-target/
Published: 2026-04-21
The most common scoping question on a yeast display intake call is some version of "how big a library do we need?" The honest answer is that it depends on where you are starting and where you are going. This post walks through the math for a 10 nM KD target, which is a typical specification for a preclinical antibody lead or a usable tool binder. The same framework applies to tighter or looser specifications with adjusted numbers.
## Starting Material Sets the Floor
The library you screen is not the library you build. The useful diversity is the diversity of the starting population that already has some probability of binding the target. This varies by an order of magnitude or more depending on how the library was constructed.
**Naive libraries.** A fully synthetic library with random CDR diversification has no bias toward your target. Most hits that enrich in a naive library sit in the 100 nM to 1 µM range, and reaching 10 nM almost always requires a second round of affinity maturation built on a first-round lead. Plan for two campaigns in sequence: a discovery campaign at 10^8 to 10^9 diversity, followed by a maturation library around a validated hit.
**Immunized libraries.** A VHH library cloned from an immunized llama or alpaca has already been selected in vivo for binding to the target. The starting population is enriched for antigen-specific B cells, and germline lineages that bind the target are over-represented. A well-constructed immunized library of 10^7 to 10^8 transformants can reach single-digit nanomolar affinity directly without a separate maturation step. The practical constraint is library construction quality, not diversity ceiling.
**Computational pools.** A library of de novo designs from RFdiffusion, BindCraft, or a related pipeline is already pre-filtered for shape complementarity and predicted binding energy. Most of the designs that make it through self-consistency filtering have a measurable probability of binding. A well-curated pool of 10^5 to 10^6 computational designs, screened on yeast, can deliver sub-100 nM hits in a single campaign, and affinity maturation on the top hits reaches 10 nM without a second discovery library.
The KD goal and the starting material together define the minimum useful library size. For a 10 nM target, a naive library needs 10^8 or more to give affinity maturation a working lead. An immunized library needs 10^7 or more to capture the rare high-affinity germlines. A computational pool needs 10^5 or more to cover the design diversity that survived filtering.
## The Coverage Math: Poisson Sampling
Library construction is a sampling process. If your transformation efficiency is 10^7 transformants per microgram and your theoretical DNA diversity is 10^8, you have not sampled 10^8 sequences; you have sampled roughly 10^7 sequences with replacement from a 10^8 pool. The Poisson distribution tells you what fraction of theoretical variants appear at least once.
For the NGS readout, the same math applies in reverse. If you want to see 95% of your library members at least once in an NGS run, you need roughly three times the library size in reads. For a 10^8-member input library, that is ~3 * 10^8 reads per round, across three to four rounds, plus the zero-selection baseline. On a NextSeq 550 or similar Illumina platform this is feasible; on a single MiSeq run it is not. Scope NGS capacity before you scope library size.
Coverage drops as the library concentrates through selection. Post-sort populations are smaller and deeper sequencing gets you finer enrichment resolution on the surviving clones. Many groups budget higher read counts for the zero-selection baseline and the final sorted population, with lighter coverage on intermediate rounds.
## Sort Gate Stringency
A FACS sort on a yeast display library is a one-dimensional selection: gate on binding signal, normalized to surface expression. The stringency of the gate sets how much of the library passes each round.
At a 10 nM target concentration, the expected fraction of displayed variants that bind at or above the gate threshold is roughly the fraction of the library with KD at or below ~10 nM. For a naive library this is typically 0.01% to 0.1% of the library. For an immunized library it can be 0.1% to 1%. For a computational pool it can be 1% to 10%. Those rates compound across rounds.
A typical four-round FACS campaign against a 10 nM target starts from 10^8 cells sorted in round 1, sorts around 10^4 to 10^6 positive cells, grows them to 10^7 to 10^8, then sorts again. By round 4 the post-sort population is dominated by a few hundred to a few thousand enriched clones. NGS resolves the ranking.
If the input library is too small relative to the expected positive fraction, the round 1 post-sort population can drop below the 10^4 cells needed to recover a reliable population. If the input library is too large relative to the expected positive fraction, round 1 oversamples bystanders and FACS throughput becomes the bottleneck. Library size and expected hit rate need to be specified together.
## The KD Ladder Across Rounds
Multi-round yeast display selection against a specific KD target uses a decreasing target concentration ladder. A common specification for a 10 nM endpoint goal:
- Round 1: target concentration at 100 nM. Five to tenfold above the goal to capture anything binding at or tighter than the goal, with margin. Relatively permissive gate.
- Round 2: target concentration at 10 nM. At the goal. Tighter gate to enrich true 10 nM binders over bystanders.
- Round 3: target concentration at 1 nM. Tenfold below the goal. Selects for the tightest binders in the enriched population. This round is sometimes replaced or paired with an off-rate selection, where the library is pre-bound at saturating target and then washed with unlabeled competitor.
- Round 4 (optional): counter-selection, specificity panel, or 100 pM for further affinity resolution. Used when the downstream readout needs specificity information (for example, cross-reactivity with close homologs) rather than raw affinity.
The ladder converts a broad 10^8 library into a ranked list of tens to low thousands of clones at the endpoint. NGS on each round lets you track enrichment and pick the clones that climb consistently across rounds rather than the ones that win a single round by luck.
## Library Size Sanity Checks
Two failure modes are worth checking before committing to synthesis.
**Theoretical diversity exceeds the yeast ceiling.** If your diversification pattern encodes more than ~10^9 theoretical DNA variants, you cannot sample it in a single yeast library. Transformation efficiency caps at around 10^8 transformants per electroporation, and pooled campaigns can stretch this to 10^9 at significant expense. If the math says 10^10 or more, you are guaranteed to undersample. Either narrow the diversification pattern, use a trimer mix to reduce redundant codons, or split into focused sub-libraries that cover different regions of the sequence space.
**Library smaller than 10^6.** If your pool fits in fewer than a million variants (for example, 500 rationally designed point mutants, or 10,000 computationally filtered designs), yeast display is often not the right screening format. Individual-clone ELISA or low-throughput SPR against a handful of candidates can deliver affinity-ranked hits faster, without the overhead of library construction, FACS, and NGS. Yeast display earns its cost when the diversity is high enough that individual-clone screening is impractical.
The [Yeast Display Library Planner](/technology/library-planner) flags both of these cases automatically against the yeast transformation ceiling and the NGS read budget you specify.
## Computational Design Plus Display: The Best Case
The campaign architecture with the cleanest library size math is a pool of 10^3 to 10^5 computationally generated designs screened on yeast. The pool is small enough to transform fully, large enough to justify FACS over individual-clone screening, and pre-filtered enough that the per-clone hit rate is meaningful.
A representative workflow: RFdiffusion generates 10,000 backbone designs against a defined hotspot patch. ProteinMPNN redesigns sequences onto the backbones. Self-consistency filtering by ESMFold, ColabFold, or Boltz-2 removes designs that do not refold to the intended backbone (typically 50% to 80% of raw output). The filtered pool of 1,000 to 3,000 designs is synthesized as an oligo pool, cloned into the yeast display vector, transformed, and sorted against the target.
Because the starting pool is already enriched for predicted binders, hit rates in round 1 typically run at 1% to 10%, not 0.01%. The FACS throughput requirement drops by two orders of magnitude. Three rounds at a decreasing KD ladder often reach 10 nM directly, without a separate affinity maturation campaign. End-to-end timeline runs three to four weeks from a ready-to-synthesize design pool to a ranked hit list with NGS enrichment data.
This architecture is the backbone of the [AI Binder Sprint](/ai-binder-sprint) flagship program, which pairs RFdiffusion, BindCraft, and similar generative tools with in-house yeast display screening.
## The Short Answer
For a 10 nM endpoint KD target:
- Naive library: 10^8 or larger, plus a follow-on affinity maturation campaign.
- Immunized library: 10^7 to 10^8, single campaign feasible.
- Computational pool: 10^3 to 10^5, single campaign feasible with a cleaner hit rate distribution than a naive library.
Library size is a downstream variable. Start with the target KD, the starting material, and the sort budget; library size falls out of that specification, not the other way around.
## Scoping a Campaign
To run these numbers against your specific target and starting material, the [Yeast Display Library Planner](/technology/library-planner) takes library inputs and a sort plan and returns theoretical diversity, achievable diversity, stop codon load, and NGS read depth per round. It flags cases where the plan is mathematically infeasible before any synthesis budget is committed.
For a scoped yeast display campaign end to end, the [Yeast Display service](/yeast-display) page covers platform capabilities. The [AI Binder Sprint](/ai-binder-sprint) pairs computational design with display validation for campaigns that benefit from a pre-enriched starting pool. To scope a project, use the [contact form](/ranomics-contact?service=yeast-display) and include the target, the starting library architecture (naive, immunized, or computational), and the KD specification.
{/* ranomics:related-services */}
## Related Ranomics services
- **[Yeast Display Library Planner](/technology/library-planner):** Free planning tool for library size, sort strategy, and NGS depth.
- **[Yeast Display](/yeast-display):** End-to-end yeast surface display CRO service for scFv, VHH, and scaffold libraries.
- **[AI Binder Sprint](/ai-binder-sprint):** Computational binder design paired with yeast display validation.
---