The AI design engine behind our binder programs
Ranomics operates an in-house GPU compute pipeline for de novo protein design. We run RFdiffusion, BindCraft, and Boltzgen as an integrated workflow — not as a black-box API wrapper.
GPU-accelerated protein design at campaign scale
In-house GPU compute, no API quotas
We run design campaigns on cloud GPU instances (NVIDIA H100/A100) provisioned on-demand. This matters because diffusion-based design requires substantial compute at scale. Wrapping an external API with a rate limit cannot run a 50,000-design campaign in a useful timeframe. We provision the infrastructure to match the campaign.
The same team that designs your binder validates it
The computational pipeline feeds directly into Ranomics' display screening platforms. Designed sequences are synthesized as pooled oligo libraries, cloned into display constructs, and screened using the same workflows we use for traditional library campaigns. No handoff between a computation vendor and a screening CRO.
Design pipeline architecture
The models we run
RFdiffusion
Backbone generation +
RFdiffusion
Backbone generationA diffusion model trained to generate protein backbone structures conditioned on a binding target. Developed by Baker Lab (University of Washington). RFdiffusion reverses a noise process to produce backbone coordinates that are geometrically complementary to a specified epitope.
We run hotspot-conditioned diffusion, specifying epitope residues that constrain where the generated scaffold must contact the target. Campaign size: typically 10,000-50,000 backbone samples per run on cloud GPU (H100/A100).
Hotspot residue specification, scaffold length range, diffusion timesteps, symmetry constraints for homodimer targets
10,000-50,000 backbone samples per run, scaled to match campaign requirements.
BindCraft
Iterative design +
BindCraft
Iterative designAn integrated binder design pipeline that combines backbone generation with iterative sequence optimization, directly optimizing for target engagement geometry and predicted affinity metrics.
BindCraft is run in parallel with RFdiffusion to diversify the design space. It is particularly useful for targets where a specific binding mode (e.g., blocking a receptor interface) is required, as it allows more direct geometric constraints on the output scaffold.
Tighter integration between scaffold generation and sequence scoring. Produces candidates with stronger initial interface metrics.
Hundreds to low thousands of candidates per run, at higher average predicted quality than RFdiffusion at the same compute cost.
Boltzgen
Generative flow +
Boltzgen
Generative flowA generative flow model for protein design that samples from a learned distribution over protein sequences and structures, enabling exploration of sequence space beyond the training distribution of diffusion models.
Boltzgen diversifies the candidate pool by generating sequences with different backbone topologies from RFdiffusion outputs. Particularly useful for targets where standard diffusion-generated scaffolds underperform or for exploring alternative binding geometries.
Diversity generation. Boltzgen candidates are filtered by the same validation pipeline as RFdiffusion outputs before advancing to synthesis.
Cloud GPU (H100/A100), provisioned on-demand. Integrated into the same filtering pipeline as all design outputs.
ProteinMPNN
Sequence design +
ProteinMPNN
Sequence designA message-passing neural network that solves the inverse folding problem: given a backbone, what amino acid sequences will fold into it? Developed at University of Washington. ProteinMPNN generates sequences that are geometrically compatible with a provided backbone structure.
We generate 8-16 sequences per RFdiffusion backbone, using temperature sampling to explore sequence diversity. Fixed-position constraints are applied at predicted hotspot contact residues to preserve binding geometry while allowing variation elsewhere.
Sampling temperature (diversity vs. stability trade-off), fixed-position residues at interface, number of sequences per backbone
8-16 sequences per backbone. Lower temperature = more conservative sequences. Higher temperature = more diverse exploration.
Boltz-2, ESMFold & ColabFold validation
Structural validation +
Boltz-2, ESMFold & ColabFold validation
Structural validationStructure prediction models used as an independent validation layer. Boltz-2, ESMFold, and ColabFold (accelerated AlphaFold2) predict the structure of each designed sequence in complex with the target. ESMFold provides rapid single-sequence folding checks, while Boltz-2 and ColabFold validate the full binder-target complex. This is a critical quality gate before any sequence advances to synthesis.
Every candidate passing initial ProteinMPNN scoring is run through complex structure prediction. Candidates are scored by interface pLDDT, predicted aligned error (PAE) at the interface, and ipTM. Only candidates passing all three thresholds advance.
Interface pLDDT, PAE at predicted binding interface, ipTM (interface predicted TM-score), per-residue confidence at hotspot contacts
Structural validation typically removes 70-90% of candidates, converting a 10,000-design pool to 500-2,000 for synthesis.
Interested in the computational pipeline?
Tell us about your target. We will assess whether the design pipeline is the right approach and propose a campaign structure within 5 business days.