Ranomics
De novo binder scaffolds bound to a target protein, generated by PXDesign's Protenix diffusion and ranked with AlphaFold2 Initial Guess
De novo binder design

PXDesign de novo binder design

PXDesign is ByteDance's open-source de novo binder design tool, built on the AlphaFold3-class Protenix model. Ranomics runs it self-serve and on our own paid wet-lab campaigns, so the candidates you generate come out in the same format we hand to the bench.

Run it self-serve, then hand the candidates to our team for wet-lab validation when you find designs worth testing.

PXDesign (ByteDance, bioRxiv 2025), built on the Protenix AlphaFold3 reproduction. Designs are ranked with AlphaFold2 Initial Guess (Bennett et al., 2023).

How it works

Target structure to ranked binder candidates

01

Define hotspots

Upload a target PDB or AlphaFold model. Specify hotspot residues that the designed binder must contact, or pull them from Epitope Scout output.

02

Generate backbones

PXDesign's Protenix-based diffusion generator produces hundreds to thousands of novel scaffolds geometrically complementary to the hotspot patch.

03

Design sequences

ProteinMPNN assigns sequences to each backbone, with fixed positions at predicted contact residues to preserve binding geometry.

04

Score and rank

AlphaFold2 Initial Guess and Protenix confidence refold each binder-target complex and rank on interface pLDDT, PAE, and ipTM. Ranked CSV plus PDBs delivered.

Methodology

Four stages, one integrated run

PXDesign is ByteDance's open-source pipeline, built on their AlphaFold3-class Protenix model. It chains the canonical stages of modern de novo binder design into one integrated run, with each stage passing its outputs directly to the next and no manual handoff between tools.

Stage 1

Backbone generation

Protenix-based diffusion. PXDesign's generator, built on ByteDance's AlphaFold3-class Protenix model, reverses a noise process to produce backbone coordinates geometrically complementary to the specified epitope. Scaffold length range and topology constraints are tunable per run.

Stage 2

ProteinMPNN sequences

Inverse-folding sequence design. Multiple sequences sampled per backbone at controllable temperature, with hotspot contact positions fixed to preserve the diffusion-defined binding geometry.

Stage 3

AlphaFold2 Initial Guess

AlphaFold2 multimer refolds each binder-target complex using the designed structure as the initial coordinate guess. This is faster and more discriminative than de novo AF2 multimer prediction, and PXDesign combines it with Protenix confidence to rank designs.

Stage 4

Interface filtering

Candidates are ranked on interface pLDDT, PAE at the predicted interface, and ipTM. Designs whose refold matches the original geometry rise to the top; designs that drift are discarded.

Output

Ranked candidate set

CSV of scored designs plus per-candidate PDB structures. Top candidates are pre-formatted for direct handoff into a Ranomics wet-lab Binder Pilot or AI Binder Sprint.

Beyond the algorithm

Why running PXDesign is not the same as downloading it

Current-generation scoring

PXDesign scores with AlphaFold2 Initial Guess and Protenix confidence rather than default AF2 multimer, which is slow and noisy when the binder is a never-seen-before designed sequence. Initial Guess seeds the model with the designed structure and asks whether it agrees the complex is real. The result is faster scoring and a success metric that tracks wet-lab binding more tightly than default multimer scores.

Wet-lab feedback loop

We run PXDesign on real campaigns and see which designs actually bind, so we know its true hit rate and where it fails. That feedback tunes how we use it: filter thresholds, scaffold length ranges, hotspot weighting, and which tool fits which target. You launch it with the settings we learned from real yeast display screens, not defaults out of the box.

Scales to full campaigns

Self-serve PXDesign runs on the same infrastructure we use internally. Hundreds of candidates per run. Need 10,000 to 50,000 designs across a full campaign? The same pipeline scales up inside an AI Binder Sprint with no rewrite and no re-validation.

When to use PXDesign

The tool we reach for first on paid campaigns

PXDesign is current-generation. Its generator is built on Protenix, an AlphaFold3-class model, and the PXDesign benchmark reports experimental hit rates of 17 to 82 percent nanomolar binders across most targets tested. We run it, we know its real hit rate, and the Initial Guess scoring step is what makes us willing to put a 100% binder guarantee on the AI Binder Sprint.

Run it self-serve when you want the same tool Ranomics uses internally, with results in a format that drops straight into a wet-lab campaign if the candidates are worth testing.

You want the same de novo binder tool Ranomics runs on its own paid wet-lab campaigns

Comparing AlphaFold2 Initial Guess scoring against your existing RFdiffusion plus default AF2 multimer pipeline

Need ranked binder candidates against a target you already have a PDB or AlphaFold model for

Planning a Binder Pilot or AI Binder Sprint and want to scope the design space before paying for wet-lab validation

Running an internal feasibility study before committing budget to a multi-week campaign

Tuning hotspot residue selection from Epitope Scout output and want to see which patches generate scoreable designs

Run the tool we run on our own campaigns

Sign in, upload a target structure, define hotspots. Get ranked binder candidates scored with AlphaFold2 Initial Guess.

PXDesign (ByteDance, bioRxiv 2025), built on the Protenix AlphaFold3 reproduction. Designs ranked with AlphaFold2 Initial Guess: Bennett, J. et al. Improving de novo protein binder design with deep learning. Nature Communications 14, 2625 (2023). See all Ranomics technology.