AffiniBind scans your protein's surface, binder design engines work in parallel, and quality filters screen every candidate before it reaches the lab. The goal is predicted binding proteins that actually work.
Containerized · Cloud GPU · Auto-scaling
WeaveSeq runs a vertically integrated pipeline from target preparation to the post-ranking developability gate. Each stage produces a schema-validated Parquet table consumed by the next: no silent fallbacks, no heuristic conversion, and complete traceability from raw structure to ranked candidate shortlist.
The target structure is cleaned and standardized: non-standard residues removed, missing atoms rebuilt, and surface properties measured. The residue table and surface metrics built here are the inputs every later stage reads from.
Four models scan the surface independently. ScanNet and MPBind use deep learning to spot binding-prone areas, MaSIF reads 3D surface shape, and APBS maps electrostatics. An XGBoost model combines their predictions into one confidence score per residue.
High-confidence residues are grouped into contiguous patches by spatial clustering; duplicates are merged and the survivors ranked by predicted quality. An interactive NGL viewer shows where each patch sits on the target surface.
Three protein design engines run simultaneously on RunPod cloud GPUs, each with a different strategy. FreeBindCraft refines designs through AF2-guided hallucination. PXDesign generates backbones shaped to fit the target interface. BoltzGen builds them by discrete diffusion. Jobs are distributed across GPU pods and managed automatically.
Each generator's own confidence metrics gate its designs first, before any expensive scoring runs. Survivors are normalized into a single candidate table, the contract the rest of the pipeline reads from.
Every surviving candidate is rescored on cloud GPUs: ESMFold2 predicts the binder–target complex and also folds the binder alone. The output is a per-design set of ten structure-derived channels: two monomer-confidence features plus eight co-fold interface-confidence features (iPTM, PAE, pLDDT and ipSAE) that the ranking stage scores.
Every run is packaged into an interactive HTML report, spreadsheets, and Parquet tables, with a JSON bundle for downstream automation. Intermediate results are retained, so any candidate can be traced back to the structure it came from.
Rescored candidates rank first: WeaveRank scores the IVTT route, pool confidence the yeast-display route. WeaveRank is a five-channel z-ensemble validated on 1,880 held-out literature designs: a median four of its top ten are confirmed binders, 1.6× the target's own base rate. Candidates without rescore data follow, ordered by generator confidence.
The developability gate runs last, on the accepted shortlist only: thirteen checks covering PRODIGY affinity, solubility, SAP3D aggregation, PROPKA charge, protease and patch risk, packing, and self-association. Every check attaches metrics and a risk flag; exclusion is per-tool opt-in, and flagged candidates are partitioned rather than dropped.
Every candidate is scored on 10 structure-derived channels: 2 ESMFold2 monomer-confidence and 8 ESMFold2 co-fold interface-confidence features. The WeaveRank five-channel z-ensemble orders the IVTT shortlist, validated on 1,880 held-out literature designs, and a post-ranking developability gate of thirteen checks attaches risk flags to the accepted set. Representative figures from a recent campaign are shown below. See the glossary for metric definitions.
Integrate AffiniBind into your discovery pipeline and work directly with our engineering team, from target analysis through assay-ready reagents.
WeaveRank Validation Across 1,880 Literature Designs