YangyangMiao1998/tractability-weights

Binder-modality tractability — model weights

Weights for the surface-patch tractability pipeline: given a region on a protein surface, predict which binder class could target it — small molecule, antibody / nanobody, peptide, or de-novo mini-protein.

The code that consumes these weights lives at https://github.com//tractability-pipeline; this repo holds only the checkpoints, because they are past GitHub's per-file limit.

Files

file	size	what it is
atomsurf_site_version69.ckpt	63 MB	AtomSurf surface encoder (version_69, NOESM). Produces the per-vertex embedding every downstream stage reads, and a protein-interaction site score
pli_site_ligandblind.ckpt	310 MB	Protein–ligand interaction site model, ligand-blind. Proposes small-molecule pockets from the apo structure
cls_4cat_ligchain_geom.pth	2.5 MB	4-category GVP-Transformer patch classifier (261-dim scalar input, geometry features included)
SHA256SUMS	—	manifest; the build verifies against it before using any weight

Usage

git clone https://github.com/<your-username>/tractability-pipeline
cd tractability-pipeline
HF_REPO=<your-username>/tractability-weights ./prepare_context.sh
apptainer build --fakeroot --ignore-fakeroot-command tractability.sif tractability.def

prepare_context.sh downloads the weights into checkpoints/ and verifies them against SHA256SUMS before the image is built. See USAGE.md in the code repo for how to run the pipeline.

What these were trained on

• Surface encoder — PINDER protein–protein complexes, site-prediction objective, no ESM features.
• PLI site model — PLINDER protein–ligand complexes, ligand-blind: the model never sees a ligand, so it applies to apo structures.
• 4-category classifier — four interface sets: SAbDab antibody/nanobody epitopes, peptide-binding sites, de-novo minibinder interfaces, and PDBbind ligand pockets, as 12 Å geodesic surface patches. Splits are at 40 % sequence identity, cluster-level.

Reliability, measured

Against an external benchmark, per class AUROC: ab+nb 0.81, small_mol 0.75, pep 0.71, minibinder ≈ 0.5 — it does not generalise, so treat a minibinder call as uninformative.

On a retrospective comparison against 4,568 human proteins with BioLiP annotations:

• the PLI site model places a predicted pocket on the real ligand site for 72.9 % of experimentally confirmed sites (random patches on the same surface: 7.3 %);
• ranking every surface point by predicted site probability gives a median AUROC of 0.928 against the experimental site;
• of the pockets that are correctly located, only about half are also called small_mol by the classifier.

The useful summary: where these models point is reliable; what the classifier calls the region is the weaker step. Performance holds on proteins with no close training homolog — site recall drops only from 81 % to 69 % between ≥95 % and <30 % sequence identity to the training set.

Caveats

• The classifier's shipped decision thresholds maximise a class-balanced F1, which ignores true negatives and is therefore permissive (small_mol at 0.10 flags 82 % of the human proteome). Recalibrating on MCC gives 0.59 / 0.26 / 0.91 / 0.24 for ab+nb / pep / minibinder / small_mol.
• The four probabilities come from a sigmoid multi-label head and do not sum to 1.
• Patch geometry (12 Å geodesic, k=20) is part of the model contract. Changing it without retraining invalidates the weights.