Live demo · 907k-parameter model · runs entirely in your browser

GNN-PCNA: cryptic-pocket residue prioritization

A dual-branch GATv2Conv graph neural network that scores every residue of a protein structure for how likely it belongs to a cryptic (transient) pocket. The example here is PCNA, the homotrimeric DNA-replication sliding clamp targeted by the anticancer molecule AOH1996. Each score shown below is the real model's output, precomputed and rendered live in the page.

A word on what the score means. It is a prioritization signal meant to generate hypotheses for follow-up study — not a claim that a druggable pocket exists, and not MD-validated. This page runs the lightweight 907k-parameter, geometry-only model. The benchmark figures further down come from the larger XL + ESM2 model. Full caveats in LIMITATIONS.md.

3D structure · colored by score

drag to rotate · scroll to zoom
low priorityhigh priority

Per-residue score profile

score ∈ [0,1] · dashed = 0.5
x = residue index (array order), one line per chain.

Top-ranked residues

highest score first
#residueaasssasascore

Predicted residue clusters

DBSCAN · eps 6Å · min 3
0.50
#sizemeanmaxresidues

Export

The painted PDB stores each residue's score (×100) in the B-factor column. Open it in PyMOL and run spectrum b, blue_white_red to reproduce the coloring shown above.

Benchmark context (XL + ESM2 model, not this demo)

The headline metrics below come from the larger XL + ESM2 model, evaluated on an MMseqs2 30%-homology-clean split. We report these instead of the earlier random-split numbers, which were superseded once we found homology leakage in that split.

ConditionTest AUPRC95% CITest AUROC
xl_esm_full (best)0.25130.127–0.3820.8649
xl_geometry0.19230.116–0.2680.8325
xl_esm_zero0.10710.047–0.1770.6815
small_geometry (this demo)0.15510.055–0.2430.7626

These are for context only and are not a like-for-like comparison, since the datasets, splits, and label definitions all differ: PocketMiner (GVP-GNN) reports AUROC 0.87 / AUPRC 0.44, and CryptoSite reports AUROC 0.83. Sequence context from ESM2 is a large contributor to our numbers, and we state that openly rather than hide it. Research by Reshwant Borra and Advay; public demo built by Advay; manuscript in preparation.