Tool   Benchmark   Validation

Retrospective validation

Before you trust a prediction, see whether loak recovers allosteric sites we already know the answer for. Five known allosteric drug–target complexes. For each one, the drug's crystal-structure binding site defines ground truth. We run the public /predict endpoint on the apo-ish PDB, score each of the top 5 predicted pockets by residue-level Jaccard (IoU) overlap with the ground-truth site, and count it a hit only if any of the top 5 clears IoU ≥ 0.20.

Cases

Methodology

Ground truth — For each PDB, we parse the bound drug's HETATM coordinates and define the ground-truth pocket as every protein residue with any atom within 5 Å of any ligand heavy atom.

Prediction — We call the live /predict endpoint on the same PDB (same code-path a user would hit). We take the top 5 predicted pockets, extract each pocket's residue list, and compute residue-level Jaccard (IoU) against the ground-truth pocket. The best score across the top 5 is the one we report.

Hit criterion — IoU ≥ 0.20 (the commonly cited 20% Jaccard threshold for pocket identification). Weaker thresholds inflate hit rates; stricter ones are for sub-pocket geometry.

Model versionallonet_gbt_v1. When LoakNet v4+ passes internal review it replaces the GBT here; this page will update automatically.

Caveats — A single drug-target complex is one structural snapshot. The ligand's pocket in a holo structure may differ from cryptic conformations in the apo state. For cryptic pockets (e.g. LoakGen's target regime), lower IoUs are still meaningful. We report raw numbers, not tuned thresholds.