Liver toxicity (DILI) β€” GAT deployment model

Binary classifier for drug-induced liver injury (DILI) concern: the probability that a compound belongs to the toxic class. This is the deployment refit of the liver programme's Arch2 GAT (graph attention network), one of the two winners selected on the locked regulatory benchmark. It is deployed as a pair with the ExtraTrees deployment model (Delower/liver-extratrees-deployment); a compound flagged by both models is the highest-concern readout.

Recipe

  • Architecture: GAT over molecular graphs (atom features, in_dim 35), hidden 128, dropout 0.2, lr 3e-3 β€” the locked run's winning configuration (validation MCC 0.2131, best epoch 37), reused unchanged with no re-tuning.
  • Final refit: fixed 37 epochs on the full deployment set (the arm's final-refit convention), positive-class weight 0.8853.
  • Flag threshold: probability β‰₯ 0.5 (the arm's own convention).

Deployment training set

All labeled liver data the programme holds, identity-deduped by InChIKey connectivity of the standardized parent: 1,891 unique compounds (1,003 positive / 888 negative), composed of train_locked.csv (1,769 rows) + train_enriched.csv additions + internal_test.csv + every regulatory benchmark compound (withdrawn, clinical failures, PASS), with benchmark labels governing conflicts. 1,769 duplicate rows dropped; 42 label conflicts (all inside train_locked, the groups the earlier merge had excluded) kept at the first train_locked row's label and logged.

Hard gate (passed before training): every deduped regulatory benchmark compound was verified compound-by-compound (InChIKey connectivity) inside the actual training matrix, post-graph-building: withdrawn 71/71, clinical failures 10/10, PASS 304/304. See gate_verification.json.

Honest benchmark scores β€” benchmark version only

The deployment refit absorbed the regulatory benchmark and is never scored on it. The frozen claim for this recipe comes from the locked-wave benchmark model (benchmark locked out of training, overlap verified 0; 385 scored compounds):

Metric Value
Pooled regulatory AUROC 0.788
Pooled balanced accuracy 0.680
Withdrawn drugs caught 66/71
Clinical failures caught 10/10
PASS compounds correct 128/304 (0.421)

This was the best pooled regulatory AUROC of any liver arm, ahead of the leakage-cleaned external Claude ensemble (0.670).

Deployment screen results

Applied to the DPP-4 programme's potent leads (flag = prob β‰₯ 0.5): 1,110/1,113 workbook leads scored, 1,027 flagged (92.5%); 4,891/4,898 of the full potent union scored, 4,496 flagged (91.9%). These flag rates are high β€” every model in this programme is a high-sensitivity, low-specificity screen, and the pair's discriminative readout is the dual flag (1,003/1,110 leads flagged by both models; only 6 pass both). Treat single-model flags as low-information on their own.

Files

  • model.pt β€” deployment weights: the epoch-37 training checkpoint (keys: state_dict β€” the exact final weights β€” plus optimizer and RNG state). The arm-format re-save of the deployment dict could not be written in the build environment, so the checkpoint written during training is the shipped artifact; config.json carries the architecture config and in_dim.
  • config.json β€” architecture, training composition, gate counts, benchmark scores.
  • deployment_common.py β€” the programme's shared featurization / graph module (graph construction + GATNet definition).
  • infer.py β€” prediction entry point (predict(smiles)).
  • validate.py + examples.json β€” runnable validation: reproduces this model's probabilities on 8 screened DPP-4 leads (tolerance 2e-3). Verified locally before upload: max deviation 4.8e-07.
  • gate_verification.json, training_manifest.json β€” hard-gate result and full training manifest.
  • absorbed_compounds.csv β€” the 1,891-compound deployment training set (the absorbed data, for audit).

Limitations

CPU/GPU-agnostic PyTorch model; requires torch, torch_geometric and rdkit. Probabilities are screening scores, not calibrated clinical risks.

Downloads last month
35
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support