YAML Metadata Warning:empty or missing yaml metadata in repo card

Check out the documentation for more information.

hERG CardioGenAI β€” DEPLOYMENT REFIT (retrained from scratch)

DEPLOYMENT REFIT β€” trained on ALL labeled data, including the regulatory/external sets (hERG-70, new-compounds 967, FDA panel 36, REGFAIL). The CardioGenAI architecture (authors' Discriminator_Model: GAT + fingerprint + frozen pretrained transformer vector) retrained from scratch β€” no published weights β€” on the complete labeled corpus:

  • classification: 14,441 compounds Β· regression: 10,370 compounds

Hard gate (corrected): PASS. Every external/regulatory compound is present in the actual training caches by InChIKey connectivity (hERG-70 462/462, new-compounds 967/967, FDA 36/36, REGFAIL 7/7 β€” see HARD_GATE_CORRECTED.md). An earlier run trained on an incomplete split-derived cache was caught by this gate, invalidated, and retrained from the full all-data files; the full story is in FIX_REPORT_EIC53.md. The 57 (cls) / 35 (reg) compounds the frozen 57-token transformer cannot encode are included with a zero transformer vector and flagged β€” never silently dropped.

Do not validate these models on the absorbed sets. Checkpoint-selection scores in validation_results.json are in-sample selection signals only. Honest locked-pool benchmarks of this architecture: classification hERG-70 ACC 0.752 / MCC 0.525, new-compounds MCC 0.270 (best of all models there), FDA panel 3 errors; regression hERG-70 RMSE 0.856.

Contents

  • weights/best_cls_checkpoint.pt, weights/best_reg_checkpoint.pt
  • src/Discriminator.py β€” authors' model/dataset module
  • vocab_recovered.npy β€” recovered 57-token transformer vocabulary
  • predict_herg_cardio.py β€” inference from a featurized h5 (authors' format)
  • config.json, validation_results.json, HARD_GATE_CORRECTED.md, FIX_REPORT_EIC53.md

Environment pin that matters: torch_geometric 2.4.0 (the checkpoints do not load correctly under pyg 2.5.x β€” project issue EIC-50). Classification cutoff 10 uM (blocker = prob >= 0.5); regression outputs predicted pIC50.

Downloads last month
41
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support