Can a reproducible, uncertainty-aware workflow rank scaffold-held-out HIV-1 protease compounds for follow-up while making its validation limits visible?
- Curate activity records — retain standardized ChEMBL IC50 measurements reported as exact nanomolar values and aggregate repeated measurements by median pActivity.
- Standardize structures — parse SMILES with RDKit, select the largest fragment, remove explicit hydrogens and canonicalize the result.
- Create a scaffold split — assign Bemis–Murcko scaffolds wholly to train or test data to reduce analogue leakage.
- Build baselines — compare an Extra Trees regressor using Morgan fingerprints with a mean-prediction baseline.
- Quantify uncertainty — use variation across individual trees as a pragmatic model-uncertainty indicator.
- Check applicability — report each candidate's maximum Tanimoto similarity to the training set and penalize low-similarity predictions.
- Prioritize transparently — combine potency, uncertainty, domain similarity and a broad molecular-weight screen in a documented score.
- Simulate learning — reveal labels for selected held-out compounds over repeated retrospective rounds. This is not experimental validation.
The holdout is scaffold-based, never a random row split. Model and baseline are evaluated on the same unseen scaffold set using MAE, RMSE and R². The small, bounded sample makes these diagnostic estimates, not claims of clinical or discovery performance.
- Fixed random seed:
42 - Versioned dependency ranges in
pyproject.toml - One-command demo:
hiv-protease-discovery demo - Executed notebook with saved outputs
- Unit tests and automated lint/test workflow
No wet-lab assay, docking result, binding claim, de novo molecule claim or clinical conclusion is presented. The optional Transformer module defines an inspectable architecture only; it does not ship a pretrained checkpoint.