Delower/dpp4-pic50-dnn_mlp

DPP-4 pIC50 Regression — dnn_mlp (best model, retrained on 100% of cleaned data)

Summary

Predicts pIC50 (DPP-4 inhibition potency, log scale) from a SMILES using a 4,608-dim fingerprint: Morgan/ECFP4-1024 + CDK-Extended alternative (RDKit path fingerprint)-1024 + ECFP4-512 + ECFP4-2048. Trained on 10,085 curated compounds (ChEMBL-derived), 80/10/10 split grouped by canonical SMILES, stratified by pIC50 band, seed 42. The FDA-12 set was excluded from training.

Uploaded weights were retrained on 100% of the cleaned data (FDA-12 still excluded). The performance quoted below is from the split run (held-out test set, n = 1,014) — it is the honest estimate, not the performance of the retrained weights themselves.

Split-run results (held-out test set, n = 1,014)

Model	Notebook	Test RMSE	Test MAE	Test R2	FDA RMSE*
dnn_mlp	A	0.5712	0.3105	0.8019	0.7269
lgbm	A	0.5750	0.3578	0.7993	0.7600
xgb	A	0.5822	0.3699	0.7943	0.8369
rf	A	0.6129	0.4097	0.7720	0.8091
ltn	C	0.7476	0.5675	0.6607	n/a
gat	C	0.8816	0.6847	0.5282	n/a
chemberta_frozen_ridge	B	0.9364	0.7327	0.4677	n/a
transformer3	B	0.9535	0.7313	0.4481	n/a
transformer4	B	0.9895	0.7712	0.4056	n/a
chemberta_ft	B	0.9928	0.7799	0.4016	n/a

Best model: dnn_mlp — test RMSE 0.5712, MAE 0.3105, R2 0.8019. See metrics.csv for full metrics (train/val/test, per-band, Pearson, Spearman, mean bias) and classification.csv for the post-hoc bioactivity classification (active >= 7.5 vs inactive <= 6.0, grey zone excluded; also >= 7.0 potent).

*FDA RMSE compares predictions against the pIC50 of the same canonical structure in the curated dataset (the FDA file has no pIC50 column), so it is NOT a fully independent holdout; it was used only as a tie-break. See fda.csv for the 12 FDA-approved drugs' predictions.

Caveat: Adjusted R2 is not interpretable on the test/validation sets here (p = 4,608 features > n), so the quoted headline metrics are RMSE / MAE / R2 only.

Files

• model.pt (torch state_dict; load into MLP(4608-768-384-192-1)) — retrained best-model weights
• config.json — architecture, feature recipe, normalisation (y_mean/y_std)
• metrics.csv — full split-run metrics for the best model
• fda.csv — FDA-12 predictions (12 FDA-approved DPP-4 inhibitors)
• model_selection.csv — all 10 models ranked by held-out test RMSE
• classification.csv — post-hoc bioactivity classification results
• infer.py — one-SMILES inference (recomputes the 4,608-dim fingerprint and prints predicted pIC50)
• requirements.txt

Usage

pip install -r requirements.txt
python infer.py "CC(C)Cc1ccccc1"   # any SMILES; prints the predicted pIC50

Limitations

• Fingerprint-only model: no 3D structure, no assay metadata, no pIC50 beyond the training range.
• Errors are largest for weakly active compounds (test per-band RMSE is highest for pIC50 < 5.5).
• Not a clinical or dosing tool — research use only.