DPP-4 pIC50 — LTN (tuned)

A Logic Tensor Network (LTN) regression model predicting DPP-4 pIC50 from molecular structure. The deployed weights are the tuning winner b256 of a 7-configuration search — the evaluated artifact, trained on the 80% training split only (8,068 rows), best-validation snapshot, FDA-12 gliptins excluded from training.

Performance (this model)

Split	RMSE	MAE	R²
Validation (n=1,009)	0.3907	0.1767	0.9053
Test (n=1,009)	0.4457	0.1870	0.8739
FDA-12 external vs wet-lab (n=11)	0.8348	0.6498	—

Derived classification (paper scheme, from predicted pIC50), test: accuracy 0.9007 · F1 macro 0.8923 · MCC 0.7866 · ROC-AUC macro 0.9948. ExtraTrees comparison: test RMSE 0.4449 · FDA wet-lab RMSE 0.9637 — the LTN ties it on the internal test and is clearly better on the FDA drugs.

Why not a full-data retrain? Retraining this configuration on all 10,097 rows (curated set + FDA-12) was attempted three times with stability fixes (gradient clipping 1.0; learning rate halved to 5e-5); all three runs diverged mid-training (epochs 49 / 68 / 25) and produced no usable weights. The evaluated tuned model is therefore the deployed one — its scores above are measured, not fit checks.

Model

Architecture: Keras MLP 768-384-192 (ELU) → 1, trained inside an LTN (equality predicate 1/(1+0.5‖a−b‖), Forall pMeanError p=2, loss 1−sat), Adam lr 1e-4, batch 256, seed 42, 85-epoch budget with best-val snapshot.
Features: 4,608 binary fingerprints — Morgan/ECFP4 (r=2) 1024 + RDKit path fingerprint 1024 + Morgan 512 + Morgan 2048. No standardisation.
Weights are raw Keras arrays (w0..w7) in .npz; infer.py rebuilds the network with TensorFlow, and a NumPy-only forward pass reproduces the predictions to <1e-5 (that is how the web app serves this model).

Data

10,085 compounds stacked from ChEMBL, BindingDB, PubChem, the ChEMBL-2024 set and DTC; censored IC50 ≥ 100,000 nM dropped; duplicate SMILES resolved by median (conflicts > 1 log dropped). Split 80/10/10, seed 42, stratified by pIC50 band, FDA-12 excluded from all splits.

Caveat: curation deduplicated by SMILES string; the 10,085 strings are 7,327 unique canonical compounds, so internal scores may be mildly optimistic. The tuning was run on this set deliberately, for comparability with the project's other models.

Usage

pip install -r requirements.txt
python infer.py "Cn1c(=O)cc(N2CCC[C@@H](N)C2)n(Cc2ccccc2C#N)c1=O"
# Alogliptin → predicted pIC50 7.74

Files

ltn_model.npz + ltn_config.json — tuned weights + config
metrics_ltn.csv — validation/test metrics · ltn_fda12.csv — FDA-12 predictions vs wet-lab values
infer.py — standalone inference · train_ltn_dpp4.py — the tuning training script (provenance)

Downloads last month: -; Downloads are not tracked for this model. How to track