J. Chem. Inf. Model.
2025
Prospective Validation of Alchemical Free Energy Perturbation Binding Affinity Predictions Across 14 Clinical Targets
Patel M, Krishnamurthy R, Chen S, Andersen MR.
We report a prospective benchmark of relative binding free energy calculations using the DrugSynq FEP pipeline across 312 congeneric pairs from 14 clinical targets. The pipeline achieved r²=0.82 (RMSE=1.02 kcal/mol) on the held-out test set. Outlier analysis identifies scaffold-hop perturbations as the primary failure mode. Evaluation code and test set SMILES are published in the SI.
JACS Au
2024
Ensemble ADMET Prediction Models with Prospective Performance in Lead Optimization Campaigns
Chen S, Patel M, Krishnamurthy R.
We introduce an ensemble gradient-boosted model architecture for ADMET prediction trained on 247,000 curated in vitro data points across 12 endpoints. Prospective AUROC values range from 0.85–0.91 across endpoints. We compare performance against publicly available ADMET tools on the same held-out sets and discuss training data curation methodology.
J. Comput. Chem.
2023
Machine-Learned Correction Terms for OPLS4 Free Energy Perturbation Calculations on Drug-Like Heterocycles
Krishnamurthy R, Patel M.
We present a hybrid approach combining OPLS4 molecular mechanics with ML correction terms trained on high-level QM reference data for 18,000 drug-like conformers. The ΔML correction reduces RMSE by 0.3 kcal/mol on kinase target test sets relative to OPLS4 alone. Implementation details and training data are open-sourced.