First-Principles Prediction of CYP3A4 Metabolite Sites Reveals Chemical Cascade Pathways Missed by Conventional Benchmarking

We present a first-principles physics engine for predicting sites of metabolism (SOM) on CYP3A4 substrates that requires no training data, force fields, or machine learning. The engine operates on deterministic 3D molecular geometry and employs three independent scoring mechanisms based on proximity to the catalytic species, standing wave antinodes, and standing wave nodes with convergent thermal energy. On 46 blind CYP3A4 substrates, the engine achieves 91.3% top-5 accuracy, 84.8% top-3, and 99% multi-site recall at top-10. Accuracy degrades with molecular weight: top-1 accuracy is 77% for drugs below 300 Da but falls to 33% above 500 Da, reflecting 3D geometry builder limitations for large molecules. Cascade reanalysis of the 21 top-1 misses reveals that 16 predictions identify chemically unstable intermediates that cascade into the experimentally reported metabolites, alternative documented metabolites, or adjacent-site matches within two bond distances. Only 5 of 46 blind drugs (10.9%) are genuine prediction failures under cascade-aware scoring. Current SOM benchmarking standards compare predictions exclusively to final stable metabolites, systematically undercounting any predictor that identifies the initial enzymatic event. We propose cascade-aware benchmarking as a field-wide standard and invite independent blinded validation.