A substrate-first-enzymology deep-learning framework for transforming new-to-nature chemicals

Enzyme activity data favor familiar substrates and positive outcomes, creating survivorship bias that limits predictions for new-to-nature chemistry. Here we introduce substrate-first enzymology: a chemically diverse substrate array is designed to broaden coverage of a predefined, synthetically relevant chemical space, independently of expected activity. Alcohol dehydrogenases (ADHs) are systematically assayed against this array using a unified high-throughput assay. The enzyme–substrate matrix records activity labels and turnover frequency ( k cat ) values; substrate-resolved profiles form an ADH k cat atlas. A deep learning model trained on these data classifies activity and predicts k cat for active pairs. The model outperforms existing approaches under matched training and evaluation conditions and prospectively prioritizes active substrates from underrepresented chemical space. Without binding-pocket annotations, the model’s attention mechanism highlights substrate reactive centers and catalytically relevant residues. By designing substrate coverage before measurement and retaining both positive and negative outcomes, this framework reduces historical sampling bias and supports interpretable k cat prediction beyond historically sampled chemistry.