Predicting enzyme kinetic parameters — turnover number, Michaelis constant, inhibition constant — from sequence alone treats the enzyme as a static object. The substrate arrives, the active site is there or it isn't, and the prediction is a lookup. But enzyme catalysis is not a lock-and-key operation. The enzyme changes shape when the substrate binds. The active site adapts. The kinetic parameter is a property of the interaction, not of the enzyme.
Wang, Zheng, Li, Wei, Liu, Hu, Bao, and Yang reformulate kinetic prediction as staged multimodal conditional modeling. Their framework, ERBA, processes substrate specificity and conformational adaptation as separate sequential stages rather than merging everything into a single representation. First, Molecular Recognition Cross-Attention injects substrate information into the enzyme representation — learning which parts of the protein respond to which molecular features. Second, Geometry-aware Mixture-of-Experts integrates active-site structure, routing through specialists that capture the induced-fit deformation. The substrate changes what the model attends to; the geometry changes how it routes.
The approach delivers consistent gains over sequence-only and shallow-fusion baselines, with stronger out-of-distribution generalization — the harder test, where the enzyme-substrate pair has no close analog in training data.
The through-claim is about the staging. Previous models treated enzyme-substrate interaction as a single inference: concatenate representations, predict. ERBA separates the interaction into two phases that mirror the physical process — recognition first, adaptation second. The staging is not just an architectural convenience. It reflects a causal structure: the enzyme cannot adapt to a substrate it hasn't recognized, and the adaptation cannot be predicted without first modeling the recognition. When the causal structure of the phenomenon is built into the architecture rather than left for the model to discover, out-of-distribution performance improves because the model doesn't have to rediscover the staging from data.