Learning Enzymatic Reaction Mechanisms with Bidirectional Flow Matching and Chemistry-Guided Search

Predicting enzymatic reaction mechanisms is a challenging structured-generation problem due to limited mechanistic data, reversible elementary steps, and the combinatorial space of chemically plausible intermediates. Existing enzyme-mechanism predictors largely rely on predefined reaction templates, limiting their ability to extrapolate to unseen chemistry. We introduce EZSolver, a generative framework that combines EZFlow, a flow-matching model of electron redistribution over Bond--Electron matrices, with evaluator-guided bidirectional search for multi-step mechanism inference. EZFlow transfers chemical knowledge from a pretrained organic-reaction model and combines standard and inverse flow to generate reversible elementary-step proposals without requiring target reaction templates. EZSolver then composes these proposals into complete pathways using bidirectional beam search and chemistry-aware evaluators. EZFlow recovers the correct elementary step for 72.4% of test cases with a sampling size of $N=200$. For full-mechanism prediction, EZSolver identifies Accurate mechanisms for 60.0% of reactions and Chemically Plausible mechanisms for 84.6% within the top-5 candidates. Under a strict out-of-distribution (OOD) benchmark, an ensemble of EZSolver runs achieves 52.7% full-mechanism accuracy, outperforming reported template-based baselines. These results suggest that a generative predictor combined with a structured, domain-constrained search can provide a mechanistic reasoning layer for biomolecular modeling beyond fixed reaction dictionaries.