A Multimodal Ensemble Framework for Optimal Mutant Prediction and Computational Enzyme Engineering

Abstract

Abstract Exploring mutational landscape of proteins to engineer improved enzymes remains a fundamental challenge. Traditional methods, whether reliant on high‐throughput experimentation, direct evolution, or on computational prediction, often face challenges to effectively model the complex epistatic and long‐range interactions that related to protein function. To address this, we present GEMS, a robust framework for enzyme engineering that leverages the ensemble zero‐shot capabilities of multiple modalities. By integrating evolutionary, structural, and sequence‐based constraints, GEMS effectively models the sequence–structure–function relationship and predicts beneficial variants. Benchmarking against state‐of‐the‐art (SOTA) methods revealed that GEMS achieves competitive performance in ranking beneficial variants, generates highly informative initial variant libraries, highlighting its strength in capturing long‐range functional constraints. We rigorously evaluated GEMS on five diverse enzyme engineering cases under both pure enzyme (Caulobacter segniscarotenoid cleavage dioxygenase from Marine gamma proteobacterium ( Mgp CSO), aldehyde dehydrogenase from Dickeya parazeae ( Dp ADA)) and whole‐cell conditions (CYP105A3 from Streptomyces carbophilus (P450 Sca ‐2), O‐methyltransferase 1 from Cnidium monnieri ( Cm OMT1), prenyltransferase from Pastinaca sativa ( Ps PT2)). Our results demonstrate that GEMS successfully identifies activity‐enhancing mutations, with single variants exhibiting 1.1 to 3.2‐fold improvements in catalytic efficiency. Collectively, our findings prove GEMS as a powerful and versatile tool for advanced enzyme engineering.