Evaluating Multi-Modal Models for Enzyme-Reaction Retrieval
Abstract
Identifying functional enzymes that can perform unannotated reactions is a major biotechnological bottleneck. While multi-modal machine learning models can be used to retrieve enzymes given target functions (reactions), existing methods have not been adequately compared to each other. Two key areas warrant further investigation: first, the optimal way to incorporate 3D protein structure and predicted binding pockets for enzyme retrieval, and second, the most effective learning objectives for training such multi-modal models. We examine these questions through experiments on Task 2 of Classification and Retrieval for Enzymes (CARE) benchmark, demonstrating that multi-modal representations combining protein structure with pocket information have better performance than sequence-only methods. Second, we evaluate learning objectives and found that contrastive learning generally provides superior performance for enzyme retrieval compared to a binary classification. Our work underscores the value of integrating structural and pocket information for precise enzyme-reaction matching and offers insights into effective training objectives for such retrieval models.
