Retrieval-Augmented Multimodal Learning for Enzyme–Substrate Interaction Prediction Under Low-Homology Shift

Predicting whether a specific enzyme will react with a given substrate is a critical task for biocatalyst discovery. However, existing models often struggle with two main issues: they fail to capture the complex, directional compatibility between enzymes and substrates, and their performance drops significantly when the test enzymes have low sequence identity to those seen during training (a phenomenon known as low-homology distribution shift).

The authors propose RAMMESI, a framework designed to address these challenges through three key innovations:

1. Multimodal Interaction Modeling: Instead of simple concatenation, RAMMESI uses a fusion module that models enzyme-to-substrate and substrate-to-enzyme interactions separately, alongside a joint complex-level transformer, to capture fine-grained compatibility.
2. Inference-Time Retrieval: To handle low-homology shifts, the model retrieves similar enzymes from a database at inference time. It then predicts the interaction between these retrieved enzymes and the query substrate, aggregating these predictions to provide robust contextual evidence.
3. Imbalance-Aware Training: To combat the scarcity of positive interaction data, the model employs a weighted binary cross-entropy objective that reduces the influence of easy negative examples.

Experiments on standard benchmarks show that RAMMESI consistently outperforms existing methods, particularly in challenging scenarios where test enzymes share very low sequence identity with training data. The retrieval module acts as a plug-and-play component, meaning it can be added to other existing ESI backbones to improve their robustness without requiring retraining of the base models.

This work provides a scalable, robust solution for biocatalyst screening. By leveraging both deep multimodal representation learning and retrieval-based inference, the framework bridges the gap between limited experimental data and the vast space of potential enzyme-substrate interactions, making it a valuable tool for industrial and academic biochemical research.