Microenvironment Awareness Pragmatic Analysis for Generative Enzyme Function Reasoning

Accurate enzyme function annotation is critical for interpreting rapidly expanding sequence data, yet experimental methods are costly and low-throughput, while computational approaches face two key limitations. Most existing methods are discriminative models confined to known EC labels, making it difficult to generalize to emerging enzymatic activities under cold-start conditions. Moreover, they typically treat enzymes as isolated molecular entities and overlook microenvironment factors such as optimal pH and temperature, which provide important contextual clues for functional annotation.

This study proposes MPA-EC, a generative framework that integrates microenvironment as pragmatic context into enzyme function reasoning. By fine-tuning the large language model on a newly constructed Enzyme Microenvironment Database, the model learns to infer microenvironment from sequence and then reason function within that context, generating EC labels with explicit reasoning chains. MEMP imputes missing pH and temperature values to complete the Enzyme Microenvironment Database. On public benchmark dataset, MPA-EC achieves an absolute F1 improvement of 0.1068 over the state-of-the-art GraphEC model. Under cold-start conditions, the model generates reference-worthy function labels, demonstrating its practical utility for emerging enzymatic activities.

The source code and data of  MPA-EC  can be accessed at  https://bliulab.net/MPA-EC,  https://zenodo.org/records/22152329  and  https://github.com/bliulab/MPA-EC.