Joint enzyme-reaction retrieval and catalytic optima prediction via multimodal fusion.

GERO achieves significant improvements in condition-aware enzyme-reaction retrieval, with higher Hit@k-TR scores over two-stage baselines across enzyme- and reaction-similarity splits.

• Why it matters: Current models either focus on matching scores or enzyme-level regressors, leaving a gap in evaluating reaction context and condition suitability simultaneously, which is crucial for experimental prioritization.
• What they did: The authors developed GERO, a multimodal fusion framework that integrates enzyme sequence semantics, pocket geometry, and reaction representations, and evaluated it using the tolerance-restricted hit rate (Hit@k-TR) metric.
• The result: GERO's approach enhances candidate ranking and assay-condition estimation, demonstrated through benchmark examples and a case study on iodinin biosynthesis, facilitating more accurate enzyme-reaction predictions.