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

Enzyme-reaction retrieval prioritizes candidate biocatalysts for experimental follow-up, where recommendations should indicate catalytic compatibility and suitable pH and temperature conditions. Existing retrieval models optimize matching scores, whereas catalytic optima predictors are typically enzyme-level regressors because condition annotations are sparse. This separation leaves ranked pairs unevaluated for condition suitability and optima predictions without reaction context.

We present GERO, a multimodal fusion framework integrating enzyme sequence semantics, sequence-derived pocket geometry, and molecular reaction representations through feature-gated fusion for condition-aware retrieval and catalytic optima estimation. We define the tolerance-restricted hit rate (Hit@k-TR), which requires both correct top-k retrieval and condition predictions within predefined tolerances. Across enzyme- and reaction-similarity splits, GERO improves Hit@k-TR over two-stage baselines under the primary tolerances. Benchmark examples and an iodinin biosynthesis case study illustrate candidate ranking and assay-condition estimation for experimental prioritization.

Source code is available at https://github.com/ykxhs/GERO.

Supplementary data are available at XXXX online.