ECHO: A multimodal deep learning framework for enzyme kcat prediction with pocket-aware and EC-weighted training

ECHO (EC-specific Catalytic Hybrid Optimizer) is a multimodal deep learning framework for predicting enzyme turnover numbers (kcat) to guide enzyme discovery in biosynthetic pathway engineering. ECHO integrates three complementary sources of biochemical information: (1) protein sequence embeddings from a fine-tuned ESM-2 language model, (2) substrate chemical representations from SchNet, and (3) catalytic pocket structural features encoded by a Geometric Vector Perceptron (GVP) network, all fused through a multi-head cross-attention mechanism. A novel training strategy incorporating EC-proximity and substrate-similarity weighting prioritizes samples relevant to target transformations, making ECHO particularly effective for pathway-specific enzyme prioritization under low-data regimes and for non-natural substrate transformations. ECHO was trained on 24,146 enzyme–substrate pairs covering 2,295 unique EC numbers from the CataPro kcat dataset. This release contains the trained model weights of ECHO, enabling users to perform enzyme activity prediction and candidate prioritization for their pathways of interest.