Enhancing biocatalytic retrosynthesis with a graph-to-graph model

Biocatalytic synthesis offers a green and sustainable route for chemical production, yet the rational design of biocatalytic routes remains challenging due to the need to jointly consider reaction feasibility and enzymatic compatibility. Here, we present BioG2G_ESR, a unified framework centered on graph-to-graph modeling for single-step biocatalytic retrosynthesis, with enzyme sequence recommendation as a downstream extension. BioG2G formulates retrosynthesis as a graph-to-graph translation task, directly predicting reactant molecular graphs from product structures while preserving molecular topology and stereochemical consistency. Extensive benchmark evaluations show that BioG2G achieves strong performance across biochemical retrosynthesis benchmarks, reaching Top-1 accuracies of 30.3% on Biochem-Plus and 55.0% on Biochem-Full. Representative literature-derived reactions absent from the Biochem-Plus training set are further examined as qualitative external examples. Building upon the predicted reactions, the enzyme sequence recommender (ESR) ranks candidate enzyme sequences primarily according to reaction similarity, with template-query matching used as an auxiliary signal for confidence stratification. Leave-one-out evaluation shows a clear association between template-match scores and Top-k accuracy across confidence groups. Together, BioG2G_ESR provides a unified and extensible framework for data-driven biocatalytic design, with graph-based retrosynthesis as the central component and enzyme recommendation as a complementary capability.