Enzyme association for environmental biotransformation reactions through contrastive learning of reaction center-specific fingerprints.

Abstract

Microbial biotransformation plays a central role in the environmental degradation of chemical contaminants, driven by the catalytic activities of diverse enzymes. However, linking specific enzymes to contaminant removal and predicting associated transformation products (TPs) under real-world conditions remain a major challenge. In this study, we present a self-supervised, contrastive fine-tuning strategy for reaction fingerprint learning, designed to improve the chemical relevance of BERT-based reaction embeddings for environmental biotransformation reactions. Specifically, we fine-tuned a BERT encoder such that the cosine similarity between its reaction fingerprints aligns with the Tanimoto similarity of traditional structure-based fingerprints. The resulting compact, 256-dimensional fingerprints, termed crxnfp, showed an improved ability to cluster reactions according to transformation type and focus attention on chemically meaningful reaction centers. Our crxnfp fingerprints were further validated in reaction classification tasks across multiple datasets, achieving superior or comparable performance relative to existing methods. Importantly, they enabled a similarity-based association of biotransformation rules and reactions from enviPath with enzyme annotations from the Rhea and UniProt databases, offering a scalable approach to enrich environmental biotransformation datasets with enzymatic information. Additionally, crxnfp was employed to identify specific enzyme classes involved in contaminant biotransformation, which were subsequently validated through experiments conducted in this study, achieving 91.3% accuracy at the third-level enzyme classification. The crxnfp fingerprints offer a promising solution to advance the understanding of contaminant biotransformation and guide the development of enzyme-informed strategies for contaminant management across diverse environmental contexts. Code is available at https://github.com/zhangky12/crxnfp and https://github.com/zhangky12/crxnfp_knn.