Curcumin is a representative aromatic β-diketone natural product, and its reduced derivatives have potential applications in functional foods, cosmetics, and pharmaceutical intermediates. CurA catalyzes the NADPH-dependent sequential reduction of curcumin to dihydrocurcumin and tetrahydrocurcumin under mild conditions. However, the sequence diversity and cofactor preferences of CurA-related enzymes remain insufficiently explored, limiting the identification of variants suitable for economical biocatalysis. Identifying divergent curcumin-reducing enzymes from underexplored resources is therefore important for expanding the biocatalytic toolkit. Reports that marine-derived strains can convert curcumin to tetrahydrocurcumin suggest that marine microbial resources may contain unexplored CurA-related enzymes.
Here, we developed an AI-guided multimodal protein-mining workflow that integrates protein language model embeddings, unsupervised clustering, structure prediction, binding-pocket comparison, and catalytic-domain electric-field analysis to identify CurA-related candidates in the Global Ocean Protein Catalog (GOPC). Approximately 40,000 candidate sequences were dereplicated at 90% sequence identity. ESM-derived embeddings of the nonredundant sequences were reduced by PCA and clustered using HDBSCAN, yielding 32 clusters. The medoid of each cluster was modeled using AlphaFold 3 and compared at the levels of overall structure and predicted substrate-binding pocket. Despite sharing only 45% average sequence identity with reported CurA enzymes, the representatives showed high structural similarity, with an average TM-score of 0.70, supporting the workflow's ability to identify structurally conserved remote homologs.
Based on sequence diversity, structural similarity, and catalytic-domain electric-field similarity scores (>0.80), five representatives were prioritized for experimental validation. Preliminary assays showed that one candidate displayed curcumin-reducing activity with NADH. Its activity with NADH was 106% of that measured with NADPH, indicating markedly greater NADH compatibility than previously reported CurA-related enzymes. Together, these results demonstrate that AI-guided multimodal mining can uncover remote, functionally relevant enzyme candidates from marine metagenomes and provide a promising starting point for developing curcumin-reducing biocatalysts compatible with the lower-cost cofactor NADH.
Keywords: Artificial intelligence; marine metagenome; ene-reductase; curcumin reduction; NADH-dependent biocatalysis