Deciphering Biosynthetic Gene Clusters with a Context-aware Protein Language Model

Microbial secondary metabolites, synthesized by biosynthetic gene clusters (BGCs), offer vast potential for biotechnological applications. Among BGC profiling techniques, computational detection methods face challenges, including time-consuming alignment and reliance on predefined profiles. To address these, we present BGC-Finder, an end-to-end pipeline utilizing protein language models for BGC detection and annotation from microbial genomes and metagenomes. This approach achieves remarkable increase in profiling speed of up to 100-fold, and employs genomic context-aware modeling to facilitate interpretable genetic essentiality assessment and large-scale BGC clustering. BGC-Finder outperformed traditional methods, successfully detecting 9.49% more biosynthetic-core genes and 27.70% more cytochrome P450s in 742 experimentally-validated BGCs. Notably, it retrieved 31 remote biosynthetic homologs from 210 polar marine metagenomes and identified 4,585 BGCs with 6,388 core genes from 256 fungal genomes. These findings highlight BGC-Finders capability to illuminate "microbial biosynthesis dark matter" (sequence-unrelated, function-similar biosynthetic enzymes) and expedite natural product discovery. HighlightsO_LIBGC-Finder is an accurate and ultrafast pipeline leveraging protein language models (pLMs) to predict and annotate biosynthetic gene clusters (BGCs) from microbial genomes and metagenomes. C_LIO_LIThe genomic context-aware model enables interpretable analysis: attention-driven identification of essential biosynthetic genes and embedding-guided BGC clustering. C_LIO_LIBGC-Finder sensitively retrieves remote homologous BGCs from both bacteria and fungi genomes, uncovering hidden microbial biosynthesis dark matter. C_LIO_LIWe discovered a non-ribosomal peptide synthetase (NRPS) family, which involved into function-specific BGCs in two evolutionarily distant fungi. C_LI