BGC-Finder

Biosynthetic gene clusters (BGCs) are the co-localized gene sets microbes use to build antibiotics, immunosuppressants, and other secondary metabolites. They have been mined for two decades with alignment-based tools such as antiSMASH, which match candidate genes against curated profile HMMs — slow enough to consume hundreds of CPU hours at metagenome scale, and blind to enzymes whose sequences have diverged past recognition even when their chemistry is conserved. BGC-Finder, from Kang Ning's group at Huazhong University of Science and Technology, drops the alignment step: it represents a contig as an ordered sequence of protein language model embeddings and learns the grammar of gene clusters from that representation.

The system is packaged as an end-to-end pipeline, but its trained contribution is a single genomic context-aware module, the BGC-annotator. Gene calls from Prodigal (bacteria) or AUGUSTUS (fungi) are embedded with ESM-2, screened by the group's earlier BGC-Prophet detector, and surviving candidates re-embedded and passed to the annotator, which assigns every gene a functional role and the cluster a product class. Because the annotator attends over genes rather than residues, it reads signal no individual protein carries — that a transporter sits beside a polyketide synthase beside a regulator.

Posted to bioRxiv in April 2025 and revised that September, it has not been peer-reviewed. It joins BGC-Transformer among models treating a gene cluster as a sentence, but its vocabulary is continuous pLM embeddings rather than Pfam domain tokens.

#Key Features

• Alignment-free profiling: Swapping HMM search for a forward pass through two neural models cuts genome-scale profiling from 5–30 minutes to seconds, up to 102-fold faster than antiSMASH and DeepBGC.
• Gene-level functional annotation: Genes are labeled biosynthetic-core, tailoring, transport, regulation, or other; clusters fall into seven product classes spanning NRP, RiPP, polyketide, terpene, alkaloid, and saccharide.
• Attention-derived essentiality: Gene-to-class attention weights score each gene's contribution to the product-class call, recovering a cluster's functional hierarchy without biological priors, including the underappreciated weight of transporters.
• Embedding-guided clustering: Context-aware cluster embeddings track BiG-SCAPE similarity more closely than raw ESM-2 embeddings, so BGCs group by cosine distance rather than alignment.
• Cross-kingdom transfer: One checkpoint runs unchanged over bacterial genomes, fungal genomes, and metagenome-assembled genomes.

#Technical Details

The BGC-annotator is a five-layer RoBERTa-style encoder with hidden size 1280, ten attention heads per layer, relative key-query positional embeddings, and a fixed 128-gene context window; the released checkpoint holds roughly 112 million parameters. Each gene enters as a mean-pooled 1280-dimensional ESM2-650M embedding, concatenated with a zero-initialized class token and augmented with token-type and relative position embeddings, and a feed-forward head emits the gene-role and product-class predictions. Parameters were initialized from the genomic language model gLM and fine-tuned on all experimentally validated BGCs in MIBiG 3.1 — 2,502 clusters spanning 26,520 bacterial and fungal genes — with Adam at 5e-5 for up to 100 epochs with early stopping and 300 held-out clusters each for validation and testing.

On the MIBiG test set it reaches AUROC 0.990 and AUPR 0.986 for gene function, and beats GECCO, antiSMASH, BiGCARP, and DeepBGC on precision and F1 for boundary detection on the Nine-Genome benchmark. Against FunBGCs, 742 experimentally validated fungal clusters, core-gene retrieval reached accuracy 0.96, precision 0.95, recall 0.82, and F1 0.88, versus recall 0.77 for antiSMASH, and it correctly annotated 731 of 733 cytochrome P450 instances where antiSMASH made 159 errors. It missed non-canonical synthases such as phosphoenolpyruvate mutase and isocyanide synthase, and called terpene cyclase tailoring rather than core.

#Applications

The model targets natural-product discovery in genomes and metagenomes where sequence divergence defeats profile search. Across 210 Arctic and Antarctic marine metagenome-assembled genomes it recovered 1.2 times more clusters than the alignment baseline and prioritized 31 core genes whose relatives were detectable only structurally. A survey of 256 fungal genomes covering 197 species from FungiDB yielded 4,585 clusters with 6,388 core genes and an NRPS family shared between Fusarium and Aspergillus. One prediction was carried to the bench: knocking out a BGC-Finder-unique core gene in Aspergillus oryzae abolished wild-type-specific metabolite peaks under LC-MS. Hosted Colab notebooks let groups without GPUs annotate their own GenBank or FASTA files.

#Impact

BGC-Finder shows that a modestly sized supervised model over protein language model embeddings captures genomic context well enough for cluster mining without a large pretrained DNA language model such as Evo, and its attention maps and embedding space make it an interpretation tool, not only a classifier. The caveats are real: inference wants a high-memory GPU, the product-class vocabulary is bounded by MIBiG's seven categories, the "biosynthesis dark matter" it surfaces is alignment-undetectable rather than chemically unprecedented, and fungal pathways whose enzymes are not co-localized fall outside a 128-gene window. Weights are released on Hugging Face under MIT, the GitHub repository carries no license file, and the work remains a preprint awaiting peer review.