f-BGM

Fungi are a rich but under-exploited source of therapeutic small molecules — penicillin, ciclosporin, and pravastatin all come from fungal secondary metabolism, and roughly a quarter of approved natural-product drugs are fungal in origin. Those metabolites are built by biosynthetic gene clusters (BGCs): groups of co-localized genes encoding the enzymes that assemble and tailor a metabolic scaffold. Nearly every computational BGC finder, however, was designed around bacterial genomes. Rule-based tools such as antiSMASH and machine-learning detectors such as ClusterFinder, DeepBGC, GECCO, and TOUCAN inherit bacterial assumptions about cluster architecture and core-enzyme repertoire, which limits their reach in fungi.

f-BGM, developed at Peking University and posted as a bioRxiv preprint in July 2025, is a deep learning framework built specifically for fungal genome mining. Its backbone is f-DLC, a self-attention model pretrained on fungal genomes to capture the inter-domain locally co-occurrent relationships that characterize fungal genomic neighborhoods. That single pretrained representation is then reused for two downstream tasks — detecting BGCs in raw genomic context and identifying the core enzymes inside a given cluster — rather than being retrained from scratch for each.

Alongside f-DLC, f-BGM incorporates f-ESM2, an ESM-2 protein language model fine-tuned on fungal protein sequences, so that predictions draw on both domain-architecture context and residue-level sequence signal.

#Key Features

• Fungi-specific pretraining: f-DLC is pretrained on fungal genomes drawn from EnsemblFungi and JGI MycoCosm, learning which Pfam domains co-occur in local genomic windows instead of importing bacterial priors.
• Hierarchical domain-to-protein tokenization: Open reading frames are linearized into Pfam domain tokens with a per-ORF prefix token, letting one model reason at domain, protein, and cluster level simultaneously.
• One backbone, two tasks: The same pretrained encoder supports ORF-level BGC membership scoring and binary classification of seven core-enzyme families found in fungal BGCs, including polyketide synthases, non-ribosomal peptide synthetases, PKS-NRPS hybrids, terpene cyclases, and prenyltransferases.
• Attention-based interpretability: Attention weights are surfaced as per-domain and per-protein importance and as inter-domain partnerships, exported as interactive HTML views for each predicted cluster.
• Complete mining pipeline: The command-line tool accepts GenBank or FASTA input, calls genes with AUGUSTUS when no GFF3 annotation is supplied, and reports putative clusters with genomic locus, member ORF count, predicted core enzymes, Pfam composition, and a confidence score.

#Technical Details

f-DLC accepts up to 26 consecutive ORFs. Each token is embedded in 72 dimensions: 64 learnable domain-identity dimensions, four encoding a domain's relative position within its protein, and four encoding domain enrichment in BGCs (enabled only for the downstream models). Four stacked transformer encoders apply different attention masks in sequence — intra-protein domain interaction, multi-protein domain interaction, domain-to-protein aggregation that yields protein-level embeddings, and a reverse protein-to-domain pass used for BERT-style pretraining in which 15% of domain tokens are masked and recovered through a multi-class head.

The detection model wraps this backbone in three stages: a short-range module that fragments genomic context into multiple window sizes, runs f-DLC over each and concatenates the result with broadcast f-ESM2 embeddings before a protein-level transformer encoder; an LSTM long-range module; and an output module producing ORF-level probabilities. The core-enzyme model mirrors the short-range module without windowing, since cluster borders are already defined. Evaluation uses AUPRC and recall at a 5% top-ratio threshold against ClusterFinder, DeepBGC, GECCO, and TOUCAN under in-distribution, cross-secondary-metabolite-class, and cross-dataset schemes, with FunBGCs as the development corpus and MIBiG held out for the cross-dataset test.

#Applications

f-BGM is aimed at natural-product chemists and fungal genomicists deciding which clusters in a newly sequenced genome are worth cloning, heterologously expressing, or knocking out. The interpretability outputs matter here: rather than a bare score, a user sees which domains and proteins drove a call, which is what makes a candidate cluster arguable at the bench. Case studies on marine fungi surfaced biosynthetic potential that antiSMASH did not flag, supported by experimental and computational validation. A hosted webserver provides inference without local installation; the command-line toolkit is Linux-only because of platform-restricted dependencies including PyHMMER.

#Impact

f-BGM addresses a real gap: fungal genome mining has largely relied on tools calibrated for bacteria, and a fungi-specific pretrained representation is a credible way to close it. The work remains a preprint awaiting peer review, and the repository is small and not yet widely adopted. Source code is MIT-licensed on GitHub, but trained parameters and the Pfam support files are deposited separately on figshare and must be downloaded before the pipeline will run. One structural limitation follows from the design: because the backbone reads proteins as sequences of Pfam domains, ORFs with no domain annotation contribute no domain-level signal, leaving the f-ESM2 branch to carry them.