PlantBGC

Plants build an enormous chemical repertoire — alkaloids, terpenoids, cyanogenic glycosides, defensive phenolics — and a share of those pathways are encoded by biosynthetic gene clusters (BGCs), contiguous runs of genes acting on one metabolite. Locating them is the entry point to natural-product chemistry, but the machine learning that transformed BGC discovery was built on bacteria and fungi. The reason is data: microbial clusters carry decades of curation, while validated plant BGCs number in the dozens — far too few for supervised detection at genome scale.

PlantBGC, from North Carolina State University, attacks that label scarcity by transfer rather than curation. Yuhan Zhao and colleagues, spanning the departments of Computer Science and of Molecular and Structural Biochemistry, represent a genome as an ordered sequence of Pfam protein-family domains and train an encoder-only Transformer to score how BGC-like each position is. Supervision comes from well-annotated microbial clusters; the model then moves into the plant domain without any plant labels, via masked language modeling over unlabeled plant genomes, then is sharpened with weak labels from GO and KEGG. The result is one fixed checkpoint applied genome-wide rather than refit per organism.

The representation is shared with BGC-Transformer, which also treats Pfam domains as vocabulary but trains a generative RoBERTa on bacterial and fungal genomes. PlantBGC is discriminative and targets the eukaryotic gap: the plant tools in wide use — plantiSMASH, PlantClusterFinder, PhytoClust — are rule- and signature-driven, making this a learned alternative where there was none.

#Key Features

• Genomes as domain sentences: Each gene contributes its Pfam domains as tokens, so context across a locus is modeled by self-attention rather than hand-written cluster rules.
• Label-free domain adaptation: Masked language modeling moves the microbially trained detector onto plant genomes without a single curated plant BGC label.
• Weak supervision from functional annotation: GO and KEGG terms become soft negatives marking primary-metabolism-like regions, suppressing the dominant false-positive mode in plants.
• Tighter locus boundaries: Against plantiSMASH on matched regions, predicted loci are markedly more compact, narrowing what a wet-lab group must clone or interrogate.
• One checkpoint, many species: The adapted model is applied genome-wide without per-species refitting.
• Released artifacts: The tagged release ships the final weights, the stage-1 pre-adaptation checkpoint, the microbial training corpus, and Pfam2vec embeddings; the plant RefSeq adaptation corpus is not included.

#Technical Details

The architecture is a stack of Transformer encoder blocks — multi-head self-attention plus position-wise feed-forward layers — over 100-dimensional Pfam2vec token embeddings, with a maximum input of 512 domain tokens and a linear-plus-sigmoid head scoring every token. Stage one trains on 1,420 microbial BGCs from MIBiG with a weighted binary cross-entropy loss; GeneSwap shuffles genes to build a negative set three times the positive set. Stage two freezes the lower blocks and continues pretraining the upper blocks with masked language modeling (mask fraction 0.15) on RefSeq coding sequences from 17 plant species. Stage three fine-tunes with a confidence-weighted soft-label objective built from GO and KEGG terms. The implementation is built on the DeepBGC codebase, retaining its package layout and third-party notices.

On microbial benchmarks the model reaches token-level AUC of 0.988 under 10-fold cross-validation and 0.979 under leave-class-out evaluation, ahead of a BiLSTM token classifier and a random forest on hand-crafted Pfam features. On 34 curated plant loci scored under a strict 100% coverage criterion, adaptation lifts known-BGC recovery from 29.4% to 67.6%. Weak supervision cuts the proxy primary-like ratio by 48.4% (GO) and 45.2% (KEGG), with consistent per-species reductions (paired Wilcoxon p = 1.53e-5), and predicted loci are shorter than plantiSMASH calls on 93.8% of matched pairs (median length ratio 0.278).

#Applications

The intended use is narrowing experimental search space in plant natural-product discovery: ranking candidate loci in a new assembly so a chemistry or metabolic-engineering group knows which regions are worth cloning, expressing heterologously, or knocking out. Tighter boundaries shrink the flanking sequence swept into each call, making construct design cheaper. The same scoring suits crop and medicinal-plant genomics, where these pathways underlie traits from pest resistance to pharmacologically active compounds.

#Impact

PlantBGC shows that microbial supervision transfers across a very large phylogenetic gap when the representation is functional domains rather than nucleotides. Code and weights are public, but the paper carries no availability statement, so the repository is found only by search. Licensing is inherited rather than authored: the LICENSE file is upstream DeepBGC's MIT text, copyright Merck Sharp & Dohme 2019, with no copyright line for the authors, and the README's License section reads "TBD". GitHub resolves the repository as MIT. Two caveats bound the work: it is a preprint awaiting peer review, and adaptation and locus-level evaluation draw on the same 17-species corpus, so performance outside that set is not measured. All evidence is in silico, with no experimental confirmation.