plantbgc-demo 0.1.0

PlantBGC: Transformer-based biosynthetic gene cluster candidate detection for plant genomes

PlantBGC: Transformer for Plant BGC Discovery via Label-Free Domain Adaptation and Weak Supervision

PlantBGC detects candidate biosynthetic gene clusters (BGCs) in plant genomes by modeling genomes as ordered Pfam-domain sequences and scoring BGC-likeness with an encoder-only Transformer. The framework supports (i) microbial supervised pretraining, (ii) label-free plant domain adaptation via masked language modeling (MLM), and (iii) optional weak-supervision strategies to reduce primary-metabolism false positives while preserving secondary-metabolism signals.

This repository contains the training/evaluation scripts used in our paper: “PlantBGC” .

---

Key ideas (paper-level summary)

• Representation: genome → ordered Pfam tokens (domain sequence)
• Stage 1 (microbial supervision): train a Transformer detector on curated microbial BGC positives vs negatives
• Stage 2 (plant adaptation, no plant labels): continue pretraining on unlabeled plant Pfam sequences with MLM to align plant Pfam co-occurrence statistics
• Stage 3 (weak supervision, optional): inject soft negatives (e.g., GO/KEGG-based) to reduce “primary-like” false positives

---

What you get

• Candidate CDS and BGC loci
• BGC-likeness scores (Transformer-based) for ranking and triage
• (Optional) GO/KEGG proxy labels for analysis / weak supervision: (Details see GO/KEGG Proxy Labels.)
  • KEGG proxy is reported as primary / secondary / mixed / review
  • GO proxy follows the project label scheme

---

Installation

Option A: conda (recommended)

conda create -n plantbgc python=3.10 -y
conda activate plantbgc
pip install -r requirements.txt

Option B: pip (if you enjoy pain)

pip install -r requirements.txt

If you rely on external gene calling / Pfam annotation tools, install them separately and ensure they are on PATH.

---

Data format

Pfam-domain TSV (required)

We assume a tokenized TSV where each row is a Pfam hit for one protein, and proteins are ordered by genomic position within a contig/chromosome.

Minimum recommended columns

• sequence_id: ID for a continuous ordered region (e.g., contig/chromosome segment)
• pfam_id: Pfam accession (token)
• protein_id (or equivalent): optional but helpful for debugging
• start, end (optional): genomic/protein coordinates if available

If you already have a pipeline that outputs *.pfam.tsv, keep it. PlantBGC mainly cares about order + Pfam IDs + sequence grouping.

---

Data acquisition (NCBI / RefSeq)

If you want to test a new genome and obtain Transformer scoring results, download FASTA from NCBI as follows.

Plant genomes

Download the RefSeq CDS file:

• Genomic coding sequences (FASTA)

This file is recommended for plant runs because PlantBGC produces CDS/gene-level scores and then aggregates them into candidate loci.

Microbial genomes

Download the genome sequence file:

• Genome sequences (FASTA)

For microbial inputs, PlantBGC can score across the genome sequence workflow as supported by the repo configuration.

---

Inputs

PlantBGC typically uses:

Required

• FASTA
  • Plants: CDS FASTA (Genomic coding sequences (FASTA))
  • Microbes: Genome FASTA (Genome sequences (FASTA))

Recommended (for genomic coordinates)

• Genome annotations (e.g., GFF/GBFF)
  Used to map CDS units back to genomic coordinates for locus-level reporting.

Optional (for proxy labeling / weak supervision)

• GO/KEGG annotation tables
  Used for proxy labeling, weak supervision, and downstream enrichment summaries.

---

Outputs

PlantBGC outputs are organized at three levels:

1. Token / Pfam-level scores
2. CDS (gene)-level scores
3. Candidate loci (candidate BGC loci) aggregated from consecutive high-scoring CDS

---

Locus aggregation rule (paper-aligned)

A candidate locus is formed by:

• pooling Pfam-level scores into a CDS-level score
• selecting CDS with score > 0.5
• aggregating strictly consecutive selected CDS into loci
• enforcing minimum length = 3 CDS and gap = 0

---

GO/KEGG Proxy Labels (optional)

PlantBGC can optionally generate proxy labels from functional annotations to support analysis and weak supervision. In this project, KEGG proxy is reported as primary / secondary / mixed / review (paper-aligned).

GO proxy will be documented in a separate section (to be added).

KEGG proxy: how to obtain KO numbers

To produce KEGG proxy labels, you first need KO numbers (K identifiers) for genes/proteins in your predicted loci.

Recommended (simplest): KEGG web mapping (BlastKOALA / GhostKOALA)

Local KO mapping can be non-trivial due to database licensing/availability, profile resources, and environment setup. For most users, we recommend using the official KEGG KO assignment service:

1. Prepare a FASTA file of coding sequences/proteins to annotate
   • For plants, we typically start from RefSeq CDS FASTA and then use the translated proteins (or protein FASTA if available).
2. Upload the FASTA to BlastKOALA (or GhostKOALA for large-scale submissions)
3. Download the KO assignment output (gene/protein → KO number mapping)
4. Feed the KO mapping file into PlantBGC to generate locus-level KEGG proxy labels

PlantBGC will then aggregate KO-derived category signals within each candidate BGC locus and output a locus-level label: primary / secondary / mixed / review.

Optional: local deployment (advanced)

If you prefer a fully local pipeline, KO mapping can be performed with tools such as:

• KofamScan (profile HMM-based KO assignment using KOfam HMMs)
• other HMMER-based KO assignment workflows

For most workflows, the web-based KO assignment above is the fastest path to obtain KO numbers.

Generating KEGG proxy labels in PlantBGC

Once you have the KO mapping (gene/protein → KO), run PlantBGC’s proxy module to:

• map KO to functional categories used by our proxy scheme
• aggregate per-gene evidence into locus-level proxy labels
• export per-locus label summaries for downstream analysis

The exact command depends on the repo version. See scripts/ (or the pipeline entrypoint) for the KEGG proxy labeling utility.

Repository structure (suggested)

PlantBGC/
  models/                # model definitions
  data_utils/            # TSV parsing, tokenization, masking, batching
  scripts/               # training/adaptation/eval entrypoints
  outputs/               # predictions + evaluation results
  README.md
  requirements.txt

---

Citation

If you use this code in academic work, please cite:

@article{plantbgc2026,
  title   = {PlantBGC: Transformer-based Detection of Plant Biosynthetic Gene Clusters from Pfam-domain Sequences},
  author  = {Zhao, Yuhan and Guo, Zhishan and Sui, Ning and others},
  year    = {2026},
  note    = {Manuscript in preparation}
}

---

License

TBD (choose one: MIT / Apache-2.0 / BSD-3-Clause)

---

Acknowledgements

This project builds on ideas from genome mining and domain-sequence modeling, and is inspired by prior BGC detection pipelines in microbes.