BGC-Bench (MIBiG 4.0)

BGC-Bench evaluates frozen biosynthetic gene cluster (BGC) representations on five tasks under random and out-of-distribution splits. Linear probes score four of the tasks and a linear contrastive head the fifth (retrieval), giving fifteen evaluation columns. This card describes release v1.0.0.

Existing BGC benchmarks mostly classify clusters into a few coarse classes on small curated datasets, which separates models poorly and says little about downstream use. BGC-Bench adds finer and chemical targets, BGC-to-molecule retrieval and A-domain substrates, scored under random, scaffold, gene-cluster-family and class holdouts with bootstrap intervals, so that encoders can be compared with statistical confidence under distribution shift.

The dataset (about 220 MB) contains MIBiG 4.0 inputs, labels and splits: ordered Pfam domains, CDS proteins, cluster DNA and A-domain sequences. The code at https://github.com/Nugkta/BGC-Bench downloads a pinned revision.

git clone https://github.com/Nugkta/BGC-Bench && cd BGC-Bench && uv sync
uv run bgcbench run --encoder deepbgc          # a built-in encoder (runs on the CPU)
uv run bgcbench evaluate --embeddings my.npz   # your own vectors (bgc_ids + embeddings)

The five tasks (fifteen columns)

task	question	metric	splits (columns)
coarse class	which of six MIBiG classes (multi-label)	macro-AUROC, one-vs-rest logistic regression	t1ml_random, t1ml_scaffold_holdout, t1ml_gcf_holdout_c07 (R, S, G)
fine class	which of 48 NPClassifier classes the cluster's products belong to (multi-label)	weighted macro-AUROC, one-vs-rest logistic regression	t2ml_stratified_5fold (fold 4 = test), t2ml_scaffold_holdout, t2ml_gcf_c07_holdout
chemical properties	13 RDKit descriptors of the first product	mean Spearman rho, ridge regression	t1_random, t1_scaffold_holdout, t1_gcf_holdout_c07
retrieval	rank the cluster's products in the test-molecule gallery	class-macro R@10 (a query is a hit when any of its products ranks in the top 10; ties scored by expectation), linear contrastive head on frozen MoLFormer molecule vectors, mean of 3 seeds	t3_random, t3_scaffold_holdout, t3_gcf_holdout_c07, t3_class_holdout (7 folds)
A-domain substrate	substrate (an amino or aryl acid) activated by an NRPS adenylation domain (28 classes)	top-1 accuracy, logistic regression	t5_bgc_grouped_5fold, t5_gcf_holdout_c07

Split axes. Random: 80/20 over BGCs (seed 42); fine classes test fold 4 of a five-fold split stratified on the 48 labels, and A-domain substrates use all five folds, grouped by BGC. Scaffold holdout: 20 % of Bemis-Murcko scaffold groups of the first product go to test (ring-free products form one group). GCF holdout: 20 % of BiG-SCAPE 2.0.3 gene cluster families at cutoff 0.7 go to test. Class holdout (retrieval): each MIBiG class, or the multi-class "Hybrid" group, is held out in turn. Each holdout controls its own axis only: a scaffold-holdout test BGC can share a family with training, and vice versa.

Files

path	content
splits/	the 15 split files (bgc_id or domain_id; split, fold, or the held-out primary_class for the retrieval class holdout; labels)
tasks/t2ml_labels.csv	fine-class labels of the 1,710 BGCs with a Pfam domain and a classified product: the NPClassifier classes of all products (classes_all), those among the 48 (labels), and whether the BGC is scored (enrolled, 1,221)
tasks/chem_descriptors.csv	13 RDKit 2026.3.1 descriptors of the first product SMILES, 1,944 BGCs
tasks/t3_bgc_compound_pairs.csv	(BGC, compound) pairs: compound_id = cmp_ + sha1(canonical SMILES)[:12], mibig_order = rank among the BGC's distinct products by first listing, multiplicity = times listed
tasks/t3_universe.csv	the 1,939 BGCs scored in retrieval (3,426 pairs, 3,110 molecules)
tasks/compounds.npz	compound_ids, canonical_smiles and 768-d MoLFormer vectors (ibm/MoLFormer-XL-both-10pct @ 7b12d946, pooler output) of the 3,118 distinct products
tasks/compounds_extraction_report.json	how the MoLFormer vectors were made (model, revision, library versions, pooling) and the count checks they pass
tasks/chem_encoders/<name>.npz	compound_ids and embeddings of the same 3,118 molecules, in the order of tasks/compounds.npz, from the paper's chemical-encoder comparison: molformer (768-d), chemberta2_mtr (ChemBERTa-2 77M-MTR, 384-d), unimol (Uni-Mol v1, 512-d), ecfp4 (2,048 bits), atompair (2,048 bits), maccs (167 keys), rdkit_desc (209 RDKit descriptors) and a random control (768-d), each also as <name>_pca128 (PCA to 128 dimensions, fitted on all 3,118 molecules)
tasks/chem_encoders/encoders.json	per file: display name, family, native or PCA, width, how the vectors were computed and, for the neural encoders, the Hub repository and revision of the weights
tasks/adomain_instances.csv	2,504 NRPS A-domains: parent BGC and gene, PF00501 envelope (env_start, env_stop, aa_seq), Stachelhaus code, substrate, position of the domain's token in the Pfam lists (token_index in the named list, token_index_raw in the accession list)
tasks/adomain_nucleotide.json	the codons of each A-domain envelope (back-mapped through the parent CDS, translation-checked)
inputs/pfam_domains.json	ordered Pfam accessions per BGC (2,631 BGCs; 5 have none)
inputs/pfam_domains_named.json	the same lists as Pfam names (2,626 BGCs), for name-keyed models (BiGCARP, Kawano)
inputs/pfam_acc_to_name.json	Pfam accession to name map (24,076 families)
inputs/protein_sequences.json	CDS translations with coordinates per BGC (2,631 BGCs)
inputs/nucleotide_seqs.json	cluster DNA per BGC (2,636 GenBank records)
inputs/bag_of_pfam_vocab.json, inputs/pfam_rand_vectors.npz	fixed files of the two Pfam baselines (3,932 accessions; N(0, 1) 256-d vectors, seed 42)
gcf/	BiG-SCAPE 2.0.3 family assignments at cutoffs 0.3-0.9 (the benchmark uses 0.7) and a per-cutoff summary
provenance/	MIBiG 4.0 metadata (classes, products, organism), NPClassifier labels and Murcko scaffolds per product; the sources of the labels and splits, not read by the scorer
manifest.json	every file with its size in bytes
croissant.json	Croissant 1.1 metadata with Responsible AI fields: one file object per file (description, size, sha256) and one record set per CSV

How it was built

• MIBiG 4.0 (Zdouc et al., Nucleic Acids Res. 53:D678, 2025). Annotations from the release mibig_json_4.0.tar.gz (dl.secondarymetabolites.org): 3,013 entries, of which the 2,437 active ones carry the labels. Sequences from the 4.0 GenBank release (mibig_gbk_4.0.tar.gz, Zenodo record 14835872): 2,636 records, the 2,437 active entries plus 199 pending ones.
• Cohorts. 2,636 DNA records; BiG-SCAPE's 500 kb length cap leaves 2,631 for the Pfam and protein inputs; 5 of those have no Pfam domain (2,626 named lists); 2,433 active BGCs have a family; 1,944 active BGCs have a product SMILES; 1,939 of them are embedded by every encoder of the paper and form the retrieval cohort.
• Classes. MIBiG 4.0's six classes (PKS, NRPS, ribosomal, terpene, saccharide, other; stored as Polyketide, NRP, RiPP, Terpene, Saccharide, Other). 456 BGCs have more than one; they are positive for each in the coarse task and form "Hybrid" where one label is needed. The chemical-property scaffold split leaves out the 61 Saccharide-only BGCs.
• Fine classes. NPClassifier (Kim et al. 2021; GNPS2 API, queried 2026-04-06 and 2026-10-08) classes of every product ("unclassified" is not a class). A BGC's labels are the union over its products, kept to the 48 classes held by at least 10 BGCs; the 1,221 BGCs with at least one of them and at least one Pfam domain are scored. Every split trains all 48 classes; a class without a test positive on a split (4 on the scaffold holdout, 8 on the GCF holdout) is not scored there. The score is the per-class AUROC averaged with weights equal to each class's test positives (weighted macro-AUROC); micro-AUROC, micro-AUPRC and unweighted macro-AUROC are reported with it.
• Molecules. SMILES from the 4.0 JSON in listing order, canonicalised with RDKit 2026.3.1 defaults (stereo tags kept); one row per distinct (BGC, molecule).
• Pfam domains. BiG-SCAPE 2.0.3 (pyhmmer, Pfam trusted cutoffs, domain overlap 0.1) on Pfam-A.hmm as downloaded on 2026-04-06 (27,481 profiles; entries dated to July 2025); domains ordered by CDS start, then envelope start. The accession-to-name map holds the ACC and NAME fields of an earlier Pfam-A release (24,076 families, version suffixes removed); the 739 domains (0.7 %) of families it lacks are left out of the named lists.
• Families. BiG-SCAPE 2.0.3 cluster on the GenBank records: region level, glocal alignment, mix weights, singletons included, affinity propagation; cutoffs 0.3-0.9.
• A-domains. MIBiG 4.0 module annotations name each A-domain's gene and substrate but give no coordinates; the i-th annotated A-domain of a gene is paired with the i-th PF00501 hit from the N terminus, and genes whose counts disagree are dropped (2,504 domains, 529 BGCs). Scored: single-substrate domains whose substrate occurs at least 10 times (level_b, 28 classes, 2,027 domains); folds are stratified on the NRPSPredictor2 substrate group (substrate_group) and grouped by BGC. evidence_tier collapses MIBiG's evidence codes (experimental > structure_inferred > homology > sequence_predicted; none when MIBiG gives no code).

Intended use and limitations

The benchmark compares frozen BGC feature extractors on the five tasks above. Its scores rank encoders on MIBiG-derived labels; they are not established as estimates of absolute performance on other clusters (for example from metagenomes or unannotated genomes), for product-structure prediction or design, or for decisions outside model comparison.

• MIBiG is small and curated: it over-represents well-studied taxa (755 of the 2,437 active entries come from Streptomyces) and well-studied classes.
• Labels inherit their tools' assumptions: NPClassifier for fine classes, RDKit for descriptors and scaffolds, BiG-SCAPE for families, a count-matching rule for A-domains (some genes with integer-named modules may be mis-paired).
• Report overlap between an evaluated model's training labels and the benchmark labels, including supervised training on MIBiG products.
• Scores rest on 1,221 to 2,437 labelled rows per task; compare encoders with the bootstrap intervals and the holdout columns, not the random split alone.
• The data describe microbial gene clusters and their products and contain no personal data. Better BGC representations serve natural-product discovery; the benchmark adds no capability for designing harmful compounds beyond what MIBiG and the evaluated models already provide.

Licence and attribution

The authors' files (splits, task tables, derived inputs) are released under CC BY 4.0. Third-party content keeps its source's terms:

• MIBiG 4.0 annotations and sequences: CC BY 4.0; please cite MIBiG together with BGC-Bench.
• Pfam accessions, names and domain annotations (inputs/pfam_*, the Pfam-derived columns of tasks/adomain_instances.csv): Pfam (Mistry et al. 2021) data are freely available under CC0; the annotations were computed with BiG-SCAPE 2.0.3 (Navarro-Muñoz et al. 2020), which also produced the families in gcf/.
• NPClassifier classes (provenance/npclassifier_labels.csv, the fine-class labels): output of the NPClassifier web service (Kim et al. 2021), which declares no licence; cite NPClassifier when using them.
• Molecule vectors: MoLFormer (ibm/MoLFormer-XL-both-10pct, Apache-2.0), Uni-Mol (dptech/Uni-Mol-Models, MIT) and ChemBERTa-2 (DeepChem/ChemBERTa-77M-MTR, no licence declared upstream; the two chemberta2_mtr files are included to reproduce the paper's chemical-encoder comparison, and CC BY 4.0 covers them only as far as the authors hold rights in them); fingerprints and descriptors from RDKit 2026.3.1 (BSD-3-Clause).

Citation

@inproceedings{xiao2026bgcbench,
  title     = {Data-Scarce, Multi-Granular, Multimodal: Benchmark Design for Biosynthetic Gene Clusters},
  author    = {Xiao, Hanlin and Takano, Eriko and {\'A}lvarez, Mauricio A. and Breitling, Rainer},
  booktitle = {Advances in Neural Information Processing Systems (Evaluations and Datasets Track)},
  year      = {2026}
}
@article{mibig4,
  title   = {{MIBiG} 4.0: advancing biosynthetic gene cluster curation through global collaboration},
  author  = {Zdouc, Mitja M. and others},
  journal = {Nucleic Acids Research}, volume = {53}, pages = {D678--D690}, year = {2025},
  doi     = {10.1093/nar/gkae1115}
}
@article{mistry2021pfam,
  title   = {{Pfam}: The Protein Families Database in 2021},
  author  = {Mistry, Jaina and Chuguransky, Sara and Williams, Lowri and others},
  journal = {Nucleic Acids Research}, volume = {49}, number = {D1}, pages = {D412--D419}, year = {2021},
  doi     = {10.1093/nar/gkaa913}
}
@article{navarro2020bigscape,
  title   = {A Computational Framework to Explore Large-Scale Biosynthetic Diversity},
  author  = {Navarro-Mu{\~n}oz, Jorge C. and Selem-Mojica, Nelly and Mullowney, Michael W. and others},
  journal = {Nature Chemical Biology}, volume = {16}, number = {1}, pages = {60--68}, year = {2020},
  doi     = {10.1038/s41589-019-0400-9}
}
@article{npclassifier,
  title   = {{NPClassifier}: A Deep Neural Network-Based Structural Classification Tool for Natural Products},
  author  = {Kim, Hyun Woo and Wang, Mingxun and Leber, Christopher A. and others},
  journal = {Journal of Natural Products}, volume = {84}, number = {11}, pages = {2795--2807}, year = {2021},
  doi     = {10.1021/acs.jnatprod.1c00399}
}

Versions and maintenance

This is release v1.0.0. Each release is a tag of this repository; tags are not moved or deleted, and a corrected or extended dataset gets a new tag. The code pins the tag it was tested with (DATASET_REVISION in src/bgcbench/__init__.py), so published results stay reproducible. Maintained by the authors at the University of Manchester; report problems as issues at https://github.com/Nugkta/BGC-Bench.