microbiome-informatics/sanntis-training-data

SanntiS training data

The data the SanntiS models are trained on: biosynthetic gene clusters (BGCs) from MIBiG as positives and a pool of BGC-free genome regions as negatives. sanntis train reads both formats directly; see the training documentation.

Layout

Path	Tag	Content
mibig/5.0rc1-train/	mibig-5.0rc1-train	MIBiG 5.0rc1 active + pending entries as a MIBiG-format archive pair (positives of the v1.0.0-beta.1 model)
negatives/v1/	negatives-v1	negative genome pool v1 (sanntis train --negatives v1)
npclassifier/v1/	npclassifier-v1	NPClassifier API answers for the compounds of mibig/5.0rc1-train (sanntis train --npc-cache v1)

mibig/5.0rc1-train/

File	Content
mibig_5.0rc1-train_json.tar.gz	the 2,795 MIBiG 5.0rc1 JSON entries, verbatim, one <accession>.json per BGC
mibig_5.0rc1-train_gbk.tar.gz	one GenBank record per BGC, <accession>.gbk, every CDS with a /translation: the MIBiG 4.0 GBK when the first locus is unchanged since 4.0 (2,621 entries), otherwise the first 5.0rc1 locus cut from the NCBI nuccore record (174 entries; /pseudo CDS, CDS cut by the slice and repeated CDS locations dropped)
manifest.tsv	per BGC: status, GBK source, locus, whether the locus changed since 4.0, length, CDS count, md5 of the sequence and of the proteins
excluded.tsv	the 2 entries without a usable record (BGC0003182, BGC0003249)
provenance.json	sha256 of the MIBiG 5.0rc1 JSON and 4.0 JSON / GBK archives and of the nuccore records the pair was built from; counts; SanntiS and Biopython versions

MIBiG 5.0rc1 ships JSON only. pending entries were active in MIBiG 4.0 and are under re-review in 5.0; they are included because the model is trained on the validated 4.0 pool plus the new 5.0 entries. sanntis train drops the 18 loci longer than 200 kb, so the model is trained on 2,777 BGCs. The archives are built, deterministically, by the runbook of objective 01 of the sanntisv1-data-and-analysis repository, which also holds the training run and its validation report; manifest.tsv is what a rebuild must reproduce.

negatives/v1/

Genomes with their native BGCs removed, annotated with InterProScan 6.0.3.0 / InterPro 107.0:

	Genomes	CDS	Clean runs
Prokaryotes: GTDB r226 species representatives, balanced over phyla (seed 42)	205	530,698	47,819
Eukaryotes: RefSeq fungi and protists (3 per fungal phylum, 2 per protozoan phylum)	23	220,759	7,844

Regions detected by antiSMASH 8.0.1 (--minimal) are masked (native_mask.parquet); the remaining CDS form the BGC-free runs (clean_runs.parquet) that give the training background and the length-matched negative chunks. The longest run of at least 50 CDS of 55 prokaryotic and 15 eukaryotic genomes is reserved as a test backbone (test_backbones.parquet) and kept out of training. cds_tokens.parquet holds the InterPro tokens per CDS and manifest.json the pins, counts, genome-wide annotation coverage and the sha256 of every file, which sanntis train verifies. The pool is rebuilt with sanntis build-negatives; the genome selection and the build are objective 00 of the sanntisv1-data-and-analysis repository.

npclassifier/v1/

The answers of the NPClassifier API (models CLASS v1, PATHWAY v1, SUPERCLASS v1, ontology npclassifier_index_v1) for the 3,842 structure skeletons of the compounds of mibig/5.0rc1-train, one JSON file per canonical SMILES named by its sha1, in the format of $SANNTIS_CACHE/npclassifier. sanntis train --npc-cache v1 seeds its cache with them, so a retraining makes no API request for these compounds and gets the labels the released model was trained with. The answers were collected up to 2026-10-30. NPClassifier: Kim et al., Journal of Natural Products 2021, doi:10.1021/acs.jnatprod.1c00399.

Use

pip install "sanntis[train] @ git+https://github.com/Finn-Lab/SanntiS@v1.0.0-beta.1"
hf download microbiome-informatics/sanntis-training-data --repo-type dataset --include "mibig/5.0rc1-train/*" --local-dir data/
sanntis train --run runs/release --mibig none \
    --external-gbk data/mibig/5.0rc1-train/mibig_5.0rc1-train_gbk.tar.gz \
    --external-json data/mibig/5.0rc1-train/mibig_5.0rc1-train_json.tar.gz \
    --negatives v1 --npc-cache v1 --container-engine docker

--negatives v1 and --npc-cache v1 fetch negatives/v1/ and npclassifier/v1/ at their tags from this dataset.

Sources and licence

• MIBiG 5.0rc1 and 4.0 (mibig.secondarymetabolites.org), CC BY 4.0. Terlouw et al., Nucleic Acids Research 2023, doi:10.1093/nar/gkac1049.
• NCBI nuccore records of the 174 re-cut loci; GTDB r226 and NCBI RefSeq assemblies of the negative pool.
• Derived files in this dataset are released under CC BY 4.0. Please cite SanntiS and MIBiG.

@article{sanchez2023sanntis,
  title   = {Expansion of novel biosynthetic gene clusters from diverse environments using SanntiS},
  author  = {Sanchez, Santiago and Rogers, Joel D. and Rogers, Alexander B. and Nassar, Maaly and McEntyre, Johanna
             and Welch, Martin and Hollfelder, Florian and Finn, Robert D.},
  journal = {bioRxiv},
  year    = {2023},
  doi     = {10.1101/2023.05.23.540769}
}