SanntiS: biosynthetic gene cluster detection

SanntiS detects biosynthetic gene clusters (BGCs) in prokaryotic and eukaryotic sequences from InterProScan domain annotations, assigns a MIBiG class and predicts NPClassifier pathways and superclasses of the products. This repository holds one model bundle: a 3-seed Conv1D + BiLSTM token detector, a RandomForest post-filter / MIBiG class head and an NPClassifier ensemble (RandomForest + logistic regression + kNN).

Use

sanntis setup --container-engine docker            # InterProScan 6 + InterPro data, once
sanntis predict genome.fna --container-engine docker --model <this repo>@<revision>
sanntis predict assembly.gbk --ips-tsv interproscan.tsv   # precomputed annotations, no containers

from sanntis import SanntisModel, Predictor
model = SanntisModel.from_pretrained('<this repo>', revision='<revision>')

Inputs: prokaryotic nucleotide FASTA (genes called with pyrodigal), GenBank with CDS translations or a protein FASTA with the GFF3 of its CDS (prokaryotes and eukaryotes), each optionally with a precomputed InterProScan 5.76+ / 6.x TSV. Outputs: GFF3 (one CLUSTER feature per region with bgc_class, p_non_bgc, npc_pathway, npc_superclass, npc_pathway_key_domains_unsupported, ...), per-CDS scores, a regions table and antiSMASH sideloader JSON.

Model

Field	Value
sanntis version	1.0.0b1
Fit	production
Features	InterProScan 6.0.3.0, InterPro 107.0; 14 member databases (pfam, ncbifam, sfld, superfamily, gene3d, hamap, coils, cdd, prints, pirsf, prositeprofiles, prositepatterns, mobidblite, smart)
Vocabulary	6422 tokens (incl. 3 specials)
Detector	3 seeds, 8 epochs (median best CV epoch)
Operating point t*	0.3 (hysteresis low 0.15); highest F2_macro on validation with BGC precision >= 0.80
RF post-filter	drop regions with P(non-BGC) >= 0.8
Heads	MIBiG class: True; NPClassifier: True

Training data

• Positives: 2795 external BGCs (MIBiG format), deduplicated (composite Dice >= 0.974, longest kept); loci > 200 kb and loci without CDS excluded.
• Negatives: genome pool v1 (GTDB r226 and RefSeq eukaryote representatives; native BGCs removed with antiSMASH 8.0.1): length-matched chunks and BGC-free background pieces.
• Entities: background euk 4233, background proka 30077, chunk euk 1608, chunk proka 6447, positive euk 536, positive proka 2149.
• Published as the dataset microbiome-informatics/sanntis-training-data (tags mibig-5.0rc1-train, negatives-v1, npclassifier-v1); see the "Training data" section of the SanntiS README for the data and the training command. The run is objective 01 of the sanntisv1-data-and-analysis repository; its out-of-fold validation report is in report/.

Performance

Metrics in ORF space. Validation = pooled out-of-fold CV predictions (training data).

Dataset	Precision_T	Recall_macro_T	F1_macro_T	BGC precision	BGC macro recall	F2_macro
Validation (CV, pooled OOF)	0.698	0.875	0.776	0.882	0.926	0.917

Full report with confidence intervals and figures: report/report.md.

Limitations

• Tokens come from InterProScan member-database hits; predictions on annotations from other InterProScan or InterPro releases drift (sanntis warns and reports coverage QC). Unannotated genes carry no signal.
• Eukaryotic negatives cover fungi and protists only (no plants); plant BGC estimates are anecdotal.
• BGC families with low similarity to MIBiG are recalled less well than near-duplicates (novelty analysis in the sanntisv1-data-and-analysis repository).
• p_non_bgc and the NPClassifier scores are model scores, not calibrated probabilities (the NPC abstention tiers target 0.9 / 0.8 precision on validation).
• npc_pathway_key_domains_unsupported is a hint, not a verdict: a called pathway none of whose key domains (or domain pairs) is in the region is wrong more often than a supported one, but most wrong calls do carry the domains (hybrids, tailoring); pathways with low rule coverage in CV are na.

Integrity

Every file is listed with its sha256 in manifest.json and verified on load. No pickles: detectors are safetensors, sklearn models skops (explicit trusted types), the kNN reference a sparse npz.

Citation

@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}
}