CHAMOIS

Finding biosynthetic gene clusters in genomes is now routine; knowing what they make is not. Public repositories hold millions of predicted clusters, while MIBiG, the reference collection of experimentally validated ones, numbers only a few thousand active entries. For the overwhelming majority the encoded metabolite is unknown, and confident guesses are possible only where strong homology to a characterized cluster exists — an assumption that breaks easily, since deleting a small accessory gene can change the product's chemistry and bioactivity.

CHAMOIS (Chemical Hierarchy Approximation for secondary Metabolism clusters Obtained In Silico), from Martin Larralde and Georg Zeller at Leiden University Medical Center and EMBL Heidelberg, sidesteps the intractable problem of generating a structure. Instead it predicts properties of the product: which classes of the ChemOnt chemical ontology the metabolite belongs to, given only the protein domains encoded in the cluster. That reframing turns metabolite inference into a tractable multi-label classification problem and, crucially, one that applies to every cluster type rather than to polyketides and non-ribosomal peptides alone, where most prior sequence-to-chemistry methods are confined.

#Key Features

• Ontology output instead of a structure: Predicting membership in interpretable, hierarchical chemical classes gives a usable chemical fingerprint without requiring the model to draw a molecule it has never seen.
• Type-agnostic coverage: Because the features are generic protein domains rather than assembly-line signatures, the same classifiers apply across all curated cluster types.
• Interpretable by construction: Sparse regression assigns most classes a handful of decisive domains, letting a user read off which enzymatic function the model thinks produces a given chemical feature.
• Hypotheses for unannotated domains: Screening high-weight domains without functional annotation yielded 106 uncharacterized domains with inferred biosynthetic roles across 292 domain-class pairs.
• Cluster-to-metabolite search: Given a metabolite of interest, the tool scores every predicted cluster in a genome by fingerprint similarity and ranks the likely producer.
• Pipeline-ready packaging: A command-line interface consumes GenBank records straight from cluster finders such as antiSMASH and GECCO.

#Technical Details

CHAMOIS is deliberately classical. Open reading frames in a cluster region are called with Pyrodigal and annotated against Pfam 38.0 with PyHMMER, producing a binary presence-or-absence vector of protein domains for each cluster — the same domain feature space that BGC-Transformer tokenizes. That vector feeds 539 independent LASSO logistic regression classifiers — scikit-learn's L1-regularized LIBLINEAR solver — one per retained ChemOnt class; domains receiving zero weight from every classifier are discarded, leaving 896 informative features in the shipped model. Labels came from ClassyFire annotations of compounds attached to 1,598 bacterial clusters in MIBiG 3.1, hitting 1,034 of 4,825 ChemOnt categories; the 539 occurring in at least five compound groups were kept.

Evaluation resists the homology inflation that plagues this field: clusters were grouped by MHFP6 fingerprint similarity of their products before five-fold stratified group cross-validation, so chemically similar metabolites never straddle a fold boundary. Median AUROC across classes was 0.79, and 120 classes were predicted at AUPRC of 0.5 or higher. Performance is not merely a function of class frequency — Phenazines reached AUPRC 1.0 from 13 positive examples, alongside Organonitrogens at 0.979. Substituting random forests for the LASSO models changed overall performance little. On 1,279 gold-standard clusters, CHAMOIS beat PRISM 4.0 at structure prediction for polyketide and non-ribosomal peptide clusters, while PRISM held a slight edge on RiPPs. In the producer-identification benchmark, 70 experimentally annotated clusters were sought among 2,331 candidate regions across complete genomes; CHAMOIS ranked the true cluster first in 31 cases and within the top five in 19 more.

#Applications

Natural product chemists can screen predicted clusters for those whose products likely carry a chemical feature of interest — a halogen, a thiazole, a hydroxamic acid — and prioritize strains for isolation. The reverse query is equally useful: when a metabolite has been characterized from an isolate but its genetic origin is unknown, CHAMOIS narrows dozens of candidate clusters to a short list. Genome miners can also group clusters by predicted chemistry rather than by sequence, measuring biosynthetic diversity in terms of products rather than genes.

#Impact

The contribution is a general, open, interpretable bridge from cluster sequence to product chemistry across cluster types, in a domain previously served mainly by rule-based systems and type-specific predictors. Its interpretability is not incidental: recovering known domain-chemistry links such as tryptophan halogenase with aryl halides, without being told any biosynthetic rules, is what makes its novel domain-function proposals worth pursuing. The authors are candid about the limits. RiPP products remain poorly predicted, since their structure depends largely on a precursor peptide that Pfam annotation does not capture. LASSO's sparsity is also double-edged: where several domains co-occur, as in enzyme complexes, the model tends to credit one and ignore the rest. Being cheap to retrain, however, CHAMOIS is positioned to improve as MIBiG grows.