bio.rodeo
ModelsOrganizationsProvidersLeaderboardAboutSign in
bio.rodeo

The authoritative source for evaluating biological foundation models. No hype, just honest analysis.

Categories
  • DNA & Gene
  • RNA
  • Protein
  • Small molecule
  • Single-cell
  • Spatial omics
  • Pathology
  • Imaging
  • Metabolomics
  • Biosignals
  • Language model
bio.rodeoModelsOrganizationsProvidersLeaderboardAboutFAQSubmit a modelContact
© 2026 Pulsatance. All rights reserved. ~
Built by Pulsatance
models / dna-gene / deepsems
DNA & GeneSmall molecule
Tongji UniversityFudan UniversityShanghai Institute of Nutrition and Health, Chinese Academy of SciencesUniversity of Chinese Academy of SciencesChinese Academy of SciencesReleased March 2025

DeepSeMS

Secondary metabolite structure prediction from microbial biosynthetic gene clusters, generating SMILES strings from Pfam functional-domain tokens.

The short version

  • —Turns a cryptic gene cluster into candidate SMILES a chemist can dock or synthesize
  • —Reads a cluster as Pfam functional-domain tokens rather than raw amino acids
  • —Feature-aligned SMILES enumeration expands a small curated set of cluster-structure pairs
  • —Ten cross-validation checkpoints vote to return ten candidate structures per cluster
82Openness2Citations
13GitHub stars
Apache-2.0License

Where to run it

No providers recorded yet. Browse all providers

DeepSeMS is a sequence-to-sequence transformer that reads a microbial biosynthetic gene cluster (BGC) and writes the SMILES string of the secondary metabolite that cluster is expected to produce. It was developed at Tongji University with collaborators at Fudan University and the Shanghai Institute of Nutrition and Health, released as a preprint in March 2025 and published in Nature Computational Science in 2026.

Genome mining has become very good at finding BGCs and very poor at saying what they make. Tools such as antiSMASH and PRISM locate clusters and then apply curated biosynthetic rules to guess a product, which works when a cluster resembles a characterized pathway and breaks down on the cryptic clusters that dominate metagenome-assembled genomes. DeepSeMS replaces the rule base with a learned translation: the gene cluster is the source language, the chemical structure the target.

The design choice that makes this tractable is the representation. Rather than feeding raw amino acid sequence — up to 50,000 tokens per cluster, beyond what the model could attend over — or whole enzymes, which collapse to as few as 50 tokens and lose the substrate detail, DeepSeMS tokenizes each cluster as the ordered Pfam functional domains its biosynthetic enzymes encode. That is the same domain-as-token idea used by BGC-Transformer, pointed at structure generation rather than cluster modeling, and it pairs naturally with detectors such as BGC-Finder.

#Key Features

  • Genome to chemistry in one step: Takes an antiSMASH or DeepBGC annotation file and emits candidate SMILES, with no requirement that the cluster resemble anything in the training set.
  • Functional-domain tokenization: A source vocabulary of 1,020 distinct Pfam identifiers represents biosynthetic enzymes at the modular level that determines substrate selection and chemistry, keeping clusters under about 250 tokens.
  • Feature-aligned data augmentation: SMILES enumeration constrained to preserve structural feature blocks expands a small curated training set while emphasizing scaffold structure, outperforming both no augmentation and plain randomized enumeration.
  • Ensemble of cross-validation checkpoints: Ten checkpoints, one per fold, are run together to return the top ten candidate structures for each input cluster.
  • Hosted, login-free web server: Two mirrored public servers run inference in a browser and ship with a searchable repository of the ocean-microbiome structures the model generated.

#Technical Details

The network is a standard encoder-decoder transformer with six encoder and six decoder layers, eight attention heads, and an embedding dimension of 512, for approximately 100 million trainable parameters. Source tokens are Pfam identifiers assigned by HMMER search against Pfam 36.0 at an e-value threshold of 0.01; target tokens are 35 distinct structural features extracted from SMILES notation. Training pairs come from MIBiG 3.1, an initial 3,029 cluster-structure pairs expanded to 55,903 by feature-aligned enumeration and trained under ten-fold cross-validation.

On 326 gold-standard BGCs curated by the PRISM 4 authors, with anything above 95% sequence similarity to training data removed, DeepSeMS produced at least one chemically valid structure for 97.55% of clusters, against 88.96% for PRISM 4 and 63.50% for antiSMASH 7. Mean structural similarity to the true metabolite was 0.60, versus 0.45 and 0.14; scaffold similarity was 0.63; 41.10% of predictions were chemically identical to the ground truth and 53.68% shared its exact scaffold. On 940 cryptic clusters from the Malaspina deep-ocean metagenome-assembled genomes, DeepSeMS returned valid structures for 96.60%, against 46.45% for PRISM 4 and 16.91% for antiSMASH 7, producing 5,104 valid structures of which 78.66% were chemically unique.

#Applications

Natural product chemists screening environmental genomes gain a way to prioritize which of thousands of uncharacterized clusters to pursue: a predicted scaffold can be searched against known compound libraries, scored for drug-likeness, and used to plan isolation or heterologous expression. The authors demonstrated this on 27,139 metagenome-assembled genomes from the Ocean Microbiomics Database, in which 46,786 BGCs yielded 65,868 unique metabolite structures, of which 97% scored as novel against known chemistry and 69% carried novel scaffolds, with predicted antibiotic activity among them exposed through the web server's built-in repository.

#Impact

DeepSeMS shows that structure prediction from biosynthetic sequence can be posed as machine translation and learned from a few thousand curated pairs, provided the input representation is chosen at the right granularity and the target language is augmented to teach chemical syntax. Its accuracy on cryptic clusters, where rule-based tools succeed less than half the time, is the result that matters for uncultivated microbial chemistry. The predictions remain computational hypotheses: the paper reports no experimental isolation of the ocean metabolites. Code is released under Apache 2.0, with the ten fold checkpoints and preprocessed data distributed through Figshare — the repository's checkpoints/ path holds a placeholder file, not a model — and a Docker image for reproducible inference. The journal article is subscription-only, so the CC BY-NC preprint is the freely readable account of the method.

At a glance

Parameters
100 Million
Released
March 2025
Category
DNA & Gene
License
Apache-2.0
Organizations
Tongji University / Fudan University / Shanghai Institute of Nutrition and Health, Chinese Academy of Sciences / University of Chinese Academy of Sciences / Chinese Academy of Sciences

Related models

  • BGC-Transformer

    Kyoto Institute of Technology / University of Tokyo

  • BGC-Finder

    Huazhong University of Science and Technology

  • CHAMOIS

    Leiden University Medical Center / European Molecular Biology Laboratory

  • BCCoE

    Agency for Science, Technology and Research (A*STAR)

  • f-BGM

    Peking University

Links

GitHub RepositoryResearch PaperbioRxiv PreprintDocumentationDemoDataset

Tags

generativelanguage_modelmolecular_generationnatural_product_discoverytransformer

Something wrong?

Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.