| name | genome-language-model |
| description | Route genome language models (gLMs) for metagenomics and microbial genomics: representation models (Nucleotide Transformer / NTv3, DNABERT-2, Caduceus), generative models (Evo / Evo 2), and mixed-modality metagenomic LMs (OMG + gLM2). Use when choosing DNA embeddings, generation, or functional-track prediction — not for classical profilers or Foldseek. Child skills: omg, nucleotide-transformer, dnabert2, caduceus, evo2. Protein LMs via protein-language-model. Trees via phylogenomics / lucaphylo.
|
| license | MIT |
| category | orchestration |
| tags | ["genome-language-model","gLM","NTv3","DNABERT","Caduceus","Evo2","OMG","gLM2"] |
Genome language model
DNA/genome foundation models are a parallel track to QC→assembly→binning.
They provide embeddings, likelihoods, generation, or track prediction — they do
not replace CheckM2, GTDB-Tk, MetaPhlAn, or Foldseek.
NTv3-style taxonomy of DNA modeling:
| Strategy | Role | Prefer here |
|---|
| Representation gLM | Embeddings / MLM | nucleotide-transformer · dnabert2 · caduceus |
| Generative (AR / diffusion) | Sequence design / long context | evo2 (± Evo 1) |
| Mixed-modality metagenomic LM | CDS AA + intergenic DNA | omg (gLM2) |
| Protein language models | Remote AA homology / embeddings | protein-language-model |
| seq2fun tracks (Borzoi-class) | Dense functional genomics | Not default MAG tools |
Analytical thinking
| Claim | Prefer | Do not treat as |
|---|
| Metagenomic corpus + gLM2 | omg | Cohort taxonomy tables |
| Multi-species DNA / NTv3 tracks | nucleotide-transformer | Experimental annotation |
| Efficient multi-species DNA LM + GUE | dnabert2 | Taxonomy profiler |
| Long-range bi-directional DNA (Mamba) | caduceus | Contig assembler |
| Generative / 1 Mb-scale DNA design | evo2 | MAG QC |
| Protein remote homology | protein-language-model | Genome LM logits as hits |
| Phylogeny | phylogenomics / lucaphylo | Raw LM distances as trees |
Hard rules:
- Pin model name, size, checkpoint, context length, and license (NT often NC)
- OMG / OpenGenome2 ≠ ENA study reanalysis (
sra-ena / mgnify)
- Do not merge gLM scores with BLAST/MMseqs/Foldseek tables unlabeled
- Prefer multi-species / microbial-aware checkpoints for MAG work
- GPU needs for large gLMs differ from KneadData/MEGAHIT nodes
Decision tree
Genome language model need?
│
├─ Metagenome mixed AA+DNA / gLM2 → omg
├─ NTv3 / multi-species tracks & NT family → nucleotide-transformer
├─ Efficient multi-species representation / GUE → dnabert2
├─ Long-range RC-aware DNA LM (Caduceus) → caduceus
├─ Generative long-context (all domains) → evo2
├─ Protein language models → protein-language-model
├─ Structure search → foldseek
├─ Trees → phylogenomics
└─ Unsure → this hub + metagenomics-llm
Related skills
omg · nucleotide-transformer · dnabert2 · caduceus · evo2 ·
protein-language-model · metagenomics-llm · lucaphylo ·
homology-search · phylogenomics · public-databases · tool-selection
