| name | metagenomics-llm |
| description | Route genomic/protein language models and Steinegger-lab structure/sequence search for shotgun metagenomics. Use when: (1) Choosing gLMs vs PLMs vs classical homology, (2) MMseqs2 vs Foldseek vs PLM search, (3) Spacedust / AlphaGEM, (4) Distinguishing foundation-model corpora from SRA/MGnify reanalysis. For DNA gLMs use genome-language-model; for protein LMs use protein-language-model (esm, esmc, saprot, prostt5, plmsearch, deepblast). Classical pipelines: metagenomics-workflow / tool-selection.
|
| license | MIT |
| category | orchestration |
| tags | ["LLM","PLM","gLM","Foldseek","MMseqs2","Folddisco","structure-search","AI"] |
Metagenomics × LMs & structure search
Foundation models and structure-aware search (Steinegger / Söding stack)
are a parallel track to QC→assembly→binning. They help when homology is
remote, functions are dark, or you need embeddings — they do not replace
CheckM2, GTDB-Tk, or abundance profilers.
For genome language models prefer genome-language-model. For protein
language models prefer protein-language-model.
Analytical thinking
Steinegger stack (sequence → fold → multimer → motif)
| Conservation level | Prefer | Typical claim |
|---|
| Amino-acid sequence | mmseqs2 (± diamond) | Catalogue search, NR, contig taxonomy |
| Single-chain fold | foldseek | AFDB/PDB structural homologs; ProstT5 from FASTA |
| Multimer / interface | foldseek easy-multimersearch | Complex TM / chain pairing |
| Local structural motif | folddisco | Catalytic / metal-site geometry |
| Gene neighborhood modules | spacedust | Conserved clusters (Foldseek-aware) |
Language models & dark metabolism
| Claim type | Prefer | Do not treat as |
|---|
| Genome language models (route all gLMs) | genome-language-model | Taxonomy or abundance tables |
| Protein language models (route all PLMs) | protein-language-model | Taxonomy or abundance tables |
| Metagenomic corpus + gLM2 | omg | ENA study reanalysis |
| ESM-2 / ESM C / SaProt / ProstT5 | esm / esmc / saprot / prostt5 | Experimental structure proof |
| Remote protein homology (sequence-only PLM) | plmsearch, deepblast | Experimental structure proof |
| GEM + structure/dark metabolism | alphagem (± gapseq/vmh) | Flux measurements |
| Classical KO/COG annotation | eggnog-mapper, dram, kegg | Obsolete — still primary for many papers |
Hard rules:
- Embeddings / LM / structure hits are hypotheses — pin model + DB + software
- OMG/Hugging Face corpora ≠ ENA study reanalysis (
sra-ena / mgnify)
- Do not merge PLM / Foldseek rankings with BLAST/MMseqs tables unlabeled
- GPU/memory needs differ wildly from KneadData/MEGAHIT nodes
- Low pLDDT structures and over-long Folddisco motifs are common failure modes
Decision tree
AI / homology need in a metagenomics project?
│
├─ Genome language models (NT, DNABERT-2, Caduceus, Evo 2, gLM2, …)
│ → genome-language-model
├─ Protein language models (ESM-2, ESMC, SaProt, ProstT5, PLMSearch, …)
│ → protein-language-model
├─ Sequence catalogue / contig taxonomy
│ → mmseqs2
├─ Structure search (AFDB/PDB) or ProstT5 from FASTA
│ → foldseek (± prostt5)
├─ Multi-chain complexes / interfaces
│ → foldseek easy-multimersearch
├─ Short discontinuous motifs
│ → folddisco
├─ Conserved gene clusters / functional modules
│ → spacedust
├─ Improve GEMs with structure / dark metabolism mining
│ → alphagem (± gapseq, vmh)
├─ Classical function at catalogue scale
│ → eggnog-mapper / dram / humann
├─ Full homology-search menu (vector DB, Phylign, …)
│ → homology-search
├─ Trees / phylogenomics
│ → phylogenomics
└─ Unsure
→ this skill + tool-selection
Where it sits in the pipeline
contigs / MAG proteins (after QC)
├─ classical: eggnog-mapper · diamond · mmseqs2
├─ structure: foldseek · folddisco · spacedust
├─ vector DB / genome warehouse search → homology-search
└─ protein language models
└─ protein-language-model → esm · esmc · saprot · prostt5 · plmsearch · deepblast
genome language models
└─ genome-language-model → omg · nucleotide-transformer · dnabert2 · caduceus · evo2
metabolic models
└─ gapseq · alphagem · vmh
trees / evolution
└─ phylogenomics
Methods caveats
- Report software version, database release, and (for LMs) checkpoint URL
- For dark-matter genes, state that LM/structure annotations lack wet-lab support
- Prefer holding out known systems when claiming de novo discovery
(
spacedust vs defensefinder / PADLOC-style tools)
- Foldseek-Multimer metagenomic CRISPR–Cas examples illustrate sensitivity at
<20% sequence identity — still validate priority systems
Related skills
genome-language-model · protein-language-model · omg ·
nucleotide-transformer · dnabert2 · caduceus · evo2 · esm ·
esmc · saprot · prostt5 · homology-search · phylogenomics ·
mmseqs2 · foldseek · folddisco · erast · plmsearch · deepblast ·
spacedust · alphagem · eggnog-mapper · public-databases ·
tool-selection · metagenomics-workflow
