| name | protein-language-model |
| description | Route protein language models (PLMs / pLMs) for metagenomics: representation models (ESM-2, ESM C / ESMC), structure-aware SaProt (Westlake), bilingual AA↔3Di (ProstT5 / ProtT5), remote-homology search (PLMSearch, TM-Vec/DeepBLAST), and PLM phylogeny (LucaPhylo). Use when choosing AA embeddings, sequence→3Di for Foldseek, or PLM search — not for classical eggNOG/KEGG tables or DNA gLMs. Child skills: esm, esmc, saprot, prostt5, plmsearch, deepblast. DNA gLMs via genome-language-model. Structure search via foldseek / homology-search.
|
| license | MIT |
| category | orchestration |
| tags | ["protein-language-model","PLM","pLM","ESM","ESMC","SaProt","ProstT5","remote-homology"] |
Protein language model
Protein foundation models are a parallel track to QC→assembly→binning and
to DNA genome-language-model. They provide AA embeddings, likelihoods,
sequence→3Di translation, or remote-homology rankings — they do not replace
eggNOG-mapper, DIAMOND/MMseqs2 catalogues, or Foldseek when structures exist.
PLM taxonomy for metagenomic proteins:
| Strategy | Role | Prefer here |
|---|
| Representation PLM (classic) | Embeddings / MLM / property heads | esm (ESM-2) |
| Representation PLM (next-gen) | Stronger sequence embeddings | esmc (ESM C / Cambrian) |
| Structure-aware PLM (AA+3Di) | SA-token embeddings (Westlake) | saprot |
| Bilingual AA ↔ 3Di | Sequence→structure tokens for Foldseek | prostt5 |
| Remote homology (sequence-only) | Dark / remote proteins | plmsearch · deepblast |
| Structure search (coords / AFDB) | Fold-level hits | foldseek (± folddisco) |
| Alignment-free PLM phylogeny | Divergent protein trees | lucaphylo |
| DNA / mixed-modality gLM | Contigs / genomes | genome-language-model |
Analytical thinking
| Claim | Prefer | Do not treat as |
|---|
| Pinned ESM-2 / Atlas-era embeddings | esm | Taxonomy profiler |
| Modern ESM C representation checkpoints | esmc | Drop-in silent swap for old ESM-2 Methods |
| Structure-aware AA+3Di PLM (Westlake) | saprot | Foldseek search engine |
| FASTA → Foldseek 3Di without AF | prostt5 / foldseek --prostt5-model | Experimental structure |
| Remote homologs, sequence-only PLM | plmsearch | Wet-lab function proof |
| TM-score-like search + structural align | deepblast | Solved PDB coordinates |
| Have AFDB/PDB structures | foldseek | PLM embedding distance alone |
| Hyperbolic PLM placement | lucaphylo | Default MAG taxonomy (gtdbtk) |
| Classical KO/COG at catalogue scale | eggnog-mapper / diamond | Obsolete — still primary |
Hard rules:
- Pin model name, size, checkpoint URL, and license (ESMC tiers differ)
- Do not merge PLM / Foldseek / BLAST–MMseqs tables unlabeled
- ProstT5 3Di DBs lack Cα — no TM-score / LDDT features that need coordinates
- SaProt 35M/650M want SA tokens; do not treat frozen AA-only as equivalent
- GPU needs for large PLMs differ from KneadData/MEGAHIT nodes
Decision tree
Protein language model need?
│
├─ AA embeddings / pinned ESM-2 → esm
├─ Modern ESM C / Cambrian embeddings → esmc
├─ Structure-aware AA+3Di PLM (Westlake SaProt) → saprot
├─ Sequence → 3Di for Foldseek → prostt5 (± foldseek)
├─ Remote homology, sequence-only → plmsearch
├─ TM-score search + structural alignment → deepblast
├─ Have structures / AFDB → foldseek (± folddisco)
├─ PLM phylogeny (no reliable MSA) → lucaphylo
├─ Classical function tables → eggnog-mapper / dram / humann
├─ DNA genome LMs → genome-language-model
└─ Unsure → this hub + metagenomics-llm / homology-search
Related skills
esm · esmc · saprot · prostt5 · plmsearch · deepblast ·
lucaphylo · foldseek · genome-language-model · metagenomics-llm ·
homology-search · eggnog-mapper · public-databases · tool-selection
