| name | plmsearch |
| description | PLMSearch — protein language model embeddings for fast, sensitive remote homology search from sequences alone. Use when: (1) Homology is too remote for BLAST/MMseqs, (2) Structure search is unavailable but structural similarity is expected, (3) Screening large query–target protein pairs. Upstream: https://github.com/maovshao/PLMSearch. Web: https://dmiip.sjtu.edu.cn/PLMSearch. Route via protein-language-model.
|
| license | MIT |
| category | analysis-tools |
| tags | ["PLM","remote-homology","search","protein"] |
| upstream | https://github.com/maovshao/PLMSearch |
| stage | function |
PLMSearch
Upstream: maovshao/PLMSearch ·
https://dmiip.sjtu.edu.cn/PLMSearch
Citation
Liu, W. et al. PLMSearch: Protein language model powers accurate and fast
sequence search for remote homology. Nat. Commun. 15, 2775 (2024).
https://doi.org/10.1038/s41467-024-46808-5
See also docs/references.md.
Analytical thinking
Classical sequence search misses many remote homologs (low identity, similar
structure). PLMSearch uses pretrained PLM representations plus a similarity
model trained with structural similarity signals — sequence-only input, structure-
aware sensitivity, MMseqs2-like speed on large pair screens in the paper’s
benchmarks.
| Tool | When |
|---|
mmseqs2 / diamond | Default catalogue search |
foldseek | Structures / ProstT5 available |
plmsearch | Remote homology / dark proteins (sequence-only PLM) |
deepblast | Structure-aware alignment + TM-score prediction (TM-Vec) |
Hits remain computational — transfer annotation cautiously.
How to run
Decision tree
Protein search sensitivity need?
├─ Standard → mmseqs2 / diamond
├─ Have structures → foldseek (± folddisco motifs)
├─ Remote homology, sequence-only → plmsearch
├─ Need explicit structural alignment from sequence → deepblast
└─ Gene neighborhood modules → spacedust
Related skills
protein-language-model · esm · prostt5 · metagenomics-llm ·
foldseek · folddisco · deepblast · mmseqs2 · diamond ·
eggnog-mapper · spacedust · tool-selection
