deepblast

name	deepblast
description	TM-Vec and DeepBLAST — deep learning for protein remote homology and structural alignment from sequences (predict TM-scores; align homologous regions without solved structures). Use when: (1) Searching structure-like similarity in large sequence DBs, (2) Aligning remote homologs for annotation, (3) Complementing PLMSearch / Foldseek workflows. Upstream: https://github.com/tymor22/tm-vec and https://github.com/flatironinstitute/deepblast. Route via protein-language-model.
license	MIT
category	analysis-tools
tags	["DeepBLAST","TM-Vec","remote-homology","structure","PLM"]
upstream	https://github.com/flatironinstitute/deepblast
stage	function

TM-Vec / DeepBLAST

Upstream: tm-vec · deepblast

Citation

Hamamsy, T. et al. Protein remote homology detection and structural alignment using deep learning. Nat. Biotechnol. 42, 975–985 (2024). https://doi.org/10.1038/s41587-023-01917-2

See also docs/references.md.

Analytical thinking

Component	Role
TM-Vec	Predict structural similarity (TM-score-like) from sequence pairs; search large DBs
DeepBLAST	Structurally align proteins from sequence by identifying homologous regions

Designed for the metagenomic reality that many proteins lack close sequence neighbors but retain structural homology. Complementary to plmsearch and to structure search (Foldseek) when structures/pLDDT models exist.

How to run

# install per wiki: https://github.com/flatironinstitute/deepblast/wiki/Installation
# typical: TM-Vec search → DeepBLAST align top hits

Decision tree

Remote homology workflow?
├─ Fast PLM embedding search → plmsearch
├─ TM-score search + structural alignment → deepblast (TM-Vec)
├─ Have structures / AFDB → foldseek (± folddisco / spacedust)
└─ Close homologs → mmseqs2 / diamond

Related skills

protein-language-model · esm · plmsearch · metagenomics-llm · foldseek · folddisco · spacedust · mmseqs2 · eggnog-mapper · tool-selection