prostt5

name	prostt5
description	ProstT5 — bilingual protein language model translating amino-acid sequences and Foldseek 3Di structure tokens (fine-tuned from ProtT5). Use when: (1) Building Foldseek DBs from FASTA via --prostt5-model, (2) Sequence-only structure-aware search without AlphaFold, (3) AA↔3Di representation tasks. Upstream: https://github.com/mheinzinger/ProstT5. HF: Rostlab/ProstT5. Parent ProtT5: Rostlab/prot_t5_xl_uniref50. Route via protein-language-model; production search usually via foldseek.
license	MIT
category	analysis-tools
tags	["ProstT5","ProtT5","PLM","protein-language-model","3Di","Foldseek","Rostlab"]
upstream	https://github.com/mheinzinger/ProstT5
stage	function

ProstT5 (ProtT5 → AA ↔ 3Di)

Upstream: mheinzinger/ProstT5 · HF: https://huggingface.co/Rostlab/ProstT5 · Parent ProtT5: https://huggingface.co/Rostlab/prot_t5_xl_uniref50

Citation

Heinzinger, M. et al. Bilingual language model for protein sequence and structure. NAR Genom. Bioinform. 6, lqae150 (2024). https://doi.org/10.1093/nargab/lqae150

ProtT5 / ProtTrans: Elnaggar, A. et al. ProtTrans: Toward Understanding the Language of Life Through Self-Supervised Learning. IEEE Trans. Pattern Anal. Mach. Intell. 44, 7112–7127 (2022). https://doi.org/10.1109/TPAMI.2021.3095381

See also docs/references.md.

Analytical thinking

ProstT5 fine-tunes ProtT5 to translate between amino acids and Foldseek 3Di tokens. In metagenomics the main production path is Foldseek createdb/easy-search with --prostt5-model: ~400–4000× faster than predicting full structures before search, but ProstT5-only DBs lack Cα — no TM-score / LDDT features that need coordinates.

vs	Prefer ProstT5 when
esm	Need 3Di / Foldseek path, not general ESM embeddings
foldseek (AFDB coords)	Only FASTA available; accept 3Di-only DB limits
plmsearch	Structure-token search vs PLM embedding homology
ColabFold (deferred)	Prefer AFDB / ProstT5+Foldseek over re-folding catalogues

How to run

# preferred production path — Foldseek wraps ProstT5
foldseek databases ProstT5 weights tmp
foldseek createdb query.fasta queryDB --prostt5-model weights
# or:
foldseek easy-search query.fasta targetDB res.m8 tmp --prostt5-model weights
# GPU: --gpu 1 (Foldseek CUDA build); see foldseek skill

# standalone HF inference: follow mheinzinger/ProstT5 README

Decision tree

Need structure-aware protein search from FASTA?
├─ Yes, Foldseek production → foldseek + prostt5 weights
├─ AA embeddings only → esm
├─ PLM embedding homology → plmsearch / deepblast
├─ Have AFDB/PDB structures → foldseek (coords)
└─ Classical sequence → mmseqs2 / diamond

Related skills

protein-language-model · foldseek · esm · esmc · saprot · plmsearch · deepblast · unicore · metagenomics-llm · homology-search · tool-selection