erast

name	erast
description	ERAST — efficient retrieval-augmented homology search combining sequence language-model embeddings with a billion-scale vector database (protein and nucleotide). Use when: (1) Classical BLAST/MMseqs/Foldseek scale poorly, (2) You need vector-DB nearest-neighbor search plus reranking, (3) Accessing the public ERAST web DB or running the local encode→predict pipeline. Upstream: https://github.com/TencentAILabHealthcare/ERAST. Web DB: https://ai4s.tencent.com/erast. Route via homology-search.
license	MIT
category	analysis-tools
tags	["ERAST","vector-DB","embedding","homology","PLM","search"]
upstream	https://github.com/TencentAILabHealthcare/ERAST
stage	function

ERAST

Upstream: TencentAILabHealthcare/ERAST · Web vector DB: https://ai4s.tencent.com/erast

Citation

Jiang, Y. et al. Scalable homology detection with ERAST. Nat. Biotechnol. (2026). https://doi.org/10.1038/s41587-026-03051-1

See also docs/references.md.

Analytical thinking

ERAST is the vector-database route: embed sequences (ESM2-class PLMs), retrieve neighbors from a ~billion-sequence index, then preretrieval / retrieval / postretrieval (rerank) stages. Supports protein and nucleotide. Complements — does not replace — mmseqs2 / foldseek for routine catalogues.

vs	Prefer ERAST when
mmseqs2 / diamond	Extreme DB scale / embedding retrieval desired
foldseek	No structures; or nt+protein vector search
plmsearch / deepblast	Using the public billion-scale ERAST index / web DB

Caveats: pin model + Pfam DB + EHSM checkpoint; web DB ≠ your private catalogue unless you rebuild locally; embedding hits need careful annotation transfer.

How to run

Web: https://ai4s.tencent.com/erast

Local pipeline (see upstream pipeline/readme.md):

# Pfam + ESM2 embeddings + homology predict (protein mode)
export PFAMDB=/path/to/pfamdb
export MODEL_PATH=/path/to/esm2_or_erast_weights
cd pipeline
python pfam_scan.py -out example/pfam_q.json -outfmt json example/q.fa "$PFAMDB"
python pfam_scan.py -out example/pfam_t.json -outfmt json example/t.fa "$PFAMDB"
python encode.py --res_path example/q_emb.npy --model_path "$MODEL_PATH" --fa_path example/q.fa
python encode.py --res_path example/t_emb.npy --model_path "$MODEL_PATH" --fa_path example/t.fa
python predict.py --fa_q example/q.fa --fa_t example/t.fa \
  --emb_q example/q_emb.npy --emb_t example/t_emb.npy \
  --pfam_q example/pfam_q.json --pfam_t example/pfam_t.json \
  --out example/out.jsonl --mode p
# nucleotide: encode/predict with --mode n (see upstream)

EHSM checkpoint: Zenodo record noted in upstream README.

Decision tree

Need homology at extreme scale / vector retrieval?
├─ Public billion-seq search → erast web DB
├─ Custom local sets → erast pipeline (encode → predict)
├─ Structure available → foldseek
├─ Routine catalogues → mmseqs2 / diamond
└─ Unsure → homology-search

Related skills

homology-search · plmsearch · deepblast · mmseqs2 · foldseek · metagenomics-llm · tool-selection