saprot

name	saprot
description	SaProt — Westlake University structure-aware protein language model using an AA+3Di (Foldseek) structure-aware vocabulary (ICLR 2024). Use when: (1) Structure-informed PLM embeddings for function / fitness / variant tasks, (2) Comparing against ESM-2 with Foldseek SA tokens available, (3) SaProt 1.3B AA-only or SA-token workflows (incl. AFDB/OMG/NCBI-trained checkpoints). Upstream: https://github.com/westlake-repl/SaProt. HF: westlake-repl/SaProt_*. Route via protein-language-model. Related Westlake tools (ProTrek, etc.) — cite separately if used.
license	MIT
category	analysis-tools
tags	["SaProt","Westlake","PLM","protein-language-model","Foldseek","3Di","ICLR"]
upstream	https://github.com/westlake-repl/SaProt
stage	function

SaProt (Westlake)

Upstream: westlake-repl/SaProt · HF: https://huggingface.co/westlake-repl · OpenReview: https://openreview.net/forum?id=6MRm3G4NiU

Citation

Su, J. et al. SaProt: Protein Language Modeling with Structure-aware Vocabulary. International Conference on Learning Representations (ICLR) (2024). https://openreview.net/forum?id=6MRm3G4NiU

Preprint: https://doi.org/10.1101/2023.10.01.560349

See also docs/references.md.

Analytical thinking

SaProt builds a structure-aware vocabulary: residue tokens combined with Foldseek 3Di structure tokens (SA tokens), trained on large AF2/PDB-scale structure–sequence corpora (Westlake REPL lab). Strong on many structure/function downstream tasks when SA input is available.

Critical upstream caveats:

• 35M / 650M: optimal with SA-token input; frozen AA-only embeddings are weak — fine-tune if forced to AA-only
• 1.3B (e.g. SaProt_1.3B_AFDB_OMG_NCBI): works better with both SA and AA-only modes; metagenome-relevant OMG/NCBI mixing noted in checkpoint names
• SA tokens need structures or Foldseek 3Di — for FASTA→3Di production search prefer foldseek + prostt5; SaProt is the PLM, not a search engine

vs	Prefer SaProt when
esm / esmc	Have structures / SA tokens and want structure-aware PLM
prostt5	Downstream PLM tasks vs building Foldseek 3Di DBs
plmsearch	Embedding/finetune model vs remote-homology search wrapper

Related Westlake releases (ProTrek, Pinal, Evolla, …) — mention only if the project uses them; keep this skill focused on SaProt.

How to run

# clone / pip per upstream README; HF checkpoints e.g.
#   westlake-repl/SaProt_650M_AF2
#   westlake-repl/SaProt_1.3B_AFDB_OMG_NCBI
# convert structures → SA tokens with Foldseek (see SaProt README)
# then HF or ESM-style forward for embeddings / fine-tunes

Decision tree

Structure-aware protein LM?
├─ AA+3Di SaProt (Westlake) → saprot
├─ Sequence-only modern embeddings → esmc (± esm)
├─ FASTA → Foldseek search → prostt5 + foldseek
└─ Remote homology ranking → plmsearch / deepblast

Related skills

protein-language-model · esmc · esm · prostt5 · foldseek · plmsearch · omg · metagenomics-llm · tool-selection