AI for Bio · Model

SaProt

Structure-aware protein language model pairing structure tokens with amino acids

ProteinsAcademic labChinaMaintained only中文

Overview

NameSaProt · aliases: SaProt, SaProt 650M, SaProt 35M, SaProt 1.3B
Full nameShort name only; neither paper nor repository gives an expanded name
Ambiguous names (not auto-linked)None
FamilyNot a family entry. 35M and 650M are two sizes of one model, not family members
OrganizationWestlake University, Westlake Representation Learning Lab
Organization typeAcademic lab
CountryChina
Model classProtein language model (structure-aware)
DomainProteins (also: Structure & co-folding)
First release† partly verified · preprint
Basis: First public bioRxiv preprint
Note: The preprint DOI is not Crossref-confirmed; the date comes from the same-day weights repository and README notes (unverified) source
Other milestones
  • First public evidence: † · Preprint posted source
  • Preprint v1: † · Preprint v1 source
  • Journal online: · SaProt/SaprotHub published in Nature Biotechnology source
  • Weights release: † · 1.3B weights noted in the README news source
Latest update · journal
Basis: Journal publication in Nature Biotechnology
SaProt/SaprotHub published in a journal; 1.3B weights earlier on 2024-12-09, README touched in 2026-03 source
The most recent sourced event, not a claim that nothing changed since.
StatusMaintained only · Journal publication and repository upkeep, no new generation
OpennessCode: Open
Weights: Open
Code and weights public on GitHub and Hugging Face; licence text not retrieved this round, so terms are left empty
Benchmarks reported3 counted (board 2, self 1)
Wet-lab validationNone recorded
Unverified fields† first_release, benchmarks_reported[2], release_milestones.weights_release, open_status.weights_license, open_status.code_license
As of2026-11-16

What it does

Combines each residue's amino acid with a local structure token into a structure-aware vocabulary, pretrained on a public predicted-structure database, for zero-shot mutation-effect scoring and fine-tuning, with no extra MSA. It is the simpler predecessor of structure-enhanced zero-shot models such as ProSST and VenusREM.

Benchmarks reported

Every row is one pairing from the survey's relation file; a score belongs to the benchmark page's ranking, where it is shown with metric, eval set, source type and date. 3 of 3 pairings are counted.

BenchmarkRelationEval set / protocolSource typeNote and evidence
02 ProteinGymOfficial leaderboard or resultszero-shot substitution DMS summaryBoardEntry on the official summary table · evidence
34 AbBiBenchOfficial leaderboard or resultsantibody binding-affinity correlation, structure-aware LM categoryBoardBenchmark-registered score †unverified · no source link
26 FLIP / FLIP2Developer report, official protocolMeltome human-cell (called Thermostability in the paper); AAV 2-vs-restSelf-reportedSelf-reported on two datasets; a third-party re-evaluation also exists · evidence

Other evaluations (outside the 44)

  • Third-party FLIP GB1 extrapolation re-evaluation (third party re evaluation) Third-party: A third-party paper re-ran the low-vs-high split and found it no better than a single-sequence baseline (figures not quoted here)
  • Time-dependence of the board position (Public benchmark not among the 44) Third-party: A maintainer blog post discusses it as an example of the structure-token route when weighing scale against gains

Developers report: Developers report that adding structure tokens to the vocabulary lets a mid-sized model beat same-size sequence-only models on zero-shot mutation effects without extra MSA.

Independent evidence: A third-party paper independently re-ran it on a FLIP extrapolation split and did not reproduce the advantage; no wet lab, no independent replication.

Training data

Group in Figure 7: Public data only.

ClassValueDetail
a. Public databasesYesAbout 40M public predicted structures and sequences, encoded into structure tokens with a structure-alignment tool source
b. Private or licensed dataNoNo private or licensed data reported
c. Own wet-lab data used for trainingNo
d. Wet lab used only for validationNo

Wet-lab validation

None. The developers report no wet-lab validation.

Background

  • Organization: An academic lab with no separate company: a representation-learning lab at Westlake University in Hangzhou working on AI for life science and on deep learning for retrieval and recommendation. The model appeared at a 2024 machine-learning conference and in a journal in 2025.
  • Funding: Paper acknowledgements not retrieved; funding left blank
  • Products and partnerships: A companion hosting and fine-tuning platform (SaprotHub) shipped with the journal version; the structure-token approach is treated as a peer baseline by ProSST and VenusREM, and third-party inference platforms offer it online.
  • Related models: ProSST, VenusREM, ESM3, ESM C

Reading notes

  1. The "once first" claim is a 2024 snapshot; board position shifts with versions and newcomers and is not a current statement.
  2. Structure inputs come from public predicted structures, so regions without structure or with low confidence are weak spots.
  3. A third-party re-evaluation on an extrapolation split shows structure tokens do not guarantee better extrapolation.
  4. Computational evaluation only, no wet lab: it speaks to ranking, not to designability.

Caveats

  1. The AbBiBench pairing rests only on this site's page (unverified).
  2. First-release evidence is thin, taken from the same-day weights repository and README (unverified).
  3. No wet-lab validation, and no independent replication of its board position.

Sources

  1. SaProt conference paper (2024) (paper)
  2. Official repository (repository)
  3. SaProt/SaprotHub journal version (journal)
  4. Weights repository (repository)
  5. ProteinGym official repository and board (leaderboard)
  6. Third-party FLIP re-evaluation paper (third-party paper)
  7. Board maintainer blog post (second-hand) (second-hand) second-hand

Corrections

None recorded.

Part of the AI for Bio benchmarks survey · all 112 models · open in the explorer. Record as of 2026-11-16.