AI for Bio · Model

ProSST

Protein language model quantizing local structure into tokens with disentangled attention

ProteinsAcademic labChinaMaintained only中文

Overview

NameProSST · aliases: ProSST, ProSST-2048
Full namePaper and repository use the short name only
Ambiguous names (not auto-linked)None
FamilyNot a family entry. Different structure-vocabulary sizes (K) are configurations of one model, not members
OrganizationShanghai Jiao Tong University, Shanghai AI Laboratory
Organization typeAcademic lab
CountryChina
Model classProtein language model (quantized structure tokens)
DomainProteins (also: Structure & co-folding)
First release† partly verified · preprint
Basis: bioRxiv preprint v1 (the conference version came later)
Note: The preprint DOI is inferred from a title match (unverified); the site previously used the conference publication as first release source
Other milestones
  • First public evidence: † · Preprint v1 posted source
  • Preprint v1: † · Preprint v1 source
  • Official announcement: † · Formal publication at a 2024 machine-learning conference (exact day not pinned) source
Latest update† partly verified · conference paper
Basis: Publication at a 2024 machine-learning conference
Formal conference publication; the 2026-07 and 2026-09 README items are project news, not model updates
Note: The exact publication day is not pinned, so it is kept to the month (unverified) source
The most recent sourced event, not a claim that nothing changed since.
StatusMaintained only · Repository maintained with project news, no new ProSST version
OpennessCode: Open
Weights: Open
Code and weights public in the official repository; licence text not retrieved this round, so terms are left empty
Benchmarks reported2 counted (board 2)
Wet-lab validationNone recorded
Unverified fields† first_release, latest_update, benchmarks_reported[1], release_milestones.official_announcement, open_status.weights_license, open_status.code_license
As of2026-11-16

What it does

A quantization module maps each residue's local 3D environment to a discrete structure token, then sequence-structure disentangled attention is pretrained for zero-shot mutation-effect prediction. It is the base model of VenusREM; unlike SaProt the structure vocabulary is learned rather than taken from an off-the-shelf alignment tool.

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. 2 of 2 pairings are counted.

BenchmarkRelationEval set / protocolSource typeNote and evidence
02 ProteinGymOfficial leaderboard or resultszero-shot substitution DMS summary table, several structure-vocabulary sizesBoardEntry on the official summary table · evidence
34 AbBiBenchOfficial leaderboard or resultsantibody binding-affinity correlation, structure-aware LM categoryBoardBenchmark-registered score †unverified · no source link

Other evaluations (outside the 44)

  • Inherited weakness on viral proteins (Public benchmark not among the 44) Third-party: A board maintainer post notes that models inheriting existing embeddings drop on viral-protein subsets, this one included
  • Retrieval-augmented successor as a control (Comparison against baselines only) Self-reported: Used repeatedly as the control for VenusREM, a comparison by the same group

Developers report: Developers report that learned quantized structure tokens plus disentangled attention put a roughly 110M-parameter model in the top tier for zero-shot mutation-effect prediction.

Independent evidence: No independent evidence found. No independent re-evaluation or third-party re-run found

Training data

Group in Figure 7: Public data only.

ClassValueDetail
a. Public databasesYesAbout 18.8M public structures 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 team with no separate company: Shanghai Jiao Tong University working with the Shanghai AI Laboratory and further university partners. The work appeared at a 2024 machine-learning conference and grew into a protein-engineering toolchain.
  • Funding: Paper acknowledgements not retrieved; funding left blank
  • Products and partnerships: The same group's Venus line adds retrieval augmentation on top of it to form VenusREM, alongside a protein-engineering system and a mining tool; it shares the structure-aware route with SaProt.
  • Related models: VenusREM, SaProt, PoET-2

Reading notes

  1. It leads the sequence-plus-structure category on the official board, but time-stamped "first open model past a threshold" claims are not carried here.
  2. It depends on predicted structures, and models inheriting existing embeddings generally drop on viral proteins.
  3. No wet lab: it speaks to DMS ranking, not to designability.
  4. The latest update is the conference publication; later repository notes are project news, not a new version.

Caveats

  1. The AbBiBench pairing rests only on this site's page (unverified).
  2. The conference date is month-precision and the preprint DOI is inferred from a title match (unverified).
  3. Retrieval and structure inputs raise scores, so it is not directly comparable with sequence-only models.

Sources

  1. Official repository (repository)
  2. ProSST preprint (DOI inferred from title match) (preprint)
  3. ProteinGym official summary table (leaderboard)
  4. Board maintainer blog post (second-hand) (second-hand) second-hand

Corrections

3 corrections
FieldWasNowBasisFound
latest_update.event_kindRecorded as none-since-releaseChanged to conference publication (2024-12)source2026-11-16
latest_update.descriptionSaid the repository only had README news and no new version, contradicting the recorded date's eventStates that 2024-12 is the conference publication and that the 2026 README items are project news, not model updatessource2026-11-16
first_releaseUsed the conference publication as first releaseChanged to the 2024-04-17 preprint v1, the earlier provable public evidencesource2026-11-16

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