AI for Bio · Model

VenusREM

Retrieval-augmented zero-shot mutation-effect model fusing sequence, structure and homologs

ProteinsAcademic labChinaActive中文

Overview

NameVenusREM · aliases: VenusREM
Full nameThe paper uses the short name only
Ambiguous names (not auto-linked)Venus, VenusREM2
FamilyNot a family entry. The successor VenusREM2 and its variants appear only in repository news and are not merged in
OrganizationShanghai Jiao Tong University, East China University of Science and Technology, Shanghai AI Laboratory
Organization typeAcademic lab
CountryChina
Model classProtein language model (retrieval-augmented zero-shot)
DomainProteins (also: Structure & co-folding)
First release · preprint
Basis: Preprint v1 posted source
Other milestones
  • First public evidence: · Preprint v1 source
  • Preprint v1: · Preprint v1 source
  • Journal online: · Journal version online source
  • Code release: † · Official repository release v1.0.0 source
Latest update† partly verified · successor
Basis: The final repository commit announces results for the successor VenusREM2
Repository news announces the successor VenusREM2's board result (README only, unverified); earlier v1.0.0 on 2026-04-25 and journal online 2025-07-15
Note: The successor's preprint identifier and score appear only in the README and are not independently verified (unverified) source
The most recent sourced event, not a claim that nothing changed since.
StatusActive · Repository release and successor news within 12 months
OpennessCode: Open
Weights: Open
Code in the official repository and weights in a public model hub; licence text not retrieved this round, so terms are left empty
Benchmarks reported1 counted (board 1)
Wet-lab validation1 recorded campaign
Unverified fields† latest_update, release_milestones.code_release, background.products_partnerships, open_status.weights_license, open_status.code_license
As of2026-11-16

What it does

Builds on the quantized-structure base of ProSST and adds retrieved homologous sequences plus structure information to score point mutations zero-shot, with no DMS labels in training. It is a multimodal sequence-plus-structure-plus-retrieval scorer, a peer of SaProt with an added retrieval path.

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. 1 of 1 pairings are counted.

BenchmarkRelationEval set / protocolSource typeNote and evidence
02 ProteinGymOfficial leaderboard or resultszero-shot substitution official summary table (Structure & MSA category)BoardEntry on the official summary table · evidence

Other evaluations (outside the 44)

  • Specialist-predictor control in an external meta-analysis (Public benchmark not among the 44) Third-party: A third-party meta-analysis uses it as a leading specialist predictor against general frontier models, abstract-level material only
  • Maintainer blog grouping of multimodal models (Comparison against baselines only) Third-party: Board maintainers group it with another multimodal model at the top, a directional observation

Developers report: Developers report that retrieval augmentation fused with structure and homolog information gives then-best zero-shot mutation-effect prediction, backed by enzyme-mutant wet-lab work.

Independent evidence: A third-party meta-analysis treats it as a leading specialist predictor against general models (abstract-level material); no independent replication of its wet lab.

Training data

Group in Figure 7: Public data + wet lab for validation only.

ClassValueDetail
a. Public databasesYesUnsupervised with no DMS training; base data as in ProSST 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 validationYesTen new DNA polymerase mutants assayed for high-temperature activity; a separate batch of antibody mutants is a retrospective analysis source

Wet-lab validation

Every rate carries its numerator and denominator. Claims that could not be verified are shown as unverified, without numbers; different paper versions are separate entries.

DNA polymerase mutants, high-temperature activity (journal full text)
Targets 1 · Designs tested 10 · Hits 8 · Success 8/10 · Best affinity best mutant about 8-fold

Denominator basis: The top ten single-point mutants from a saturation scan were assayed (small n, a single enzyme target)
Developer-reported, not an independent replication: eight mutants significantly improved activity and three improved thermostability and activity together. A separate batch of about 31 antibody mutants is a retrospective analysis of already published data, not a new experiment.

Self-reported · · verified · source

Background

  • Organization: An academic team with no separate company: released in 2025-07 by Shanghai Jiao Tong University with a chemical-technology university and the Shanghai AI Laboratory, largely the same author group as ProSST. The public model-hub entry sits under a research organization account.
  • Funding: Paper acknowledgements not retrieved; funding left blank
  • Products and partnerships: Built on ProSST with retrieval augmentation; the same Venus line includes a protein-engineering system and a mining tool, and the successor VenusREM2 appears only in repository news (unverified).
  • Related models: ProSST, SaProt, PoET-2

Reading notes

  1. It shares the official board's top tier with another retrieval-augmented model; differences inside that tier sit at noise level, so do not read rank as a gap.
  2. Retrieval augmentation and structure inputs raise scores markedly, so it is not directly comparable with sequence-only models.
  3. The only new experiment is ten enzyme mutants (8/10 improved activity): a single target with small n, which says nothing about design hit rates.
  4. The roughly 31 antibody mutants are a retrospective analysis and must never be added to the new experiment.

Caveats

  1. The successor VenusREM2's result and preprint identifier appear only in repository news (unverified).
  2. The wet-lab work is developer-reported, not an independent replication.
  3. The paper's state-of-the-art claim and the board's tie at the top are two different statements.

Sources

  1. VenusREM journal paper (journal)
  2. Journal full text, source of the wet-lab figures (journal)
  3. VenusREM preprint v1 (preprint)
  4. Official repository (repository)
  5. ProteinGym official summary table (leaderboard)
  6. Public model-hub entry (repository)

Corrections

3 corrections
FieldWasNowBasisFound
wetlab_validationThe site said per-mutant data were absent and no hit rate could be assessedThe journal full text reports 8 of 10 mutants significantly improved, the best about 8-fold, and 3 improving thermostability and activity togethersource2026-11-16
wetlab_validationAbout 31 antibody mutants were treated as another experimentRecorded as a retrospective analysis of published data, not a new experiment, and not added to wet-lab scalesource2026-11-16
latest_update.date2026-09-302026-09-29, the date of the final repository commitsource2026-11-16

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