AI for Bio · Model

Protenix

Open co-folding series reproducing and trying to exceed the contemporary closed model, with v2 weights turned proprietary

Structure & co-foldingBig-tech labChinaActive中文

Overview

NameProtenix · aliases: Protenix, Protenix-v1, Protenix-v2
Full nameOfficial materials use Protenix; the companion tools are named PXDesign and PXMeter
Ambiguous names (not auto-linked)Protenix-Mini
FamilyNot a family entry. The v0.5 to v2 line is one record; PXDesign and PXMeter are companion tools and SeedFold has its own record
OrganizationByteDance Seed, ByteDance Seed AI for Science
Organization typeBig-tech lab
CountryChina
Model classCo-folding (plus the design tool PXDesign)
DomainStructure & co-folding
First release† partly verified · code release
Basis: First GitHub release v0.1.0; preprint 2025-01-11; the repository creation date is excluded by policy
Note: A possible 2024-11 announcement was not checked this round (partially verified) source
Other milestones
  • Preprint v1: · Preprint v1 source
  • Code release: · First GitHub release v0.1.0 (v0.3.2 on 2024-12-18) source
  • Weights release: † · The v1-line weights were open with the code source
Latest update · new version
Basis: GitHub release v2.0.0
v2 released with improved antibody-antigen prediction and proprietary weights not under an open licence; later commits are docs and maintenance source
The most recent sourced event, not a claim that nothing changed since.
StatusActive · A new version in 2026-04 with maintenance commits since
OpennessCode: Partly open · Apache-2.0
Weights: Closed · v1 weights open; v2 weights proprietary, not under an open licence
Code is Apache-2.0; v1 weights are open while v2 weights are proprietary and not released under an open licence, so third parties cannot re-check the new version
Benchmarks reported4 counted (board 1, self 1, modified protocol 1, retrospective 1, unconfirmed 1)
Wet-lab validation1 recorded campaign
Unverified fields† first_release, release_milestones.weights_release, benchmarks_reported[1], background.products_partnerships
As of2026-11-16

What it does

An open reproduction and extension of AlphaFold 3 aiming to match or beat it under the same data cutoff, with about 368M parameters in v1 and 464M in v2 plus template and RNA alignment features. Unlike OpenFold3 it uses public data only, and unlike Boltz-2 it has no affinity head.

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. 4 of 5 pairings are counted.

BenchmarkRelationEval set / protocolSource typeNote and evidence
31 FoldBenchOfficial leaderboard or resultsmaintainer leaderboardBoardMaintainer leaderboard · evidence
05 PoseBustersDeveloper report, official protocolofficial set, early version only (via the Mini paper)Self-reportedOnly an early version's table; current versions report no score on this set, so the matrix cell overstated it †unverified · evidence
30 PLINDER / PINDER / Runs N' PosesDeveloper report, modified protocolv1 self-report on the subset where all compared models have results (figure only)Self-reportedv1 self-report on the subset where all models have results, figure only; plus a third-party figure in the Pearl paper †partly verified · evidence
33 PoseXRetrospective dataset useself-docking set, benchmark authors' runThird-partyPoseX authors ran Protenix under the same protocol †partly verified · evidence
12 DUD-E / LIT-PCBAUnconfirmed (not counted)—Unconfirmed sourceSecond-hand LIT-PCBA mention, no numbers †disputed · no source link

Other evaluations (outside the 44)

  • Own ligand evaluation set (companion tool) (Own task set) Self-reported: A 623-entry in-house set scored with public validity criteria after manual removal of experimental artefacts; scores are not shown here
  • The design-side tool PXDesign (Wet-lab validation) Self-reported: The wet-lab validation belongs to the companion design tool and is not a result of the structure-prediction model
  • Third-party tables and second-hand paraphrase (third party evaluation) Third-party: Several third-party tables list v1 alongside the contemporary closed model, partly relayed by second-hand platforms under inconsistent protocols

Developers report: parity with the contemporary closed model under the same data cutoff, clearly improved antibody-antigen prediction in v2, and five seeds achieving what v1 needed a thousand seeds for.

Independent evidence: Third-party re-runs (the benchmark's authors and one comparison paper) place it below the open controls it is compared with, and with v2 weights proprietary later versions cannot be re-checked by third parties.

Training data

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

ClassValueDetail
a. Public databasesYesPDB (cutoff 2021-09-30, plus a 2025-06-30 version), alignments and distillation data; the training data and alignment pipeline are open source
b. Private or licensed dataNo
c. Own wet-lab data used for trainingNo
d. Wet lab used only for validationYesWet lab validates only the companion design tool PXDesign; the structure-prediction side has none 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.

PXDesign binder campaign (design preprint v3 (2025-12-17))
Targets 7 · Designs tested — · Hits — · Success nanomolar hits on 6/7 targets; hit rates 17-82% · Best affinity —

Denominator basis: Seven targets, six of which yielded nanomolar hits; per-target denominators were not opened in primary material
The range published earlier came from a stale repository README; corrected to the design preprint and cited to the design paper rather than the model repository; developer-reported, not an independent replication

Self-reported · † · verified · source

Background

  • Organization: It comes from a large internet company's foundation-model research division, whose science agenda spans biological foundation models, quantum chemistry and molecular dynamics, funded from the company's internal budget.
  • Funding: Big-tech internal budget, no separate fundraising
  • Products and partnerships: The series and its companion design and evaluation tools are open, and the same company also has SeedFold; a 2026-06 report says it is spinning out an AI drug-discovery unit of about 50 people with the structure team joining, without stating whether this series moves (second-hand, unverified).
  • Related models: AlphaFold 3, OpenFold3, RF3, Boltz-2, SeedFold

Reading notes

  1. On the maintainer board v1 sits in the same band as the contemporary closed model under official framing, which supports parity but not superiority.
  2. With v2 weights closed, later versions rest on official figures alone and cannot be re-checked by third parties.
  3. The only summary figure available on the site comes from a third-party paper, as the official v1 report gives binned figures only.
  4. The developers prefer their own off-site evaluations, which sit closer to antibody and RNA use cases and must not be mixed with site entries.

Caveats

  1. The v2 weights are proprietary, so openness must be read separately for code and weights.
  2. The public pose-set pairing holds only for an early version and must carry a version qualifier.
  3. The wet-lab figures belong to the companion design tool, not to the structure-prediction model.
  4. One pairing rests only on the site's own page (unverified).
  5. The report on the company's spin-out is second-hand and does not say whether Protenix moves with it (unverified).

Sources

  1. Official repository (repository)
  2. Official repository releases (repository)
  3. The model's preprint (preprint)
  4. Design tool preprint and project page (preprint)
  5. The team's own compact-version paper (preprint)
  6. Third-party re-run paper (third-party paper)
  7. Re-run paper by the benchmark's authors (third-party paper)
  8. Coverage of the company's spin-out (second-hand) (second-hand) second-hand

Corrections

4 corrections
FieldWasNowBasisFound
first_releaseFirst release taken as the preprint date 2025-01-11Changed to the first code release on 2024-12-12; the repository creation date is excluded by policysource2026-11-16
latest_updateThe v2 release was dated 2026-04-08Changed to 2026-04-07 per the release page; later commits are documentationsource2026-11-16
wetlab_validationThe design-side wet-lab hit-rate range was 20-73%, cited to the model repositoryChanged to the design preprint v3 range of 17-82% with nanomolar hits on 6 of 7 targets, cited to the design paper rather than the model repositorysource2026-11-16
benchmarks_reportedThe public pose-set cell was shown as a self-report for the current versionVersion-qualified: it holds only for an early version, as neither v1 nor v2 reports on that test setsource2026-11-16

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