AI for Bio · Model

OpenComplex2

All-atom generative model that samples conformational distributions of proteins, nucleic acids and complexes by graph diffusion rather than predicting one structure

Structure & co-foldingNonprofit instituteChinaMaintained only中文

Overview

NameOpenComplex2 · aliases: OC2, OpenComplex-2
Full nameThe name is a project name; the preprint gives no expanded form
Ambiguous names (not auto-linked)OpenComplex
FamilyPart of the OpenComplex family
OrganizationBeijing Academy of Artificial Intelligence
Organization typeNonprofit institute
CountryChina
Model classConformational ensemble generation (co-folding with diffusion sampling)
DomainStructure & co-folding (also: Genomics & RNA)
First release · preprint
Basis: Preprint v1 posted
Note: The site recorded 2025-05-01, which is the DOI string date, not a posting date source
Other milestones
  • First public evidence: · Preprint v1 posted source
  • Preprint v1: · Preprint v1 posted source
Latest update · no later event
Basis: Preprint v2 posted; no sourced substantive event since
The code repository still holds only a readme stating that code will be released, and no journal version was found (as of 2026-11-16) source
The most recent sourced event, not a claim that nothing changed since.
StatusMaintained only · No new version or code release found within 12 months
OpennessCode: Unclear
Weights: Unclear
A free research trial is open to outside users; neither the paper nor official channels state whether code or weights will be released, and the repository holds only a readme
Benchmarks reported1 counted (board 1)
Wet-lab validationNone recorded
Unverified fields† open_status.code, open_status.weights, training_data.c_own_wetlab_for_training, background.products_partnerships
As of2026-11-16

What it does

Represents different molecule types as one graph form and runs diffusion sampling at atom, residue and motif levels, shifting the goal from predicting one structure to sampling a conformational distribution, and claims to handle very large assemblies and multiple binding sites. It complements single-structure co-folding such as AlphaFold 3 and Protenix, and unlike BioEmu-2, which also targets dynamics, it additionally claims nucleic-acid and small-molecule coverage.

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
01 CASPOfficial leaderboard or results—BoardCASP16 participant with a talk at the CASP16 meeting; official ranks not found †partly verified · evidence

Other evaluations (outside the 44)

  • Own ensemble-reproduction claims (Own task set) Self-reported: The abstract claims reproduction of known protein and RNA ensembles and capture of allostery and cryptic pockets, with no citable figures
  • A general claim of competitive structure prediction (Comparison against baselines only) Self-reported: The abstract only says competitive, naming no evaluation set or metric

Developers report: Developers report producing conformational landscapes within hours across proteins, nucleic acids and small molecules, scaling to very large assemblies.

Independent evidence: No independent evidence found. No third-party evaluation was found, and the comparison figures in the participation talk are the developers' own

Training data

Group in Figure 7: Not disclosed.

ClassValueDetail
a. Public databasesNot disclosedThe paper describes a purely data-driven system without disclosing dataset composition source
b. Private or licensed dataNot disclosed source
c. Own wet-lab data used for trainingNot disclosedThe abstract mentions incorporating experimental data without stating source or scale source
d. Wet lab used only for validationNot disclosed source

Wet-lab validation

None. The developers report no wet-lab validation.

Background

  • Organization: The developer is a nonprofit research institute founded in Beijing in 2018 as a non-enterprise unit, set up with local science-authority backing; this model belongs to its life-science model series.
  • Funding: A nonprofit with no funding rounds; the team's budget composition is not public.
  • Products and partnerships: The previous generation was that series' structure-prediction model; officially the research trial is free and registered users span over three hundred institutions, a registration count rather than usage, and no pharmaceutical partnerships are public.
  • Related models: BioEmu-2, AlphaFold 3, Protenix

Reading notes

  1. Its target is a conformational distribution while most existing benchmarks score a single structure, so the protocols do not line up.
  2. Participation is documented by an official talk, but no citable rank or value exists.
  3. The institution count is an official registration figure, not usage and not performance.
  4. Neither code nor weights have a public release statement, so its claim of parity with dynamics simulation cannot be reproduced externally.

Caveats

  1. The site's first-release date used the DOI string; it is corrected to the preprint posting date.
  2. Training data is entirely undisclosed and cannot be filed under any known source class.
  3. The abstract's capability statements carry no figures and should not be restated as evaluation results.

Sources

  1. Preprint (aggregator page) (preprint)
  2. Preprint version record (registry)
  3. Discussion-day talk at the assessment (presentation)
  4. Code repository (readme only) (repository)

Corrections

1 correction
FieldWasNowBasisFound
first_release2025-05-012025-05-02 (preprint v1 posted; v2 on 2025-05-03)source2026-11-16

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