AI for Bio · Model

OpenDDE

Fully open 655M-parameter co-folding model tuned for antibody-antigen, billed as an open discovery engine

Structure & co-foldingAI-native biotechUnited States, ChinaActive中文

Overview

NameOpenDDE · aliases: OpenDDE
Full nameOpen Drug Discovery Engine
Ambiguous names (not auto-linked)DDE
FamilyNot a family entry. Covers the open OpenDDE weights only; the company's internal antibody design model is related but different
OrganizationAureka AI Research, Aureka Biotechnologies
Organization typeAI-native biotech
CountryUnited States, China
Model classCo-folding (antibody-leaning)
DomainStructure & co-folding (also: Antibodies)
First release · preprint
Basis: arXiv technical report (the HF card notes a 07-03 preview; public announcement 07-06) source
Other milestones
  • First public evidence: † · GitHub preview release (per the HF card) source
  • Preprint v1: · arXiv technical report v1 source
  • Code release: † · GitHub preview release source
  • Weights release: † · Hugging Face weights source
  • Official announcement: † · Public announcement (per press coverage) source
Latest update · code release
Basis: GitHub release 1.2.0
1.2.0 adds constraint-guided inference with examples, after 1.1.0 (2026-08-16) and several 1.0.x releases source
The most recent sourced event, not a claim that nothing changed since.
StatusActive · Shipping releases through its first year
OpennessCode: Open · Apache-2.0
Weights: Open · Apache-2.0
Code, weights and benchmarks all ship under Apache-2.0, making it a fully reproducible counterpart to closed engines
Benchmarks reported1 counted (modified protocol 1, unconfirmed 1)
Wet-lab validationNone recorded
Unverified fields† benchmarks_reported[1], background.funding
As of2026-11-16

What it does

An all-atom biomolecular foundation model that predicts protein, nucleic-acid and small-molecule complexes as the entry point of a planned discovery engine, tuned for antibody-antigen generalisation. Unlike the proprietary weights of Protenix or the closed IsoDDE, it releases training code, inference pipeline, checkpoints and benchmarks together.

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 2 pairings are counted.

BenchmarkRelationEval set / protocolSource typeNote and evidence
31 FoldBenchDeveloper report, modified protocolself-defined FoldBench antibody-antigen subsetSelf-reportedSelf-defined FoldBench-AB antibody-antigen subset; not comparable to official board · evidence
36 AIntibodyUnconfirmed (not counted)Challenge 1 (computational affinity maturation)Unconfirmed sourceThe placement belongs to that company's internal model and cannot be credited to the open weights; the site matrix and model page disagree. †disputed · evidence

Other evaluations (outside the 44)

  • Self-built antibody evaluation sets (self report) Self-reported: The report also self-reports on two antibody evaluation sets, one partly self-built, each under both a top-ranked and an oracle selection protocol
  • The competition paper's overall finding (third party reading) Third-party: The AIntibody paper concludes computational design did not beat experimental methods overall; the press release's wording overstates it
  • The open-version claim (second hand) Third-party: The claim that this is the open version of the internal model appears only in a company release and one second-hand article, never in the report or repository

Developers report: Developers report the first fully open co-folding model at closed-engine level, leading on antibody-antigen tasks (self-reported, on partly self-built sets).

Independent evidence: No independent evidence found. Only self-reported tables exist; no third-party rerun found

Training data

Group in Figure 7: Public data only.

ClassValueDetail
a. Public databasesYesWeighted PDB, AFDB-multimer, distillation data, MGnify, Swiss-Prot and SAbDab source
b. Private or licensed dataNot disclosedThe paper does not mention private or licensed data source
c. Own wet-lab data used for trainingNoNot mentioned in the paper; the company has a high-throughput wet lab and calls the feedback loop a later feature source
d. Wet lab used only for validationNo

Wet-lab validation

None. The developers report no wet-lab validation.

Background

  • Organization: Aureka Biotechnologies was founded in 2023, headquartered in Laguna Hills, California with a Shanghai team; an AI-native biotech running a dry-wet closed loop plus its own pipeline, with headcount undisclosed.
  • Funding: A Series A in 2025-11 (amounts differ across second-hand sources), a $35M Series A+ on 2026-04-28, and a $100M Series B announced 2026-08-10, bringing the total to nearly $200M.
  • Products and partnerships: Open-sources this model as a design-engine module, runs its own cardiometabolic, immunology and CNS pipeline, and is building a new wet lab in Irvine; no pharma partnership disclosed.
  • Related models: IsoDDE, Protenix, ESMFold2

Reading notes

  1. The antibody figures are high but partly rest on self-built sets and need independent replication.
  2. Each antibody evaluation comes in two selection protocols; any citation must say which one.
  3. Without PoseBusters or Runs N' Poses there is no comparison of small-molecule pose accuracy.
  4. Openness is its firmest advantage: the only fully reproducible engine-level counterpart.

Caveats

  1. The competition placing belongs to the internal model, not these weights; the pairing remains an open inconsistency (unverified).
  2. The open-version framing rests on company wording plus one second-hand article.
  3. The report is internally inconsistent about one baseline model's values.
  4. The company calls the wet-lab feedback loop a later feature; this version uses no in-house data.

Sources

  1. OpenDDE technical report (preprint) ·
  2. OpenDDE GitHub releases (repository)
  3. Hugging Face weights (repository)
  4. Antibody design competition paper (journal)
  5. Press report on the open module and the company (press)
  6. Syndicated company press release (second-hand) (second-hand) second-hand

Corrections

1 correction
FieldWasNowBasisFound
benchmarks_reported.bench-36-aintibodyThe matrix marked it as an official board entryChanged to unconfirmed: the placing belongs to the company's internal antibody model, not these weights (unverified)source2026-11-16

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