What researchers are trying
OpenDDE proposes a shared all-atom co-folding foundation with released training and inference pipelines, checkpoints, data processing and scaling analyses.
Open reproductions of frontier drug-design engines
Can the community reproduce and extend proprietary all-atom drug-design engines with open training code, checkpoints and benchmarks?
Why it matters
Open implementations make replication, ablation, safety review, fine-tuning and domain-specific extension possible instead of limiting progress to hosted inference.
Evidence boundary
OpenDDE is a very recent July 2026 preprint. Its claimed parity has not yet received broad independent evaluation.
Organizations represented
Aureka AI OpenDDE project
Demonstrated evidence
What has actually been shown.
- Preprint-reported IsoDDE-level co-folding accuracy and released reproducibility assets.
Unresolved questions
- Can external teams reproduce the reported training and benchmark results?
- What data provenance and leakage controls are documented?
- How do open checkpoints perform in prospective discovery projects?
Signals to watch next
- Independent reproductions
- Checkpoint audits
- Training-data documentation
- Community fine-tunes
Connected evidence graph
Related BioAtlas model passports.
AlphaFold 2 / 3
The model that solved the 50-year protein-folding problem.
4/7 evidence fields documentedBoltz-1 / Boltz-2
Open-source AF3-quality structure — plus binding affinity.
4/7 evidence fields documentedChai-1 / Chai-2
An AlphaFold3-class complex predictor, made freely usable.
4/7 evidence fields documentedOpenFold
A fully open, trainable reproduction of AlphaFold 2.
3/7 evidence fields documentedProtenix
ByteDance's open reproduction of AlphaFold 3.
2/7 evidence fields documentedPrimary and evaluation sources