OpenDDE v1.2.0 Adds Constraint Guided Antibody Design

What changed

Aureka Research posted OpenDDE v1.2.0 on 9 October 2026. The GitHub release and CHANGELOG focus on constraint guided inference for antibody and antigen pairs.

The new path uses --use_tfg_guidance true. The input JSON constraint field then steers the sampler without changing model weights. You can list contact pairs with distance windows near 3.5 to 8 angstroms. You can name which antibody chains move together as one rigid block. You can also name antigen sites the antibody must touch.

The code runs guidance in an early step band and a late step band on a fixed list of steps. On NVIDIA GPUs it can call optional Triton kernels when TF32 is off. If CUDA or Triton is missing, it falls back to a plain torch path with no error.

The release adds trunk output caching, new examples/tfg inputs, and an OpenBench2026 target list. Default inference dtype is now bf16 and TF32 defaults to off. That can shift numeric results compared with older fp32 runs.

Why it matters

Many antibody programs start with partial lab data. You might know two chains must touch but not the full fold. Guidance that honors those hints can focus GPU time on useful samples. That matters when a screen runs thousands of poses per week.

Rigid body moves are also easy to explain in a lab meeting. You can show which chain shifted and whether the code refused a clash. That beats a black box score alone.

Who is affected

Computational biologists who run OpenDDE for humanization or de novo design need the new flag and JSON rules. DevOps teams that queue OpenDDE jobs must handle bf16 defaults and pickle cache files in safe folders. Benchmark groups can line up runs with OpenBench2026 lists shipped in this tag.

Medicinal chemists who only read final structures should still know that constraints now change outputs. Old runs that ignored the constraint field are not comparable to guided runs.

What to do next

Upgrade in staging before you touch production queues. Read docs/tfg_constraint_guidance.md and docs/infer_json_format.md. Run one Fab or VHH case from examples/tfg with guidance off, then on. Check contact satisfaction with the helper script named in the release notes.

If you must match older numeric results, pass --dtype fp32 --enable_tf32 true as the changelog describes. Point OPENDDE_TRUNK_CACHE only at a directory you trust because cache files are pickles. Review disk use if many jobs share the same trunk inputs.

What to watch

Watch the OpenDDE repo for a patch after 1.2.0 that tunes guidance steps or cache modes. Watch OpenBench2026 posts for version tags and guidance settings. Watch Mac teams on older macOS builds where bf16 may fall back to fp32.

Also watch whether your own structures show fewer clashes when guidance is on but diversity drops. That tradeoff is a lab policy call, not a vendor default.

Teams that batch many JSON inputs should test trunk cache hit rates before they enable write mode on shared NFS. A cache miss still runs the full trunk, but a slow disk can erase expected savings. Legal teams should note the technical report PDF linked from the release when they file patent summaries that cite guided sampling.

Training groups unaffected by inference can stay on older tags for now, but any production pose generator should plan a cutover window. Mixed tags across clusters have caused confused bench results in other structure tools, and OpenDDE now treats constraint JSON as hard error instead of a warning.

Record the OpenDDE git hash in ELN entries when guidance steers a lead series. Auditors may ask which sampler path produced a pose later in the year.

The release moves co-folding examples into examples/cofold, so old paths in internal wiki pages need a quick search and replace before onboarding new hires.

Photo: National Cancer Institute / Pexels

Sources

• Primary — Aureka Research, OpenDDE v1.2.0 release (9 October 2026). Feature list for TFG guidance and defaults.
• Primary — Aureka Research, OpenDDE CHANGELOG 1.2.0 (9 October 2026). Constraint fields, cache behavior, and dtype changes.
• Secondary — Aureka Research, Conditional inference technical report PDF (2026). Linked from the release page for method detail.