Protenix-v2
Predict target–binder complex cohorts.
Results
Loading starting example…
Method & evidence
Co-fold target and binder sequences with Protenix-v2, explicit sampling seeds, checksum-bound structures, per-token PAE, confidence files, and row-level failures.
Scope & limitations
- The released deterministic profile is byte-repeatable for identical input and seed on the qualified H100 image.
- Experimental binding and orthogonal structural validation remain required.
Interpreting results
- Different seeds are deliberate inference-time samples, not reproducibility defects.
- PAE and confidence are model evidence and do not establish binding affinity.
- Compare candidates within a target and prediction method before applying score cutoffs.
References
Input requirements & outputs
Input checks
- 20–2,000 target residues and 10–500 binder residues
- Up to 8 rows in one native cohort
- One to five distinct non-negative sampling seeds
Expected results
- Seed-expanded predicted complexes — One structure and PAE binding per successful row and seed.
- Prediction Artifacts — Checksum-bound mmCIF, full PAE, and confidence evidence.
- Native cohort evidence — Complete model, runtime, seed, per-row status, request, result, and Artifact provenance.
