What's changed

🚀 Key Highlights

• Protenix-v2 Model Released: Introduced protenix-v2, an enhanced-capacity model (464M parameters). It delivers significant accuracy improvements in predicting challenging antibody-antigen complex structures and updates ligand-related plausibility.
• Training-Free Guidance (TFG) Module: Introduced a powerful new guidance module enforcing geometric and physical constraints (Steric, Torsion, Bond, etc.) during diffusion sampling without the need for retraining.

✨ New Features & Enhancements

• Inference Efficiency Breakthrough: protenix-v2 shows remarkable efficiency gains. Utilizing only 5 sampling seeds, it successfully
outperforms protenix-v1 at 1000 seeds.
• Configurable TFG Capabilities: Exposed via the --use_tfg_guidance CLI flag. Supported geometries include VinaStericPotential,
ExperimentalTorsionPotential, and PairwiseDistancePotential.

📖 Documentation & Assets

• Bumped protenix version to 2.0.0.
• Published the new Protenix-v2 Technical Report (docs/PX2.pdf).
• Updated README.md and docs/supported_models.md with the latest Protenix-v2 benchmarks, showcasing a 9 to 13 percentage points absolute success rate gain over Protenix-v1 at the DockQ > 0.23 threshold.