Skip to content- Sources: primary
- Summary: In arXiv:2611.08249, submitted 2026-11-05 and accepted at the NeurIPS 2026 AI4DD workshop, Brown and four co-authors compare broad continuation training of DiffDock against torsional supervision focused on Cartesian consequence or conditional target consistency, and against a broad and focused mixture, all under a fixed budget. Evaluation separates fixed score fidelity, generated-pose quality, best-of-10 sampling capacity and confidence-selected Top-1 performance. At under 5 Angstrom, broad continuation had the highest mean Top-1 success among trained policies and the focused or mixed variants showed no consistent advantage. The starting model had better fixed score fidelity while the broad-continuation model had higher mean Top-1 success at the same aggregate best-of-10 recovery.
- Why it matters: For teams fine-tuning diffusion docking models, the result argues against reweighting training toward geometrically important corrections without new evidence.
- Background: Molecular docking predicts how a small molecule binds a protein pocket. Diffusion docking models such as DiffDock corrupt a known pose with noise and learn a score field that reverses the corruption, so training coverage across the whole corruption range matters. Reported docking accuracy usually counts poses within 2 or 5 Angstrom RMSD of the crystal structure, and a confidence model selects which sampled pose to return, which is why score fidelity and returned-pose success can move in opposite directions.