Five Pharma Firms Federally Fine-tune OpenFold3, Lifting Ligand-Pose Accuracy 29% to 47%
rbhar90 · x · 2026-09-17
Five pharma companies fine-tuned OpenFold3 in a federated setting on 20,000+ structures, raising correct ligand-pose predictions from 29% to 47% — the model improved dramatically once trained on private pharma data. The post argues protein-ligand structure prediction isn't solved: models only work well where training data is dense, and claims of broad generalization in other bio foundation models deserve skepticism given protein structure's unusually good data curation, labels, and physical constraints.
More from Research
- Mozilla's 91-page report: open-weight AI now only ~4 months behind the frontier — rohanpaul_ai · 2026-09-17
- Protein Models ESMC and ESMFold2 Land on HuggingFace with NVIDIA-Built Fused Triton Kernels — AllThingsApx · 2026-09-17
- Panel Discussion on AI in Mathematical Research from Sept 15 Comes Highly Recommended — MindlessPapaya8463 · 2026-09-17
- OpenAI hints at millennium problem breakthrough ahead of Sept 29 DevDay — haider1 · 2026-09-17
- Vitalik: cybersecurity favors defense in the AI-hacking era, thanks to formal verification — jessi_cata · 2026-09-17
- PufferLib 5.0 hits 60M steps/sec single-GPU RL training, solves Breakout in under a second — jsuarez · 2026-09-17