Five Pharma Firms Federally Fine-Tune OpenFold3 on Private Structures, Beating Public-Data Models

According to Nature, a consortium of five pharma companies including AbbVie and Astex, organized by Apheris, used federated learning to fine-tune OpenFold3 Preview 2 (an open-source reproduction of AlphaFold3) on 20,167 proprietary protein structures, with no data ever leaving each company's local environment. Tests show the new model substantially outperforms comparable models trained only on public data, marking a landmark application of federated learning in drug discovery.

Confirmed

• Training used a combined 20,167 proprietary structures from five pharma companies via federated fine-tuning, with no need to share sensitive internal structural data
• On 1,056 held-out structures, correct ligand pose prediction rose from 29% to 47%, and high-quality interface prediction rose from 35.6% to 52.1%
• The new model clearly beats public-data-trained counterparts as well as versions trained on each company's isolated dataset alone
• The work proceeded under the OpenFold coalition backed by roughly 50 companies; @AllThingsApx credits open-source AI and Mohammed AlQuraishi's team's work as a key foundation
• Per @AllThingsApx, neither the fine-tuned model nor the data will be publicly released

Why it matters

• A landmark demonstration that pharma companies can jointly train stronger folding models without exposing proprietary structural data, offering a template for model collaboration in data-sensitive industries
• The results show that scaling use of private structural data materially improves drug-discovery-relevant prediction accuracy, with direct value for pharma R&D workflows relying on AlphaFold-family tools
• @AllThingsApx notes that if protein structure prediction is still in the early innings of its scaling law, there remains large headroom for de novo discovery that does not depend on known structures; reposts from @MoAlQuraishi and others reflect broad attention across the AI–biomedicine community

2026-09-14 ~ 2026-09-15 · 6 related posts

• [source] Five pharma firms federated-finetune OpenFold3 on 20,000+ private structures, lifting ligand-pose accuracy from 29% to 47% — AllThingsApx · 2026-09-14
• [source] Federated fine-tuning of OpenFold3 by five pharma companies lifts interface predictions from 35.6% to 52.1% on private structures — AllThingsApx · 2026-09-14
• AI trained on 20,000 pharma-secret protein structures beats public-data AlphaFold models — MoAlQuraishi · 2026-09-14

3 near-duplicate retellings: AllThingsApx · TheMoonMidas · AllThingsApx