Five Pharma Firms Federated-Tune OpenFold3 on 20,167 Private Structures, Beating Public-Data Models

A consortium of five pharmaceutical companies, including AbbVie and Astex, organized by Apheris, has fine-tuned OpenFold3 Preview 2, an open-source reproduction of AlphaFold3, using federated learning on 20,167 private protein structures — all without any data leaving each company's local environment. The work was reported by Nature.

Confirmed

• Training data consisted of 20,167 proprietary protein structures pooled from the five pharma companies, fine-tuned via federated learning so that sensitive internal structural data never had to be shared
• On a held-out test set of 1,056 structures: correct ligand pose predictions rose from 29% to 47%, and high-quality interface predictions improved from 35.6% to 52.1%
• The new model significantly outperforms comparable models trained only on public data, as well as versions each company trained on its own siloed datasets
• The work advanced under the backdrop of roughly 50 companies supporting the OpenFold alliance; @AllThingsApx specifically noted that open-source AI and the work of Mohammed AlQuraishi's team were key foundations
• According to @AllThingsApx, neither the trained model nor the data will be made public

Why it matters

• This is a landmark deployment of federated learning in drug discovery: it demonstrates that pharma companies can collaboratively train stronger protein folding models without leaking proprietary structural data, offering a template for data-sensitive industries to build models together
• The results show that leveraging private structural data at scale can substantially improve prediction accuracy relevant to drug discovery, with direct value for pharma R&D pipelines that rely on AlphaFold-family tools
• Extensive discussion and reshares from @MoAlQuraishi and other bloggers reflect the broad attention the result drew across the AI–biomedicine community

2026-09-14 ~ 2026-09-14 · 5 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
• [source] AI trained on 20,000 pharma-secret protein structures beats public-data AlphaFold models — MoAlQuraishi · 2026-09-14

2 near-duplicate retellings: AllThingsApx · TheMoonMidas