GoKawiilPharma consortium's private protein data boosts AlphaFold-style AI models
A group of drug companies trained an open-source AlphaFold 3 replica called OpenFold3 on more than 20,000 proprietary protein structures pooled from their private vaults. The resulting model outperformed versions trained only on public data or on any single firm's siloed dataset, according to an unpublished, non-peer-reviewed blog post.
GoKawiil's interpretation of the reporting above, not reported fact.
Public repositories like the Protein Data Bank are running low on new structures to fuel further gains in AI protein-folding accuracy, so unlocking industry-held data could be key to the next leap in drug discovery tools. The result also bolsters arguments for publicly funded efforts, such as the UK's OpenBind project, to generate fresh open datasets rather than relying solely on companies to share proprietary information.
- A pharma consortium trained OpenFold3 on over 20,000 private protein structures, beating models built on public data alone.
- The unpublished result suggests pooled industry data can meaningfully improve AI protein-structure prediction.
- UK-funded OpenBind is separately building public datasets to keep such models improving without relying on corporate data.
Source: nature.com — Callaway, 2026-09-14
Published there as: “Drug firms’ secret data supercharge AI protein models”
Read the original report → The summary and analysis above are GoKawiil's own, written from reporting by the source above. Facts and quotes belong to the original publisher.