Pharma data improve experimental AI protein model, consortium reports
A pharmaceutical consortium reported that adding 20,167 private protein structures improved an OpenFold3 model's test performance, though the work is not peer-reviewed.
The AI Structural Biology Network fine-tuned OpenFold3 with 20,167 proprietary structures showing proteins bound to potential drugs. Nature reported that the resulting model was more accurate on a held-out test set than versions trained only on public or individual-company data.
The findings were described in a blog post, have not been peer-reviewed, and the model is not publicly available. The group plans to submit a paper for peer review.