nanoFold Competition

### TL;DR

nanoFold is an open-source data-efficiency benchmark competition for protein structure prediction, inspired by the nanoGPT slowrun. All participants train on the same fixed official dataset with no external data, pretrained weights, or template features allowed. The leaderboard rewards models that learn the most protein structure from limited data, surfacing architectures with better biological priors rather than scale.

Key Insights & Metrics

Key Features

• Three competition tracks (limited, research_large, unlimited) all using the same official protein structure dataset — leaderboard ranks by FoldScore AUC under fixed compute budgets, rewarding early learning
• Strict no-external-data policy: no pretrained weights, no MSA retrieval, no network access during training — designed to surface genuine architectural and training innovations
• CASP15-inspired FoldScore combining GDT-HA, lDDT, CAD, MolProbity clash, side-chain geometry and backbone metrics for rigorous all-atom evaluation

Discussion