Geodesic Intelligence's NovaAtom-Lite predicts 79.07% of FoldBench antibody–antigen interfaces, ahead of AlphaFold 3
The all-atom structure model from UCLA professor Quanquan Gu's startup uses AlphaFold 3's September 2021 training cutoff; the benchmark results are company-reported and not yet independently checked.
Why it matters
Predicting where antibodies bind their targets and how small molecules sit in protein pockets is a bottleneck in drug design, and AlphaFold 3 has been the reference model. Geodesic says NovaAtom-Lite beats both AlphaFold 3 and IsoDDE on these interface tasks while sharing AlphaFold 3's training cutoff, so test structures postdate its data. It follows a preview released in September and Gu's earlier role leading ByteDance's AI drug-discovery work. Watch for independent benchmark runs and the larger NovaAtom models the company says are in training.
Key facts
- 79.07% of FoldBench's 172 antibody–antigen interfaces (DockQ ≥ 0.23)
- AlphaFold 3: 47.90% on 167 interfaces, per Geodesic
- Training cutoff 30 Sep 2021, same as AlphaFold 3
- Predicts proteins, DNA, RNA and ligands in one pass
- Preview version released in September 2026
- Founder Quanquan Gu is a UCLA associate professor