NovaAtom-Lite debuts: all-atom structure model beats AlphaFold 3 on antibody–antigen prediction
QuanquanGu · x · 2026-11-11
Geodesic Intelligence released NovaAtom-Lite, an all-atom structure prediction model for drug discovery that predicts proteins, nucleic acids, and ligands together as full complexes, supporting covalent bindings and contact pockets, with per-residue pLDDT/PAE/ipTM confidence outputs.
Benchmark results (FoldBench, antibody–antigen, 172 interfaces, DockQ ≥ 0.23):
- NovaAtom-Lite: 79.07%
- IsoDDE: 75.58%
- AlphaFold 3: 47.90%
On the protein–ligand task (538 interfaces) it scores 77.06%, beating IsoDDE (75.99%) and AlphaFold 3 (64.90%); on PXMeter 22to25 it reaches 66.05% vs OpenDDE's 51.0%. Evaluation uses 5 seeds × 5 samples with top-1 selected by the model's own ranking score. For novel approximate ligands, multi-sample plus pLDDT filtering lifts success from 74.3% (1 seed) to 83.1% (100 seeds); pLDDT>90 corresponds to 0.47 Å central Cα error. Available now via the Geodesic API and NovaDDE.
More from Models
- Frontier model speed gap is underrated: 1.8 min per task vs 12.7 for DeepSeek — rickasaurus · 2026-11-11
- Rumor: OpenAI holding GPT-6.2 Astra until Anthropic's Nov 19 IPO to steal the spotlight — kimmonismus · 2026-11-11
- Opus 5.5 is finally good at subtle humor, from joke documentaries to silly PSAs — andrey_kurenkov · 2026-11-11
- Paying 25x for 0.06 points: blind test shows max reasoning effort barely beats medium — aakashgupta · 2026-11-11
- Jason Wei maps which benchmarks frontier labs actually reported in 2026 — _jasonwei · 2026-11-11
- OpenAI report: 5 eval agents broke into its research infra using leftovers from prior runs — tszzl · 2026-11-11