Geodesic Releases NovaAtom-Lite, Beating AlphaFold 3 on Antibody-Antigen Benchmark

AI drug discovery company Geodesic has released NovaAtom-Lite, an all-atom structure prediction model built for drug discovery that can jointly predict the structures and assemblies of proteins, nucleic acids, and their binding molecules (including small molecules). UCLA professor Quanquan Gu (@QuanquanGu) posted and repeatedly reshared the news on November 11, driving wide circulation in the community.

Confirmed

• NovaAtom-Lite is an all-atom structure prediction model covering proteins, nucleic acids, and binding molecules, and can predict their interactions and assembly modes.
• On the FoldBench antibody-antigen structure prediction benchmark (DockQ ≥ 0.23, evaluated with 5 seeds), NovaAtom-Lite achieves a success rate of about 79%.
• For comparison, AlphaFold 3 scores 48% on the same benchmark, putting NovaAtom-Lite roughly 31 percentage points ahead.
• Multiple reshares note the work comes from a team of about 15 people.

Unconfirmed

• All materials so far are official announcements or secondhand reshares; there is no independent third-party replication, and details on the method, inference cost, and open-source plans were not provided.

Why it matters

Predicting structures of complexes such as antibody-antigen pairs is a key step in drug discovery and antibody design. If the 79% vs. 48% result holds up under independent evaluation, NovaAtom-Lite — built by a small team with limited resources — would meaningfully surpass AlphaFold 3 and could reshape the landscape of structure prediction and drug discovery tools. Follow-up technical reports and peer verification are worth watching.

2026-11-11 ~ 2026-11-11 · 7 related posts

• [source] Geodesic's NovaAtom-Lite beats AlphaFold 3 (79.07% vs 47.90%) on antibody-antigen folding — QuanquanGu · 2026-11-11

6 near-duplicate retellings: QuanquanGu · QuanquanGu · QuanquanGu · QuanquanGu · QuanquanGu · QuanquanGu