On November 10, Geodesic Intelligence, founded by UCLA professor Quanquan Gu, released NovaAtom-Lite, an all-atom structure prediction model that predicts every heavy atom in complexes of proteins, DNA/RNA, and small-molecule ligands, with pLDDT, PAE, and ipTM confidence scores. By the company's figures, it predicts 79.07% of FoldBench antibody-antigen interfaces correctly, ahead of IsoDDE at 75.58% and AlphaFold 3 at 47.90%, and also leads protein-ligand at 77.06%. Its training data stops at September 30, 2021, the same cutoff as AlphaFold 3. It is available today through the Geodesic API and the NovaDDE platform; weights are not released.

Key Takeaways

  • ✓FoldBench antibody-antigen (DockQ >= 0.23, 172 interfaces): NovaAtom-Lite 79.07%, IsoDDE 75.58% (as reported by its developers), AlphaFold 3 47.90% (FoldBench's own evaluation, 167 interfaces) (vendor figures)
  • ✓Protein-ligand 77.06% (IsoDDE 75.99%, AlphaFold 3 64.90%) and protein monomer LDDT 0.891; it does not lead 5 other tasks, and its protein-DNA score of 74.24% trails AlphaFold 3 at 79.18%
  • ✓On the harder PXM-22to25 antibody-antigen set it scores 66.05%, about 15 points above OpenDDE at 51.0%
  • ✓Test-time scaling: going from 1 to 100 seeds lifts the ranked antibody-antigen success rate from 74.3% to 83.1%
  • ✓Training cutoff 2021-09-30 (same as AlphaFold 3); available today via the Geodesic API and NovaDDE, and the founder says NovaAtom-Pro is coming in the next few months
Geodesic Launches All-Atom Structure Prediction Model NovaAtom-Lite: 79.07% on FoldBench Antibody-Antigen, Ahead of IsoDDE and AlphaFold 3, Available Today via API
🖼️Official Media
Click to view high-res
🧭

Turn your technical choice into a development budget

Compare 40 dev plans & simulate token costs vs $20/mo subscriptions

🔬

In-Depth Technical Analysis

Core Background and Industry Pain Points

Drug discovery often starts with a structure: where an antibody grips its antigen, or whether a small molecule fits a pocket, decides which wet-lab experiments come next. Since AlphaFold 3, models such as Isomorphic Labs' IsoDDE and ByteDance's Protenix have pushed structure prediction forward, yet antibody-antigen interfaces are still where models differ most. On November 10, Geodesic Intelligence, founded by UCLA professor Quanquan Gu, released NovaAtom-Lite to target exactly these interfaces.

Architecture Highlights and Underlying Mechanisms

The input is the sequences of the protein, DNA, and RNA chains in a complex plus its small-molecule ligands, given by chemical component code or SMILES. The output is the coordinates of every heavy atom with pLDDT, PAE, and ipTM confidence. It supports cyclic peptides and covalent bonds and accepts hints about contact residues and binding pockets. Geodesic has not disclosed the parameter count or network design; Gu says the team trained it in a few months on a few hundred GPUs.

Authoritative Benchmarks and Measured Comparisons

These are Geodesic's figures. On FoldBench antibody-antigen, NovaAtom-Lite succeeds on 79.07% of interfaces, versus 75.58% for IsoDDE and 47.90% for AlphaFold 3, and it leads protein-ligand at 77.06%. It leads only 3 of 8 tasks, though: its protein-DNA score of 74.24% trails AlphaFold 3 at 79.18%, and it also trails OpenDDE on protein-RNA. Sample counts differ between reports, and Geodesic itself says a gap of a point or two should be read as a tie.

Developer Adoption and Getting Started

The model is available today through the Geodesic API (platform.geodesiclab.com) and the NovaDDE platform; weights are not released. Geodesic suggests using pLDDT to decide which regions to trust and which to check before going to the wet lab. Sampling more seeds keeps improving accuracy, but compute cost rises with it. A larger NovaAtom-Pro is expected in the next few months.

Action HubReady to adopt this in production?

Benchmark side-by-side against alternatives, or calculate monthly token cost vs subscription break-even.

Up NextAnthropic shares early results from Claude in drug development: six partners including BMS, plus its own programs via Mithril Medicine
Read Now→