DEV Community

Cover image for NovaAtom-Lite vs AlphaFold 3: Antibody Win, Not a Sweep
Shamuddin
Shamuddin

Posted on Originally published at aitecharchive.com

NovaAtom-Lite vs AlphaFold 3: Antibody Win, Not a Sweep

Verdict: On antibody-antigen interfaces, NovaAtom-Lite beats AlphaFold 3 by a wide margin on FoldBench: 79.07% versus 47.90% success, according to its developer, Geodesic Intelligence. It also leads on protein-ligand interfaces, so antibody and small-molecule drug-discovery teams have the most to gain. It does not win everything, because AlphaFold 3 still leads on protein-DNA, and every figure here is vendor-reported with no independent replication cited yet.

Last verified: 14 November 2026. Author: Sham Uddin, founder and editor of aitecharchive.com.

TL;DR

  • NovaAtom-Lite, introduced on 10 November 2026, is Geodesic Intelligence's first all-atom structure model, covering proteins, DNA, RNA and small molecules.
  • Vendor-reported FoldBench results put it ahead of AlphaFold 3 on antibody-antigen and protein-ligand interfaces.
  • AlphaFold 3 keeps the lead on protein-DNA interfaces, and another model, OpenDDE, leads protein-RNA and protein-protein.
  • All numbers are self-reported, including the IsoDDE comparison figure, which comes from IsoDDE's own developers.
  • Weights and pricing are not published, so labs cannot yet reproduce the result locally.

What is NovaAtom-Lite and who built it?

NovaAtom-Lite is the first model in the NovaAtom family from Geodesic Intelligence. According to the launch post, it predicts the 3D position of every heavy atom in assemblies of proteins, DNA, RNA and small-molecule ligands, and reports pLDDT, PAE and ipTM confidence scores. A task-by-task write-up describes the architecture as a recycled pairformer trunk with atomic diffusion, which is the same broad family of design as AlphaFold 3 from Google DeepMind.

The company was founded by Quanquan Gu, a UCLA professor and former ByteDance AI scientist. South China Morning Post reported on 12 November that he said on social media the model had topped a third-party benchmark, and Lets Data Science's write-up relays the same 79.1% claim against IsoDDE at 75.6% and AlphaFold 3 at 47.9%, with Isomorphic Labs' IsoDDE as the incumbent Isomorphic model. That is the first independent media coverage, but it relays a claim rather than testing it.

How does NovaAtom-Lite vs AlphaFold 3 compare on FoldBench?

FoldBench is a low-homology benchmark of 1,522 biological assemblies from PDB entries deposited between 13 January 2023 and 1 November 2024, filtered for similarity to AlphaFold 3's training set. The table below collects the figures reported by the vendor and its coverage. Interface tasks use DockQ of 0.23 or higher as the success threshold.

Task NovaAtom-Lite IsoDDE AlphaFold 3 Leader
Antibody-antigen interfaces 79.07% (n=172) 75.58% 47.90% (n=167) NovaAtom-Lite
Protein-ligand interfaces 77.06% (n=558) 75.99% 64.90% NovaAtom-Lite
Protein monomer (LDDT) 0.891 (n=334) not stated 0.880 NovaAtom-Lite
Protein-DNA interfaces 74.24% not stated 79.18% AlphaFold 3

The gaps are not uniform. The antibody-antigen lead over AlphaFold 3 (79.07% against 47.90%) is the widest in the table, while the monomer lead (0.891 against 0.880 LDDT) is small enough that few structural biologists would change a workflow over it. The protein-ligand lead (77.06% against 64.90%) is meaningful but smaller than the antibody result. Note also that AlphaFold 3 was scored on 167 antibody-antigen interfaces by FoldBench's own evaluation, while NovaAtom-Lite was scored on all 172, so the sets are close but not identical.

Where does NovaAtom-Lite not lead?

According to the petagenius breakdown, NovaAtom-Lite ranks first on 3 of the 8 charted FoldBench tasks and in the top three on 7 of 8. AlphaFold 3 leads protein-DNA at 79.18%, and OpenDDE leads protein-RNA at 73.5% and protein-protein at 76.9%. Anyone working on transcription factor complexes or RNA-binding proteins should not read the headline as a reason to switch.

The stricter tier is more favourable. At DockQ of 0.80 or higher, the same source reports 41.28% for NovaAtom-Lite against 19.2% for AlphaFold 3. If that holds up, it suggests the antibody gain is about precise interface geometry and not only about getting roughly the right pose.

Sampling depth also moves the score. Raising the sample count from one seed to 100 lifts the ranked antibody-antigen success rate from 74.3% to 83.1%, according to Geodesic's own protocol figures. On the harder PXM-22to25 antibody-antigen set, the model scored 66.05% against OpenDDE's 51.0%, a roughly 15-point gap on targets newer than the standard FoldBench window. Both figures are vendor-reported.

Can you trust the NovaAtom-Lite numbers?

Treat them as claims. Four reasons:

  1. Every figure is vendor-reported. The IsoDDE number is developer-reported too, so the comparison mixes two companies' own measurements.
  2. The result was announced on social media, and SCMP's report quotes the founder rather than an outside evaluator.
  3. No peer review or independent replication is cited in any source I found.
  4. The sampling protocol favours the new model slightly. The vendor ranks the top prediction out of 25 samples, so a like-for-like comparison depends on AlphaFold 3 being run under a comparable budget.

There is one genuine fairness control. The training cutoff is 30 September 2021, the same as AlphaFold 3, and all benchmark targets were deposited after that date. That limits the most obvious form of leakage, though it does not replace outside testing.

What does this change for antibody and drug-discovery teams?

For research teams, the practical points are modest and concrete:

  • Treat it as a second opinion. Run NovaAtom-Lite alongside AlphaFold 3 on antibody-antigen complexes and look at where they disagree, rather than replacing your current pipeline.
  • Plan for multiple samples. The reported result depends on ranking several samples per target, so budget compute accordingly and use the ipTM and PAE outputs to filter.
  • Check access first. Access is through the Geodesic API and the NovaDDE platform. Weights and code are not released and pricing is not published, per the same coverage.
  • Wait for NovaAtom-Pro. A larger model is expected in the coming months, which may change the picture again.

This is research tooling for computational structure work, and nothing here is clinical guidance. For related context on how AI-for-science claims get tested, see our piece on AlphaGenome 2 and variant effects, the BioBasicsBench wet-lab benchmark, the Hydrogen 1.0 vs GPT-6 Astra science benchmarks, and our look at AI cheminformatics agents.

Which model should you use today?

For antibody-antigen and protein-ligand work, NovaAtom-Lite is the model to trial first, because the reported margins are large enough to justify the effort even after discounting for vendor bias. For protein-DNA, stay with AlphaFold 3. For protein-RNA and protein-protein, the reported leader is OpenDDE. Revisit all of this once an independent group publishes its own FoldBench run.

FAQ

Q: Is NovaAtom-Lite better than AlphaFold 3?
A: On antibody-antigen and protein-ligand interfaces, the vendor-reported FoldBench numbers say yes. On protein-DNA interfaces, AlphaFold 3 scores higher, and no independent replication of the comparison has been cited yet.

Q: What did NovaAtom-Lite score on antibody-antigen prediction?
A: Geodesic Intelligence reports 79.07% success on 172 interfaces, against 75.58% for IsoDDE and 47.90% for AlphaFold 3. SCMP rounds the figure to 79.1%.

Q: What is FoldBench?
A: It is a low-homology benchmark of 1,522 biological assemblies from PDB entries deposited between January 2023 and November 2024, filtered for similarity to AlphaFold 3's training data.

Q: Can I download NovaAtom-Lite weights?
A: No. Weights are not released, and access is through the Geodesic API and NovaDDE platform. Pricing is not published.

Q: Who founded Geodesic Intelligence?
A: Quanquan Gu, a UCLA professor and former ByteDance AI scientist, according to SCMP.

Corrections and disclosure

No corrections have been issued for this article. It was produced with AI assistance and reviewed against the primary sources linked above; figures marked as vendor-reported have not been independently verified.

Top comments (0)