Geodesic Intelligence says its NovaAtom-Lite model has beaten AlphaFold 3 and Isomorphic Labs’ IsoDDE on several biomolecule prediction tests. The results, reported by 36Kr, are a notable benchmark claim, not evidence that the model can produce medicines or shorten drug development to four years.
Watch Desk analysis
What happened
36Kr reports that NovaAtom-Lite ranked first on the FoldBench tests it describes, including antibody-antigen and protein-ligand prediction. The article says that, on 172 antibody-antigen interfaces, the model’s success rate was 3.5 percentage points above IsoDDE’s and 31 points above AlphaFold 3’s. At a stricter accuracy threshold, it reports NovaAtom-Lite’s success rate was 2.2 times AlphaFold 3’s.
For protein-ligand prediction, the article says NovaAtom-Lite ranked first on two measures across 558 interfaces. It also reports that on a separate test of unfamiliar molecules, its success rate was 47.6%, against 24% for AlphaFold 3. These are figures reported in the article, not independently established results here. Read 36Kr’s report.
Why it matters
Predicting how proteins, antibodies and small molecules fit together is relevant to early drug research. If the reported results hold up, a model that performs better on unfamiliar molecules could help researchers prioritise what to test in the lab. That is a useful scientific lead, not a shortcut past experiments, clinical trials or regulatory approval.
Our read
The numbers are interesting enough to merit attention, especially the test on unfamiliar molecules. But benchmark leadership is a starting point for scrutiny, not a medicinal miracle in a lab coat. The important next question is whether other researchers can reproduce the results and find them useful in real drug-discovery work.
What to watch
- Whether independent researchers reproduce NovaAtom-Lite’s results on the named tests.
- How performance compares on new targets and in experimental laboratory work.
- Whether Geodesic Intelligence publishes further technical detail on the model and evaluations.
Discussion spark: What should count as convincing evidence that a biomolecule-prediction model can help drug discovery: strong benchmark scores, independent replication, or results in the lab?
Sources and evidence
- 10 Years Is Too Slow: Gu Quanquan & Team Challenge Dario – Compress Cancer Cure Timeline to Just 4 Years – eu.36kr.com (16 November 2026, 06:11 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.