Predicting the structure of molecular assemblies with NovaAtom-Lite

From sequences, NovaAtom-Lite predicts biomolecular assemblies. Its reported gains focus primarily on the interfaces between their components.

Proteins, DNA, RNA, and small molecules can be modeled together using NovaAtom-Lite, Geodesic Intelligence’s new structure predictor. Given sequences and descriptions of the associated molecules, it estimates the coordinates of their heavy atoms, meaning all atoms other than hydrogen. The prediction covers the assembly and the interfaces between its components.

Access is advertised through the Geodesic API or NovaDDE, the provider’s drug discovery workspace. The resulting structures include confidence metrics to help teams examine predicted regions and interactions before planning their experiments. What the 79.07% score measures

In its results overview, the lab highlights a 79.07% success rate across FoldBench’s 172 antibody–antigen interfaces. For each target, it generates 25 predictions from five random seeds, each producing five samples. Only the prediction ranked first by the model’s internal score is evaluated.

Here, success means a DockQ score of at least 0.23. The official DockQ scale places this threshold in the “acceptable” quality category: it does not mean that every atomic position is reproduced perfectly. At the more demanding threshold of 0.80, classified as “high quality,” the reported success rate for NovaAtom-Lite is 41.28%.

The provider sets the training structure cutoff at September 30, 2021. The FoldBench assemblies used in this comparison come from structures deposited in 2023 and 2024. This time gap is part of the protocol for evaluating cases that postdate the declared training data. A comparison across publications, with different sample sizes

The 75.58% rate cited for IsoDDE appears in the Isomorphic Labs technical report. The 47.90% figure for AlphaFold 3 comes from the FoldBench evaluation reproduced in the table. That score covers 167 evaluable interfaces, compared with 172 for NovaAtom-Lite. The three figures therefore do not come from a new comparison conducted entirely by an independent evaluator on exactly the same dataset.

The reported lead over IsoDDE is 3.49 percentage points for antibody–antigen interfaces. For protein–ligand interfaces, the lab reports 77.06%, compared with 75.99% for IsoDDE, a difference of 1.07 percentage points. This second task uses different criteria, combining the accuracy of the ligand’s position with the quality of its local interactions. Rankings vary by category: the model does not lead every protein-related task or every task involving nucleic acids presented in the publication. Generating more structures does not fully solve the ranking problem

The study also examines the effect of the compute budget. With five samples per seed, the average antibody–antigen success rate rises from 74.3% with one seed to 83.1% with 100. The average for five seeds reaches 78.3%; the headline figure of 79.07% comes from a particular run. The curve is calculated by drawing subsets from the predictions, with uncertainty intervals obtained through resampling.

The lab distinguishes selection by internal score from an “oracle” selection, which retrospectively picks the structure closest to the known reference. This second measure points to room for improvement in ranking: generating a good structure does not guarantee that the model can identify it among its candidates.

To examine the outputs, researchers have access to pLDDT, which estimates local confidence per residue, along with metrics for relative errors and interface confidence. Additional information can also guide predictions: covalent bonds, closure of a cyclic peptide, or a proposed contact site. The publication distinguishes bonds specified in the input from pocket and contact hints, which are treated as clues rather than guaranteed constraints.