Research release

Introducing NovaAtom-Lite

All-atom structure prediction for proteins, nucleic acids and the molecules they bind.

Geodesic Intelligence Try NovaAtom-Lite

Today, we’re introducing NovaAtom-Lite, Geodesic’s most capable model yet for all-atom structure prediction. Given the sequences of a biomolecular assembly’s protein, DNA and RNA chains, together with its small-molecule ligands, NovaAtom-Lite predicts the position of every heavy atom, with confidence for each residue and each pair of residues. Many drug-discovery questions start with a structure, and NovaAtom-Lite’s clearest advantage over other models is at the interfaces they turn on: between antibodies and their protein antigens, and between proteins and small-molecule ligands. On these critical drug-discovery tasks, NovaAtom-Lite outperforms both IsoDDE and AlphaFold 3, setting a new bar for molecular structure prediction. NovaAtom-Lite is trained on structures released up to September 30, 2021, the same cutoff as AlphaFold 3[1], and every benchmark below is drawn from structures released after that date.

FoldBench, antibody–antigen

79.07%

of FoldBench’s 172 antibody–antigen interfaces predicted correctly (DockQ ≥ 0.23) by NovaAtom-Lite’s top-ranked prediction, chosen from 25 per target.

  • NovaAtom-Lite 79.07% all 172 interfaces; our run, scored with FoldBench’s official evaluation
  • IsoDDE 75.58% as reported by its developers[2]
  • AlphaFold 3 47.90% FoldBench’s own evaluation[3], on the 167 interfaces it could assess

All eight FoldBench tasks The antibody–antigen cases

What NovaAtom-Lite predicts.

Proteins, nucleic acids and small molecules, folded together in one prediction.

NovaAtom-Lite, in eight predicted molecules

  1. N:8QABThree helices
  2. O:8Q1MAChBP pentamer
  3. V:8A51BRME1 helices
  4. A:7Z5KMYF5 dimer
  5. A:8IPGHP101 peptides
  6. T:8P6KM3 protein
  7. O:9ZENNucleosome DNA
  8. M:9ZENHistone H3 pair
Seven NovaAtom-Lite predictions of benchmark targets, each shown first with its partner chains; then eight of their molecules spell NOVAATOM. Ligands are not drawn. Under each letter: the PDB entry and the molecule.

It reads

Protein
Any number of chains, including modified and non-standard residues.
DNA and RNA
Single- and double-stranded DNA, structured RNA, with their protein partners.
Ligands
Small molecules, cofactors, metals and ions, by chemical component code or SMILES.

It returns

Coordinates
Every heavy atom, several independent samples per target, ranked.
pLDDT
Per-residue confidence in the local structure.
PAE and ipTM
Expected error between residues, and per-interface reliability.

It also takes constraints and hints

Cyclic and covalent bonds are built into the input; contacts and pockets are hints the model weighs, not constraints.

  • Cyclic peptides

    Mark a peptide as cyclic, and the bond that closes the ring is added for you.

  • Covalent bonds

    Link two atoms, as for a covalent inhibitor, a glycan or a cross-link.

  • Contacts and pockets

    Point to residues that touch, or to the pocket where a ligand binds.

How NovaAtom-Lite performs.

Structure prediction across interfaces and monomers, against published baselines.

NovaAtom-Lite on FoldBench

FoldBench[3] is a low-homology benchmark of 1,522 biological assemblies from PDB entries deposited between January 13, 2023 and November 1, 2024, filtered for similarity to AlphaFold 3’s training set. We chart eight of its tasks: five interface types scored by success rate, and protein, DNA and RNA monomers scored by LDDT.

Charted: antibody–antigen, protein–ligand, protein–protein, protein–RNA and protein–DNA interfaces, and protein, DNA and RNA monomers.

  • 2023–2024 PDB deposit window, January 13, 2023 to November 1, 2024
  • 1,522 biological assemblies in FoldBench
  • 8 tasks charted

Antibody–antigen

Success rate (%), DockQ ≥ 0.23Higher is better; official n = 172 interfaces

  • DockQ ≥ 0.80
  • 0.49 ≤ DockQ < 0.80
  • 0.23 ≤ DockQ < 0.49
  • NovaAtom-Lite, sample 17279.07 (best value); segments show the stricter tiers
  • IsoDDE, sample 17275.58; tier split not reported, drawn whole
  • OpenDDE, sample 104 entries / 160 clusters70.0; segments show the stricter tiers
  • SeedFold, the task's full set, 172; its report states no count53.2; tier split not reported, drawn whole
  • ESMFold2, sample 17253; segments show the stricter tiers
  • Protenix-v1, sample 104 entries / 160 clusters52.3; segments show the stricter tiers
  • AlphaFold 3, sample 16747.90; segments show the stricter tiers
  • Boltz-1, sample 16433.54; segments show the stricter tiers

Protein–ligand

Success rate (%), ligand RMSD < 2 Å and LDDT-PLI > 0.8Higher is better; official n = 558 interfaces

  • NovaAtom-Lite, sample 55877.06 (best value)
  • IsoDDE, sample 55875.99
  • AlphaFold 3, sample 54764.90
  • SeedFold, the task's full set, 558; its report states no count63.12
  • Protenix-v1, sample 55862.5
  • ESMFold2, sample 51161
  • OpenDDE, sample 55860.1
  • Boltz-1, sample 47655.04

Protein monomer

LDDTHigher is better; official n = 334 chains

  • NovaAtom-Lite, sample 3340.891 (best value)
  • ESMFold2, sample 3330.890
  • OpenDDE, sample 3300.890
  • SeedFold, the task's full set, 334; its report states no count0.889
  • Protenix-v1, sample 3300.886
  • AlphaFold 3, sample 3340.880
  • Boltz-1, sample 3300.870

Protein–protein

Success rate (%), DockQ ≥ 0.23Higher is better; official n = 279 interfaces

  • DockQ ≥ 0.80
  • 0.49 ≤ DockQ < 0.80
  • 0.23 ≤ DockQ < 0.49
  • NovaAtom-Lite, sample 27975.63; segments show the stricter tiers
  • OpenDDE, sample 27976.9 (best value); tier split not reported, drawn whole
  • ESMFold2, sample 27876; segments show the stricter tiers
  • IsoDDE, sample 27974.19; tier split not reported, drawn whole
  • SeedFold, the task's full set, 279; its report states no count74.03; tier split not reported, drawn whole
  • AlphaFold 3, sample 26672.93; segments show the stricter tiers
  • Protenix-v1, sample 237 entries72.7; tier split not reported, drawn whole
  • Boltz-1, sample 25268.25; segments show the stricter tiers

Protein–RNA

Success rate (%), DockQ ≥ 0.23Higher is better; official n = 70 interfaces

  • DockQ ≥ 0.80
  • 0.49 ≤ DockQ < 0.80
  • 0.23 ≤ DockQ < 0.49
  • NovaAtom-Lite, sample 7065.71; segments show the stricter tiers
  • OpenDDE, sample 7073.5 (best value); tier split not reported, drawn whole
  • ESMFold2, sample 3972; segments show the stricter tiers
  • SeedFold, the task's full set, 70; its report states no count65.31; tier split not reported, drawn whole
  • AlphaFold 3, sample 6962.32; segments show the stricter tiers
  • Boltz-1, sample 5856.90; segments show the stricter tiers

Protein–DNA

Success rate (%), DockQ ≥ 0.23Higher is better; official n = 330 interfaces

  • DockQ ≥ 0.80
  • 0.49 ≤ DockQ < 0.80
  • 0.23 ≤ DockQ < 0.49
  • NovaAtom-Lite, sample 33074.24; segments show the stricter tiers
  • AlphaFold 3, sample 31779.18 (best value); segments show the stricter tiers
  • ESMFold2, sample 27379; segments show the stricter tiers
  • OpenDDE, sample 32276.1; segments show the stricter tiers
  • SeedFold, the task's full set, 330; its report states no count72.60; tier split not reported, drawn whole
  • Boltz-1, sample 31070.97; segments show the stricter tiers

DNA monomer

LDDTHigher is better; official n = 14 chains

  • NovaAtom-Lite, sample 140.521
  • AlphaFold 3, sample 140.530 (best value)
  • OpenDDE, sample 140.504
  • SeedFold, the task's full set, 14; its report states no count0.499
  • ESMFold2, sample 140.470
  • Boltz-1, sample 140.340

RNA monomer

LDDTHigher is better; official n = 15 chains

  • NovaAtom-Lite, sample 150.620
  • OpenDDE, sample 150.660 (best value)
  • AlphaFold 3, sample 150.610
  • SeedFold, the task's full set, 15; its report states no count0.590
  • ESMFold2, sample 150.580
  • Boltz-1, sample 150.440

NovaAtom-Lite’s values come from five seeds of five samples per target, scoring the top-ranked prediction. NovaAtom-Lite leads the two interface tasks this release is aimed at — 79.07% on antibody–antigen against 75.58% for IsoDDE, and 77.06% on protein–ligand against IsoDDE’s 75.99% — and on both the step down to the third model is the larger one: 5.6 points on antibody–antigen, 11.1 on protein–ligand. AlphaFold 3, in FoldBench’s own evaluation, reaches 47.90% and 64.90% on the same two tasks. General protein–protein interfaces are closer, 75.63% against 76.9% for OpenDDE and 76% for ESMFold2, with five models inside three points, and protein monomers are level across the field, every model here between 0.870 and 0.891 LDDT. Antibody–antigen is where the models still differ most, from 33.54% to 79.07% across the eight drawn; monomer folding is where they no longer do. The counts under the bars differ, because each report scored the targets it had, so read a gap of a point or two as a tie. The margin grows in relative terms as the bar is raised: against AlphaFold 3, NovaAtom-Lite is 1.7 times its success rate at DockQ ≥ 0.23 and 2.2 times at ≥ 0.80, where it reaches 41.28% against 19.2%. More than half the antibody–antigen interfaces it gets right are at that quality, where two in five of AlphaFold 3’s are. A bar drawn solid is a model whose report gives no split.

  • #1 on 3 tasks Antibody–antigen, protein–ligand, protein monomer
  • Top 3 on 7 / 8 In the top three on 7 of the 8 FoldBench tasks in the table below
  • +3.49 / +1.07 pts Over IsoDDE on antibody–antigen / protein–ligand, as IsoDDE’s developers report[2]

All eight tasks at a glance

FoldBench: one row per task, one column per model, NovaAtom-Lite first. Each value is the one its bar prints in the chart above. “not reported” means the source gives no value, never zero.
TaskNovaAtom-LiteIsoDDE[2]OpenDDE[4]SeedFold[5]ESMFold2[6]Protenix-v1[7]AlphaFold 3[3]Boltz-1[3]
Antibody–antigenSuccess rate (%), DockQ ≥ 0.23172 interfaces79.07 (best in row)75.5870.0, sample 16053.25352.3, sample 16047.90, sample 16733.54, sample 164
Protein–ligandSuccess rate (%), ligand RMSD < 2 Å and LDDT-PLI > 0.8558 interfaces77.06 (best in row)75.9960.163.1261, sample 51162.564.90, sample 54755.04, sample 476
Protein monomerLDDT334 chains0.891 (best in row)not reported0.890, sample 3300.8890.890, sample 3330.886, sample 3300.8800.870, sample 330
Protein–proteinSuccess rate (%), DockQ ≥ 0.23279 interfaces75.6374.1976.9 (best in row)74.0376, sample 27872.7, sample 23772.93, sample 26668.25, sample 252
Protein–RNASuccess rate (%), DockQ ≥ 0.2370 interfaces65.71not reported73.5 (best in row)65.3172, sample 39not reported62.32, sample 6956.90, sample 58
Protein–DNASuccess rate (%), DockQ ≥ 0.23330 interfaces74.24not reported76.1, sample 32272.6079, sample 273not reported79.18 (best in row), sample 31770.97, sample 310
DNA monomerLDDT14 chains0.521not reported0.5040.4990.470not reported0.530 (best in row)0.340
RNA monomerLDDT15 chains0.620not reported0.660 (best in row)0.5900.580not reported0.6100.440

Each value is the one its bar shows above. A dash means not reported, never zero. Small figures give the sample a value was computed on when it is not the task’s full set, given under each task. Bold marks the best value in each row.

Test-time scaling on FoldBench antibody–antigen

Ranking the prediction the model’s own score ranks first

Ranking: success rate by number of seeds
SeedsDockQ ≥ 0.23DockQ ≥ 0.49DockQ ≥ 0.80
174.3% (95% interval 66.9 to 81.4)65.4% (95% interval 57.0 to 73.3)41.0% (95% interval 32.6 to 48.8)
276.3% (95% interval 69.2 to 83.1)67.5% (95% interval 59.3 to 75.6)42.0% (95% interval 34.3 to 50.0)
578.3% (95% interval 71.5 to 84.9)69.5% (95% interval 61.6 to 76.7)43.4% (95% interval 34.9 to 51.2)
1079.9% (95% interval 73.3 to 86.0)70.9% (95% interval 63.4 to 77.9)44.1% (95% interval 36.0 to 52.3)
2081.2% (95% interval 75.0 to 87.2)71.8% (95% interval 64.5 to 79.1)44.8% (95% interval 37.2 to 52.9)
5082.6% (95% interval 76.2 to 88.4)73.0% (95% interval 66.3 to 79.7)46.2% (95% interval 38.4 to 54.1)
10083.1% (95% interval 77.9 to 88.4)73.8% (95% interval 67.4 to 80.2)46.5% (95% interval 39.0 to 54.1)

Oracle the best of all samples, an upper bound on ranking

Oracle: success rate by number of seeds
SeedsDockQ ≥ 0.23DockQ ≥ 0.49DockQ ≥ 0.80
177.0% (95% interval 69.8 to 83.7)68.4% (95% interval 59.9 to 76.2)47.0% (95% interval 38.4 to 55.2)
280.9% (95% interval 74.4 to 87.2)72.1% (95% interval 64.0 to 79.7)50.1% (95% interval 42.4 to 58.1)
584.1% (95% interval 77.3 to 90.1)76.0% (95% interval 68.6 to 82.6)53.5% (95% interval 45.9 to 61.0)
1085.7% (95% interval 79.7 to 90.7)78.2% (95% interval 71.5 to 84.3)55.3% (95% interval 47.7 to 62.8)
2086.5% (95% interval 80.8 to 91.3)79.8% (95% interval 73.8 to 85.5)56.3% (95% interval 48.3 to 64.0)
5087.5% (95% interval 82.0 to 92.4)81.5% (95% interval 75.0 to 87.2)57.6% (95% interval 50.0 to 64.5)
10088.4% (95% interval 83.7 to 92.4)83.1% (95% interval 77.9 to 88.4)58.1% (95% interval 51.1 to 65.1)

NovaAtom-Lite scales with test-time compute on antibody–antigen: sampling more seeds keeps making the prediction better, over the same 172 antibody–antigen interfaces throughout. Its ranked success rate rises 8.8 points, from 74.3% at one seed to 83.1% at a hundred, and the gain holds at the stricter bars — 8.4 points at DockQ ≥ 0.49 and 5.5 at ≥ 0.80, where accuracy is hardest to buy. Five seeds, the budget the charts above use, already carry 4.0 of those 8.8 points, so a small budget captures much of the benefit and a larger one keeps adding to it. At five seeds the curve is an average over random draws, 78.3%; the single run charted above scored 79.07%. The oracle panel, the best of all samples, shows how much more is within reach: it runs 2.7 points above the ranked curve at one seed and 5.3 at a hundred, and 6.0 to 11.6 at DockQ ≥ 0.80 — samples that are already better than the ranking score picks out.

Success rate on FoldBench’s 172 antibody–antigen interfaces as NovaAtom-Lite samples more seeds, 5 samples each. Each point is the mean over 500 random subsets of that many seeds out of 100; bands are 95% bootstrap intervals over interfaces (2,000 resamples).

NovaAtom-Lite on PXMeter

PXMeter[8] is an open-source toolkit for consistent, reproducible evaluation of predicted structures. Its PXM-22to25 dataset collects PDB entries released between January 1, 2022 and December 31, 2025, with low homology to earlier entries; PXM-22to25-AB is its antibody–antigen set. Success is the rate of predictions reaching DockQ ≥ 0.23 under the PXMeter protocol.

Charted: antibody–antigen interfaces in the PXM-22to25-AB set.

  • 2022–2025 PDB release window, January 1, 2022 to December 31, 2025
  • 528 antibody–antigen PDB entries across PXM-22to25
  • 383 clusters, with 1,209 antibody–antigen interfaces in all

PXM-22to25 antibody–antigen

Success rate (%), DockQ ≥ 0.23Higher is better; official n = 516 entries

  • DockQ ≥ 0.80
  • 0.49 ≤ DockQ < 0.80
  • 0.23 ≤ DockQ < 0.49
  • NovaAtom-Lite, sample 528 entries / 383 clusters66.05 (best value); segments show the stricter tiers
  • OpenDDE, sample 494 entries / 364 clusters51.0; segments show the stricter tiers
  • ESMFold2, sample 494 entries / 364 clusters42.9; segments show the stricter tiers
  • Protenix-v1, sample 516 entries / 376 clusters40.2; segments show the stricter tiers
  • Boltz-1, sample 516 entries / 376 clusters19.6; segments show the stricter tiers

The PXM-22to25 set is drawn from entries released between 2022 and 2025 with low homology to earlier ones, and it is the harder of the two antibody–antigen comparisons here: every model that appears on both scores lower on this one, by 10 to 19 points. NovaAtom-Lite reaches 66.05% against 51.0% for OpenDDE, a 15-point margin, and the ordering below them is the one FoldBench gives. AlphaFold 3 is absent because its authors report no value on this set. Counting units differ between reports — NovaAtom-Lite’s run is scored per cluster over 528 entries in 383 clusters, the baselines over 516 or 494 entries from their own reports[4] — so read the ordering and the size of the gap, not the decimal.

NovaAtom-Lite on Runs N’ Poses

Runs N’ Poses[9] tests whether co-folding methods generalize beyond ligand poses seen in training. Its 2,600 high-resolution protein–ligand systems were released in the PDB between October 6, 2021 and June 5, 2024, after the training cutoff of AlphaFold 3, Chai-1 and Boltz-1[10, 11]. We chart the 2,272 ligands that all five methods predicted, so each line is scored on the same set; success (ligand RMSD < 2 Å, LDDT-PLI > 0.8) is charted against pocket and ligand-pose similarity to the training set.

Charted: the success rate in each similarity bin, with 95% bootstrap intervals. We recomputed AlphaFold 3, Chai-1 and Boltz-1 from the authors’ published predictions[12], and scored NovaAtom-Lite and OpenDDE from our own runs.

  • 2021–2024 PDB release window, October 6, 2021 to June 5, 2024
  • 2,272 ligands all five methods predicted, of the 2,600 systems
  • 8 bins of similarity to the training set, from 0 to 100

Success by similarity to the training set

Success rate (%)Higher is better

  • NovaAtom-Lite
  • AlphaFold 3
  • Chai-1
  • Boltz-1
  • OpenDDE
Success rate (%) by similarity to the training set: one row per bin with its number of ligands; each value is followed by its 95% bootstrap interval.
SimilarityLigandsNovaAtom-LiteAlphaFold 3Chai-1Boltz-1OpenDDE
0–206347.6% (95% interval 36.5 to 60.3)23.8% (95% interval 14.3 to 34.9)9.5% (95% interval 3.2 to 17.5)12.7% (95% interval 4.8 to 20.7)15.9% (95% interval 7.9 to 25.4)
20–309140.7% (95% interval 29.7 to 50.5)26.4% (95% interval 17.6 to 35.2)17.6% (95% interval 9.9 to 25.3)7.7% (95% interval 2.2 to 13.2)19.8% (95% interval 12.1 to 28.6)
30–4014641.1% (95% interval 32.9 to 48.6)28.8% (95% interval 21.2 to 35.6)13.7% (95% interval 8.2 to 19.2)24.0% (95% interval 17.1 to 30.8)28.1% (95% interval 20.6 to 34.9)
40–5025757.2% (95% interval 51.4 to 63.4)42.4% (95% interval 36.6 to 48.2)29.6% (95% interval 23.7 to 35.0)35.4% (95% interval 29.6 to 41.2)32.3% (95% interval 26.8 to 37.7)
50–6031066.1% (95% interval 61.3 to 71.6)55.8% (95% interval 50.3 to 61.0)45.5% (95% interval 40.0 to 51.0)45.5% (95% interval 40.0 to 51.3)48.1% (95% interval 41.9 to 53.6)
60–7037681.4% (95% interval 77.1 to 85.4)67.8% (95% interval 63.3 to 72.3)58.5% (95% interval 53.5 to 63.0)54.8% (95% interval 50.0 to 59.6)64.4% (95% interval 59.8 to 69.2)
70–8036486.3% (95% interval 82.7 to 89.6)74.2% (95% interval 70.0 to 78.6)70.9% (95% interval 65.9 to 75.5)67.0% (95% interval 62.4 to 71.7)69.5% (95% interval 64.8 to 73.9)
80–10066491.0% (95% interval 88.7 to 93.2)88.2% (95% interval 86.0 to 90.5)84.9% (95% interval 82.4 to 87.8)80.9% (95% interval 77.9 to 83.7)81.2% (95% interval 78.2 to 83.9)

Each bin groups ligands by how much they resemble the training set, so the left of the chart is the harder test and the right is closer to recall. All five curves rise from left to right, but by different amounts: NovaAtom-Lite’s rises 43 points from the lowest bin to the highest, where AlphaFold 3’s rises 64, OpenDDE’s 65, Boltz-1’s 68 and Chai-1’s 75. The models converge on the right, 91.0% against 88.2% for AlphaFold 3 in the 80–100 bin, and separate on the left: 47.6% against 23.8% for AlphaFold 3 and 15.9% for OpenDDE in the 0–20 bin, and 40.7% against 26.4% and 19.8% in 20–30. Those two bins hold the fewest ligands, 63 and 91 of 2,272, which is why their bands are the widest drawn, where the three on the right hold 376, 364 and 664. Every line is scored on the same ligands, the ones all five methods predicted.

Case studies.

Six FoldBench[3] targets that NovaAtom-Lite predicts correctly and OpenDDE does not, each for a different reason. Each panel is the model’s top-ranked prediction of 25 (five seeds of five samples, ranked by its own score), superposed onto the deposited structure, which is drawn in grey.

Antibody–antigen

Superposed on the antigen, drawn as a surface. A prediction succeeds at DockQ ≥ 0.23; 8SLB’s Fab has two interfaces, scored heavy chain, then light chain, and must pass both.

Where an antibody binds, and with which loops, is what these three cases test: it is the interface where structure models still differ most, as the FoldBench results above show.

  • Deposited structure
  • NovaAtom-Lite
  • OpenDDE

  • 7ZAU Fascin-1 with nanobody 3E11Released September 27, 2023

    NovaAtom-Lite DockQ 0.805, a success
    OpenDDE DockQ 0.121, a miss

    A flat epitope, no anchor

    Fascin-1 is an actin-bundling protein of four β-trefoil domains, and 3E11 binds across several of them — 30 epitope residues in eight stretches, a flat interface with nothing to seat the nanobody in. No loop dominates either: H1, H2 and H3 make 5, 5 and 6 contacts. NovaAtom-Lite lands 2.6° from the crystal orientation, with the binding residues 1.4 Å from their place on average; 25 of 25 samples succeed. OpenDDE finds the same face but turns the nanobody 107° about the binding axis.

  • 8SLB CorA N-terminal domain with synthetic antibody C12 (Fab)Released June 7, 2023

    NovaAtom-Lite DockQ 0.710 / 0.787, a success
    OpenDDE DockQ 0.014 / 0.013, a miss

    A 23-residue loop decides

    C12 is a synthetic Fab, raised by phage display rather than by an immune system, and recognition runs through its CDR-H3 — 23 residues long, making 10 of the contacts to a 35-residue epitope. Build that loop wrong and the docking cannot be right. NovaAtom-Lite builds it to 1.9 Å, all six CDRs between 0.8 and 1.9 Å, and seats the Fab within 3 Å of the crystal pose, recovering 83% of the epitope at 97% precision. OpenDDE’s H3 is off by 8.5 Å and its Fab lands 42 Å away, at ipTM 0.70.

  • 8IDO MERS-CoV receptor-binding domain with nanobody VHH-T148Released February 28, 2024

    NovaAtom-Lite DockQ 0.839, a success
    OpenDDE DockQ 0.016, a miss

    One long loop does the work

    VHH-T148 reads the receptor-binding domain with a 22-residue CDR-H3, held in shape by an extra CDR1–CDR3 disulfide; that one loop makes 12 of the 18 epitope contacts. NovaAtom-Lite covers all 18 epitope residues, builds H3 to 0.56 Å, and sits 0.6 Å and 7.9° from the crystal pose; 25 of 25 samples succeed. OpenDDE samples the right pose once, then ranks it fourth.

Protein–ligand

Superposed on the pocket. A prediction succeeds at ligand RMSD < 2 Å and LDDT-PLI > 0.8.

  • Deposited structure
  • NovaAtom-Lite
  • OpenDDE

  • 8UN5 KRAS G13D with a switch-II pocket inhibitorReleased December 20, 2023

    NovaAtom-Lite RMSD 0.98 Å, a success
    OpenDDE RMSD 17.74 Å, a miss

    A pocket that must be opened

    The switch-II pocket is not formed in unliganded GDP-bound KRAS; it opens only as switch-II (residues 58–76) swings aside for the ligand. A model has to produce the new protein conformation and the ligand pose together, each the precondition for the other, with GDP only 4.1 Å away. NovaAtom-Lite opens it: switch-II within 0.4 to 0.6 Å of the crystal in all 25 samples, all 12 native polar contacts reproduced, no ligand atom closer than 2.2 Å to the protein. OpenDDE leaves it shut (3.3 to 4.8 Å) and its ligand is pushed onto the surface, 17.7 Å away.

  • 7FWP Human FABP4 with a furanone inhibitorReleased June 14, 2023

    NovaAtom-Lite RMSD 1.27 Å, a success
    OpenDDE RMSD 2.43 Å, a miss

    One bond sets the direction

    FABP4 encloses its ligand in a closed β-barrel. OT9 is a dumbbell — phenyl, ethyl arm, furanone core, propyl arm, methoxyphenyl — whose two ends are close enough in size that the molecule fits the cavity almost as well reversed. Only one feature fixes the direction: the furanone’s 3-hydroxyl, bonded to Arg126 and Tyr128. NovaAtom-Lite keeps every fragment on its own side in all 25 samples, 0.3 to 1.8 Å from the crystal positions — and does so while parking an extra sulfate on that very anchor. OpenDDE centres the molecule to within 0.7 Å but places the molecule end to end reversed in 23 of its 25 samples.

  • 8F8E DCAF1, WDR domain with OICR-8268Released March 1, 2023

    NovaAtom-Lite RMSD 1.48 Å, a success
    OpenDDE RMSD 5.55 Å, a miss

    Buried in the channel

    OICR-8268 sits in the central channel of DCAF1’s seven-bladed WD40 propeller, completely buried: no ligand atom is more than 4.17 Å from protein, and the 28 pocket residues arrive in ten stretches across its seven blades. With no solvent-exposed face to absorb error, and two near-identical chlorofluorophenyl ends, placement has to be exact. NovaAtom-Lite puts all eight fragments on their own positions in every one of 25 samples, 0.4 to 1.2 Å out, ligand pLDDT 89. OpenDDE builds the channel well — the pocket backbone within 0.60 Å — but never settles: 17 different arrangements across 25 samples.

Loading the 3D view…

Drag to rotate, scroll or pinch to zoom. Click a residue to see its contacts; click empty space to clear. Grey is the deposited structure.

Confidence that tracks accuracy.

NovaAtom-Lite’s per-residue confidence, pLDDT, follows its actual error closely: confident regions tend to match the experiment, and low-confidence ones are mostly regions the experiment did not resolve, or ones the prediction gets wrong.

pLDDT against observed error, NovaAtom-Lite and OpenDDE

691,945 residues from FoldBench, the same residues for NovaAtom-Lite and OpenDDE, each matched to its deposited structure

  • NovaAtom-Lite (solid bar, filled dot)
  • OpenDDE (hatched bar, hollow square)
  • Very high, above 90
  • Confident, 70 to 90
  • Low, 50 to 70
  • Very low, below 50

Not resolved in the deposited structure

A bar chart comparing NovaAtom-Lite and OpenDDE over the same 691,945 FoldBench residues, binned by each model's pLDDT: the share of residues the deposited structure does not resolve, below pLDDT 50 67% for NovaAtom-Lite and 52% for OpenDDE, at 50–70 58% and 61%, at 70–90 16% and 15%, above 90 1% for both. Colours: dark blue, pLDDT above 90; light blue, 70 to 90; yellow, 50 to 70; orange, below 50.
Model < 5050–7070–90> 90
NovaAtom-Lite 67%58%16%1%
OpenDDE 52%61%15%1%

Cα error of the residues the experiment did resolvemarker = median · whisker = 25th–75th percentile

A line chart over the same residues, NovaAtom-Lite and OpenDDE: the median Cα error of the resolved residues, below pLDDT 50 9.32 Å for NovaAtom-Lite and 13.5 Å for OpenDDE, at 50–70 3.00 Å and 4.08 Å, at 70–90 1.20 Å and 1.15 Å, above 90 0.47 Å for both. Colours: dark blue, pLDDT above 90; light blue, 70 to 90; yellow, 50 to 70; orange, below 50.
Model < 5050–7070–90> 90
NovaAtom-Lite 9.32 Å3.00 Å1.20 Å0.47 Å
OpenDDE 13.5 Å4.08 Å1.15 Å0.47 Å

Above pLDDT 70 the two models’ confidence means almost the same thing. Below 70 they part. At the same low pLDDT, the residues OpenDDE places that the experiment did resolve are further off: a median Cα error of 13.5 Å below pLDDT 50, against 9.32 Å for NovaAtom-Lite. And fewer of OpenDDE’s very-low-confidence residues are ones the experiment could not see, 52% against 67%: when NovaAtom-Lite is unsure, the region is more often one the experiment did not resolve.

The same 691,945 FoldBench residues for both models, each matched to its deposited structure; the whiskers are the 25th to 75th percentile.

The structures below are coloured by pLDDT in the standard colours: dark blue above 90, light blue from 70 to 90, yellow from 50 to 70 and orange below 50. Blue marks the regions the model is confident in; yellow and orange flag where it is not. Grey is the deposited structure.

Six cases from the test set, coloured by pLDDT

Colour: model pLDDT · grey: deposited structure · Å: median Cα error

  • Very high, above 90
  • Confident, 70 to 90
  • Low, 50 to 70
  • Very low, below 50
  • Deposited structure

Six predictions coloured by pLDDT over the deposited structure in grey: PDB 7QRJ, 9GWT, 8V91, 7UL5, 9QE8 and 9CB3. In each, the very-high-confidence regions lie on the experimental structure, with a median Cα error between 0.27 and 1.12 Å, and the very-low-confidence regions are largely residues the experiment left unresolved (from 5% to 100% of them) or, where it did resolve them, residues the prediction places several ångströms off. Colours: dark blue, pLDDT above 90; light blue, 70 to 90; yellow, 50 to 70; orange, below 50.

  • 7QRJ Zamilon virophage protein Zav_19 · trimer

    pLDDT < 50: 5% of 155 residues unresolvedthe resolved ones 13.4 Å off · pLDDT > 90: 0.27 Å off

  • 9GWT Fab 23ME-00610 · human CD200R1

    pLDDT < 50: 87% of 78 residues unresolvedthe resolved ones 17.5 Å off · pLDDT > 90: 0.53 Å off

  • 8V91 aquaporin-4 tetramer · autoantibody rAb58 Fab · cryo-EM

    pLDDT < 50: 100% of 189 residues unresolvedpLDDT > 90: 0.55 Å off

  • 7UL5 somatostatin receptor 2 (inactive) · nanobody Nb6 · cryo-EM

    pLDDT < 50: 100% of 72 residues unresolvedpLDDT > 90: 0.77 Å off

  • 9QE8 SMAD4 MH2 domain · nanobody NbA7

    pLDDT < 50: 95% of 41 residues unresolvedthe resolved ones 2.8 Å off · pLDDT > 90: 0.50 Å off

  • 9CB3 Cyclin F · E2F1 peptide

    pLDDT < 50: 100% of 108 residues unresolvedpLDDT > 90: 1.12 Å off

Six targets, each coloured by NovaAtom-Lite’s pLDDT and superposed onto the deposited structure (grey). Where it is confident the prediction lies on the experiment; where it is not, the region is one the experiment could not see, such as a flexible loop or terminus, or one the prediction gets wrong.

In practice, this is how the confidence is used: to decide which parts of a prediction to rely on, and which to check before taking a structure to the wet lab. NovaAtom-Lite also ranks its own samples; the test-time scaling figure above shows how close its top-ranked pick comes to the best of them.

References

  1. Abramson, J. et al. Accurate structure prediction of biomolecular interactions with AlphaFold 3. Nature 630, 493–500 (2024). doi.org/10.1038/s41586-024-07487-w
  2. Isomorphic Labs Team. Accurate Predictions of Novel Biomolecular Interactions with IsoDDE. Technical report, Zenodo 10.5281/zenodo.19699685 (February 2026); FoldBench values from its Table 1. doi.org/10.5281/zenodo.19699685
  3. Xu, S. et al. Benchmarking all-atom biomolecular structure prediction with FoldBench. Nature Communications (2025). doi.org/10.1038/s41467-025-67127-3
  4. OpenDDE Project, Aureka AI Research. Folding, Reasoning, and Scaling with Open-source Drug Discovery Engine (OpenDDE technical report). arXiv:2607.03787 (2026). arxiv.org/abs/2607.03787
  5. Zhou, Y., Lu, C. et al. (ByteDance Seed). SeedFold: Scaling Biomolecular Structure Prediction. arXiv:2512.24354 (2025), Table 3. arxiv.org/abs/2512.24354
  6. Candido, S. et al. Language Modeling Materializes a World Model of Protein Biology. bioRxiv 10.64898/2026.06.03.729735 (2026). www.biorxiv.org/content/10.64898/2026.06.03.729735v1
  7. ByteDance Seed. Protenix-v1: Toward High-Accuracy Open-Source Biomolecular Structure Prediction. bioRxiv 10.64898/2026.02.05.703733 (2026), Fig. 1a. www.biorxiv.org/content/10.64898/2026.02.05.703733v3
  8. Ma, W. et al. From Dataset Curation to Unified Evaluation: Revisiting Structure Prediction Benchmarks with PXMeter. bioRxiv 10.1101/2025.07.17.664878 (2025); the PXM-22to25 dataset, with the toolkit at github.com/bytedance/PXMeter. doi.org/10.1101/2025.07.17.664878
  9. Škrinjar, P. et al. Evaluating generalization in protein–ligand cofolding methods. Nature Structural & Molecular Biology 33, 782–794 (2026). doi.org/10.1038/s41594-026-01797-5
  10. Chai Discovery. Chai-1: Decoding the molecular interactions of life. bioRxiv 10.1101/2024.10.10.615955 (2024). doi.org/10.1101/2024.10.10.615955
  11. Wohlwend, J. et al. Boltz-1: Democratizing Biomolecular Interaction Modeling. bioRxiv 10.1101/2024.11.19.624167 (2024). doi.org/10.1101/2024.11.19.624167
  12. Runs N’ Poses: dataset and published predictions. Zenodo (2025). doi.org/10.5281/zenodo.14794785

Where to use NovaAtom-Lite.

From your own code through the Geodesic API, or with no code at all in NovaDDE.

In the long run, we build models like NovaAtom to make AGI improve human health.