PaperlayerView exact revision on arXiv

Source-linked AI summary

The Open Molecules 2025 (OMol25) Dataset, Evaluations, and Models

Daniel S. Levine, Muhammed Shuaibi, Evan Walter Clark Spotte-Smith, Michael G. Taylor, Muhammad R. Hasyim, Kyle Michel, Ilyes Batatia, Gábor Csányi, Misko Dzamba, Peter Eastman, Nathan C. Frey, Xiang Fu, Vahe Gharakhanyan, Aditi S. Krishnapriyan, Joshua A. Rackers, Sanjeev Raja, Ammar Rizvi, Andrew S. Rosen, Zachary Ulissi, Santiago Vargas, C. Lawrence Zitnick, Samuel M. Blau, Brandon M. Wood

arXiv:2505.08762v2physics.chem-ph

TL;DR

Broad chemical coverage and practical evaluation remain limited by the cost and scope of existing DFT datasets. This paper introduces OMol25, a large-scale high-level DFT dataset with diverse chemistry, baseline models, and domain-informed evaluations; the evaluations reveal strong performance in some domains but substantial challenges for charge, spin, and long-range interactions.

  • Problem

    Existing molecular datasets lack the combined scale, chemical diversity, accuracy, and evaluation coverage needed for broadly capable ML models and practical chemistry applications.

  • Method

    The paper constructs OMol25 from diverse high-level DFT calculations and existing datasets, then evaluates baseline ML models on practical chemistry tasks.

  • Results

    Baseline performance is strongest on standard tests and some biomolecular tasks, while ionization energies, electron affinities, spin gaps, and long-range scaling remain difficult.

  • Takeaways & Limitations

    OMol25 provides a broad resource and evaluation framework for developing molecular ML models beyond conventional energy, force, and in-distribution metrics.

  • Takeaways & Limitations

    OMol25 omits radioactive elements after bismuth and polymer-material structures, while long-range models can show serious potential-energy-surface discontinuities at cutoffs.

Abstract

from arXiv · show

Machine learning (ML) models hold the promise of transforming atomic simulations by delivering quantum chemical accuracy at a fraction of the computational cost. Realization of this potential would enable high-throughout, high-accuracy molecular screening campaigns to explore vast regions of chemical space and facilitate ab initio simulations at sizes and time scales that were previously inaccessible. However, a fundamental challenge to creating ML models that perform well across molecular chemistry is the lack of comprehensive data for training. Despite substantial efforts in data generation, no large-scale molecular dataset exists that combines broad chemical diversity with a high level of accuracy. To address this gap, Meta FAIR introduces Open Molecules 2025 (OMol25), a large-scale dataset composed of more than 100 million density functional theory (DFT) calculations at the $ω$B97M-V/def2-TZVPD level of theory, representing billions of CPU core-hours of compute. OMol25 uniquely blends elemental, chemical, and structural diversity including: 83 elements, a wide-range of intra- and intermolecular interactions, explicit solvation, variable charge/spin, conformers, and reactive structures. There are ~83M unique molecular systems in OMol25 covering small molecules, biomolecules, metal complexes, and electrolytes, including structures obtained from existing datasets. OMol25 also greatly expands on the size of systems typically included in DFT datasets, with systems of up to 350 atoms. In addition to the public release of the data, we provide baseline models and a comprehensive set of model evaluations to encourage community engagement in developing the next-generation ML models for molecular chemistry.

1 Introduction

Molecular ML models need broad, high-quality data and application-relevant evaluations to serve as DFT surrogates across chemistry. OMol25 addresses these needs with a large, diverse DFT dataset and evaluation tasks spanning practical chemical objectives.

  • DFT’s approximately cubic scaling with electron number limits large-scale screening, long simulations, and systems exceeding a few hundred atoms.
  • ML interatomic potentials can approach DFT accuracy at a fraction of its computational cost, but further progress depends on large-scale, high-quality training data.
  • Existing datasets are generally small or narrow, while most larger efforts remain focused on neutral, isolated organic molecules with fewer than 50 atoms.
  • Random-split energy, force, and structure metrics do not fully assess physical-property preservation or out-of-distribution generalization for practical chemistry.
  • OMol25 provides over 140 million high-level DFT single-point calculations covering 83 elements, systems up to 350 atoms, variable charge and spin, solvation, reactivity, and intermolecular interactions.
  • The paper supplements the dataset with evaluation tasks including conformer ranking, ionization energies, spin gaps, and distance scaling.

2 Open Molecules 2025 Dataset

OMol25 combines multiple chemistry domains with diverse sampling strategies designed to represent charge, spin, conformational variation, reactivity, solvation, and intermolecular structure. Its more than 100 million calculations include broad system sizes and extensive chemical information.

  • OMol25 spans biomolecules, metal complexes, electrolytes, and main-group molecules while re-evaluating existing community datasets and derivatives.
  • At least 14 conceptually distinct data-generation methods target charge, spin, molecular conformations, and reactivity across the dataset’s domains.
  • Protein-interaction data covers protein–ligand, protein–protein, nucleic acid–nucleic acid, and protein–nucleic acid systems using extracted and simulated molecular fragments.
  • Metal-complex sampling uses specified metal centers, oxidation states, coordination numbers, ligands, and multiple generated geometries and conformations.
  • Reactive metal-complex structures are generated by substituting metals and ligands in existing datasets, then sampling AFIR reaction paths for reactive geometries.
  • Electrolyte structures are sampled from periodic-box MD across temperatures and concentrations, with ion environments extracted and selected clusters assigned added or removed electrons.
  • Intermolecular distance scaling is sampled by dilating clusters while keeping intramolecular distances fixed.
  • The final dataset contains more than 100 million DFT calculations, systems from 2 to 350 atoms, charges from -10 to +10, and spin multiplicities from 1 to 11.

3 Evaluations

The evaluations span practical molecular-chemistry behaviors beyond conventional energy, force, and structure errors, including binding, conformational, protonation, electronic-state, spin, and interaction-distance tests.

  • At least 1,000 structures are used per evaluation task to provide robust statistics across OMol25’s targeted chemistry.
  • Protein-ligand interactions: Protein-ligand binding is evaluated with ligand-pocket interaction energy and interaction-force MAEs, while full binding free-energy prediction remains outside scope.
  • Ligand strain: Ligand strain energy measures the energy difference between a bioactive-conformation local minimum and the ligand’s global minimum.
  • Conformers: Conformer evaluations compare tightly optimized DFT and MLIP conformers after RMSD-based assignment to test recovery of low-energy structures.
  • Electronic and spin properties: Protonation, electronic-state, and spin evaluations use energy and force differences to probe protonation energies, ionization energy, electron affinity, and spin surfaces.
  • Distance scaling: Distance-scaling tests evaluate whether intermolecular energy and force changes follow interaction-specific distance behavior, including 1/r, 1/r^3, and 1/r^6 scaling.

4 Baseline Models

The baseline study evaluates representative equivariant and invariant MLIPs on OMol25 validation, test, and specialized evaluation sets, while adapting them for charge and spin diversity.

  • The baseline set includes eSEN, GemNet-OC, MACE, and UMA as representative equivariant and invariant models.
  • Model architecture: Baseline MLIPs are message-passing graph neural networks that use atomic elements and 3D positions to predict energies and, depending on design, forces.
  • Charge and spin: Charge- and spin-aware inputs are required because identical 3D structures can occur with different total charges or spins and therefore different energies.
  • Charge and spin: The baselines incorporate total charge and spin through a combined embedding added to node embeddings.
  • Training: eSEN models are trained at multiple sizes, using multi-step training for medium and large models to improve efficiency at OMol25 scale.

5 Results

OMol25 models are evaluated on structure-to-energy and force prediction across diverse test splits and chemistry-focused tasks. Performance is generally strongest for standard organic and biomolecular systems, while protonation, metal charge/spin changes, and long-range interactions remain more difficult.

  • Test performance: Per-atom energy and force MAE are evaluated across All and 4M training splits using multiple test splits.The reported results compare total energy and force errors across the two training-set sizes.
  • Test performance: Conserving models outperform direct counterparts across all test splits and metrics, while larger models outperform smaller variants.GemNet-OC outperforms eSEN-sm across all splits and is comparable to eSEN-md across most metrics.
  • Test performance: 50% to 100% improved performance is observed for models trained on All versus 4M, while trends correlate well enough to support development on 4M.Averaged across all splits, UMA-M-1.1 reaches 1.38 kcal/mol energy MAE and 0.13 kcal/mol/Å force MAE on OMol-0.
  • Domain generalization: Neutral organics and biomolecules have lower errors than electrolytes and metal complexes on the out-of-distribution composition test.Table 4 reports total energy and force MAE across biomolecules, electrolytes, metal complexes, and neutral organics for All and 4M training splits.
  • Optimization evaluations: Ligand-strain and conformer evaluations remain within chemical accuracy, with ligand-strain errors of 0.19 to 0.07 kcal/mol and best conformer errors below 0.1 kcal/mol.Models trained on 4M show only small degradation for ligand strain and remain within chemical accuracy for conformer ranking.
  • Single-point evaluations: Protonation is the hardest optimization evaluation by energy, with differences from 0.55 to about 1 kcal/mol, while metal charge/spin and long-range interaction tasks remain severe challenges.Metal IE/EA and spin-gap errors are high, and distance-scaling errors increase dramatically beyond model cutoffs, producing cutoff-related PES discontinuities.

6 Outlook and future directions

OMol25 broadens molecular DFT data and evaluation while identifying important coverage and accuracy gaps that future MLIP development must address.

  • OMol25 spans major chemistry domains at high theoretical level, but excludes radioactive elements after bismuth and polymer-material structures.
  • OMol25 coverage is relatively limited for lanthanide complexes, multimetallic structures, and solvated protonated organic molecules and metal complexes.
  • Baseline OMol25 models achieve average energy and force errors as low as 1.38 kcal/mol and 1.07 kcal/mol across all test splits.These results are reported for UMA-M-1.1 and UMA-S-1.2, respectively.
  • Evaluation tasks still find errors as high as 4-9 kcal/mol for ionization energies/electron affinity, spin-gap, and long-range scaling.The authors identify charge, spin, and long-range-interaction architecture as especially important areas for improvement.
  • A public leaderboard was released to support community engagement and rapid method development on OMol25 evaluation tasks.

A.2 Computed Properties

OMol25 records a broad set of electronic-structure outputs for each calculated molecular system, including energies, densities, charges, spins, convergence information, and associated files.

  • Total energy is recorded in eV for each calculated molecular system.
  • The dataset includes ECP-electron counts, basis-function counts, and whether calculations used unrestricted or restricted treatment.
  • SCF convergence information includes the number of SCF steps, VV10 energy, and deviation of S2 from its ideal value.
  • Integrated density is recorded and should remain very close to the total number of electrons.
  • The properties include maximum per-atom force magnitude in a given direction, Mulliken charges and spins, and Loewdin charges and spins when unrestricted.
  • NBO charges and spins are included for systems with at most 70 atoms, alongside ORCA warnings and planned releases of ORCA .gbw files and densities.
  • Additional properties such as multipole moments, Fock matrices, and orbital energies are planned for future dataset versions.

B OMol Dataset Versions

OMol-1 expanded the pre-release OMol-0 dataset with 40M additional DFT datapoints and broader coverage of metal complexes, electrolytes, and main-group molecules. The curation pipelines also prepared diverse protein–ligand systems through deduplication, protonation-state enumeration, pocket extraction, and molecular-dynamics sampling.

  • Protein–ligand systems: 258,467 unique ligand-pocket pairs were elaborated with up to 10 low-energy protonation/tautomer states and receptor-residue optimization.The extracted systems included the ligand, identified receptor residues, and nearby residues within 2.5Å, including coordinated waters and metals.
  • Protein–ligand systems: 416,324 extracted protein pockets remained after physically consistent charge and spin assignment and restricting systems to 350 atoms or fewer.Charge and spin were assigned using formal atomic charges and determined metal oxidation states.
  • Protein–ligand systems: 100ps Brownian dynamics, 12ps restrained Langevin dynamics, and subsequent unrestrained molecular dynamics were used to increase structural diversity in extracted pockets.The protocol began at 10K with heavy atoms restrained before restraints were removed.
  • Protein–ligand systems: 750k experimental ligand–protein datapoints were exhausted, motivating synthetic ligand–protein structures to broaden interaction sampling and docking-relevant coverage.The synthetic workflow identified 32,833 likely druggable pockets and docked drug-like ligands into them.

D.2 Complexes from the Crystallography Open Database (COD)

The COD-derived evaluation and electrolyte components extend OMol25 beyond isolated molecular structures. COD metal complexes were filtered and assigned electronic states, while electrolyte data combine literature and randomized formulations with molecular-dynamics sampling and deliberately broad configuration diversity.

  • COD complexes: COD systems were discarded when they had multiple metal centers, molecular charge magnitude above 2, actinides, more than 250 atoms, colliding atoms, or implausible bonding patterns.Multimetallic complexes were excluded because accurate oxidation-state assignment would require substantial manual intervention.
  • COD complexes: 70,004 COD structures passed filters for trustworthy structure and electronic assignment, yielding 90,359 DFT inputs across low- and high-spin configurations.Second- and third-row complexes received low-spin configurations; first-row complexes were calculated in highest and lowest reasonable spin states.
  • Electrolytes: Electrolyte sampling uses classical molecular dynamics because isolated molecules cannot capture dynamic properties such as transport, (de)solvation, and wetting.The resulting training structures include clusters of noncovalently bound ions and solvent molecules.
  • Electrolytes: Literature electrolyte systems were supplemented with random mixtures spanning protic, aprotic, ionic-liquid, aqueous, and molten-salt classes.The class probabilities were 40% protic, 40% aprotic, 10% ionic liquid, 5% molten salt, and 5% aqueous.
  • ML-sampled structures: 1.6M metal-complex, 1.3M electrolyte, and 1.1M biomolecular structures were calculated with DFT from ML-sampled configurations.Chemically unreasonable structures generated along trajectories were retained as useful examples of unstable regions of the potential-energy surface.

G.1 Interpolated Reactivity Datasets

OMol25 increases reactive-geometry coverage using interpolated and force-guided paths across organic, main-group, and other molecular systems. The procedures generate configurations near bond breaking and formation, while filtering and subsampling control redundancy and structural validity.

  • Main-group diversity: Random substitutions introduced Se/Te, P/As/Sb, and Si/Ge into common O-, N-, and C-containing motifs, followed by charge and spin-state sampling.The procedure includes +1, -1, +2, and neutral triplet states, as well as singlet and triplet states for +2 species.
  • Reactive geometries: Four distinct approaches were used to generate reactive geometries for OMol25.The approaches combine interpolations, force-guided paths, and chemically diverse substitutions across several source datasets.
  • Interpolated datasets: 126,857 RGD1 reactant–product–transition-state triplets were expanded by geodesic interpolation with ten configurations on each reactant-to-transition-state and transition-state-to-product path.RGD1 contains neutral, closed-shell species composed of C, H, O, and N with up to ten heavy atoms.
  • AFIR sampling: AFIR paths push atoms involved in forming bonds together and pull apart atoms in breaking bonds, producing optimization trajectories that are subsampled for DFT.MACE-MP-0 supplies energies and forces, while assigned charge and spin affect the subsequent DFT calculations.
  • AFIR sampling: Trajectory frames are selected using force magnitude, neighboring-frame energy changes, and minimum RMSD to favor reactive and nonredundant structures.Selection begins with the highest-energy trajectory structure.

H.3 Metal Reactivity

Metal-reactivity data were generated by swapping metals and selected ligands in curated reaction templates, then producing reactive paths and DFT snapshots. The workflow reserves ligand-incompatible reactions for testing and uses Popcornn for electrolyte reactions when other potentials fail.

  • Template preparation: 125 template reactions remained after manual inspection and validation of 137 MOR41, ROST61, and MOBH35 reactions.Two additional reactions were obtained by manually relocating I2 species in MOR36 and MOR23.
  • Reaction enumeration: Ligand swaps matched charge state and denticity, changed one ligand at a time, and excluded ligands involved in bond breaking or forming.Only monodentate and bidentate ligands were swapped to reduce overlap and steric clashes.
  • Reaction enumeration: 99 templates contained swappable ligands, whereas 34 did not; the resulting 1,436 reactions without swappable ligands were reserved as a test set.The ligand-swappable templates produced approximately 1.1M possible reaction templates.
  • Reaction sampling: 146,962 reactions were obtained after metal swapping, and approximately 250k reactions were targeted for roughly 5M AFIR snapshots.Sampling was selected to distribute coverage as evenly as possible across reactions within the computational budget.
  • Electrolyte reactivity: MACE-MP-0 and B97-3c produced poor electrolyte-reaction snapshots with undercoordinated metals and radicals, so Popcornn generated optimized paths instead.Seventeen snapshots were sampled uniformly from each optimized path, with the middle 15 selected for simulation.
  • Out-of-distribution testing: Two explicitly out-of-distribution reactive test sets contained geodesic interpolations for 1,782 organic reactions and AFIR snapshots for 1,436 metal-template reactions.Together, these sets contained approximately 67k snapshots.

I Community Datasets

Existing datasets were recalculated at the OMol level, while solvated protein fragments were diversified by charge and lightly optimized. Additional filtering was applied to remove unsuitable ANI-1xBB structures.

  • SPICE, Transition-1x, ANI-2x, and OrbNet Denali were recalculated at the OMol level without modifications.
  • 10% of solvated protein fragments received an electron, 10% lost one, and 80% remained unchanged.
  • Solvated protein fragments were optimized with Sella for up to two geometry optimization steps.

J Evaluation Details

The evaluations measure model behavior across conformers, protonation, charge, spin, distance scaling, and protein–ligand systems. They use standardized optimization, matching, sampling, and baseline-model procedures.

  • Drug-like protein–ligand pockets were selected using QED greater than 0.5 and 3–12 rotatable bonds.
  • Conformer evaluations: Ligand strain and conformer evaluations retain five lowest-energy and five randomly selected conformers after xTB deduplication and DFT optimization.Conformers are deduplicated using a 0.25 Å maximum corresponding-atom distance and 1 kcal/mol energy tolerance.
  • Conformer evaluations: Boltzmann-weighted ensemble RMSD computes the dot product of DFT Boltzmann weights and RMSDs from the conformer mapping.Because the Boltzmann weights sum to 1, no further averaging is required.
  • Specialized evaluations: Charge, spin, and distance-scaling evaluations compare energies, forces, or optimized geometries across altered molecular states and separations.Distance scans identify short-range and long-range structures using a neighbor graph with a 6 Å cutoff.
  • Baseline models: eSEN, GemNet-OC, and charge-neutral MACE provide baseline models, with eSEN evaluated in small, medium, and large variants.All models receive architectural embeddings for total charge and spin to support comparisons on OMol25.

K.2 Total Energy Test Results

Total-energy results are reported for the test splits, together with corresponding breakdown results in Tables 3 and 4.

  • Tables 3 and 4 report total-energy test results and their corresponding breakdowns.

K.3 Validation Results

Validation results for the out-of-distribution composition set are reported in Table 13.

  • Table 13 reports results on the out-of-distribution composition validation set.

K.4 Additional Results

Additional evaluations test OMol25-trained baselines on out-of-distribution compositions, structures, and strained conformers. The baselines achieve errors no greater than 1 kcal/mol against the Wiggle150 reference theory, while model and training configurations are documented for reproducibility.

  • Additional test splits: Out-of-distribution evaluations cover TorsionNet500, curated cations, solvents, and ions-plus-solvents, measuring energy and force prediction.These results are reported across additional test splits.
  • Composition validation: Composition validation reports total-energy and force mean absolute errors for biomolecules, electrolytes, metal complexes, and neutral organics across All and 4M training splits.The results are broken down by molecular category and training-set size.
  • Additional test splits: Energy-per-atom and force mean absolute errors are reported across the All and 4M training splits for additional structure-to-energy-and-force test sets.These metrics complement the composition-validation results with further structure-based evaluations.
  • Wiggle150: ≤1 kcal/mol errors were obtained by all OMol25 baseline models against the Wiggle150 reference theory.The original Wiggle150 study reported this accuracy only for ωB97M-V and two double-hybrid functionals.
  • Training configurations: GemNet-OC was trained with both 12 and 6 Å cutoff configurations, with the latter matching eSEN-sm+md; hyperparameters for eSEN, MACE, and GemNet-OC are provided.The All-set GemNet-OC training required doubling the batch size relative to its 4M variant.
Loading 2505.08762v2…