Chemistry & ADMET · Berkeley Lab, Cambridge and collaborators
Zatom-2
Unified atomistic model generating 3D molecules, materials, and proteins with energy and force predictions.
What Zatom-2 is
Zatom-2 is a unified atomistic foundation model developed by the Zatom-AI collaboration, with work at Lawrence Berkeley National Laboratory. While conventional generative models in computational chemistry and structural biology are tailored exclusively to individual domains—such as small drug molecules, crystalline materials, or macro-biomolecules—Zatom-2 is pretrained across all three domains simultaneously under the principle that physical atomic interactions are universal.
The model performs both generative modelling and representation learning. In its small-molecule setting, it samples three-dimensional coordinates based on specified atomic counts, overall molecular charges, and spin multiplicities. Released weights also include property prediction heads, allowing researchers to evaluate conformational strain, system stability, and atomic forces directly alongside coordinate sampling without immediately invoking external quantum-mechanical solvers.
How it works
Zatom-2 uses a multiscale transformer architecture integrated with conditional flow matching. By parameterising velocity fields over continuous spatial coordinates, the flow-matching framework learns to transport random distributions into physically realistic atomic configurations. The model incorporates force conditioning and supports multitask training objectives, including structure generation and interatomic machine-learned potential predictions across diverse systems.
For molecular generation, the network is trained on the OMol25 distribution of density functional theory (DFT) relaxed small molecules. Samples are generated according to input parameters defining the number of atoms (including hydrogens), total formal charge, and spin multiplicity. Property heads trained jointly with the generative backbone infer total system energy in electronvolts and calculate residual force vectors on every atom.
Key facts
- Architecture
- Multiscale transformer with conditional flow matching
- Tasks
- Molecular generation, periodic material design, protein transfer, interatomic potential prediction
- Supported modalities
- Small molecules, periodic materials, proteins
- Pretraining benchmarks
- OMol25, OMat24, SCOPe
- Generation inputs
- Atom count, sample count, total charge, spin multiplicity
- Outputs
- SMILES, 3D structures, predicted energy (eV), atomic forces
- GPU memory
- 24 GB
What you can use it for
- Unconditional small-molecule generation
- Sample 3D molecular conformations based on defined atom numbers, charges, and spin multiplicities.
- Interatomic energy and force estimation
- Evaluate structural stability and identify conformational strain in generated structures using joint property prediction heads.
- Periodic material exploration
- Design crystalline materials and evaluate generative fidelity against benchmark distributions such as OMat24.
- Protein structure sampling
- Explore protein design tasks using specialised transfer checkpoints pretrained on SCOPe domain representations.
Run it here
No installation and no GPU of your own: fill in the inputs and it runs on our servers. Results stay in your browser unless you sign in.
Runs on On-demand GPU
Runs on our servers with on-demand compute. A first run needs time to load the model; active capacity can be reused and scales down when idle.Run this model
Zatom-2
Strengths and limits
Good at
- Unifies molecular, crystalline, and macromolecular generative tasks within a shared atomistic architecture.
- Reports estimated total energy and atomic forces jointly with generated coordinates to indicate structural relaxation.
- Permits conditioning on explicit physical boundary constraints, including net molecular charge and spin multiplicity.
Watch out for
- Generated molecules are drawn from the OMol25 distribution and are not pre-filtered for synthetic accessibility, novelty, or drug-likeness.
- Coordinate generation may place disconnected fragments, such as isolated water molecules, alongside the target structure.
- Sampling does not always resolve into valid bonded chemical graphs, with success dropping as molecular size increases (tested at three in four valid molecules at 24 atoms and two in four at 40 atoms).
- Predicted energy values are machine-learning estimates rather than density functional calculations, and they cannot be compared across molecules with different atomic compositions.
Inputs and outputs
You provide
Atoms (including hydrogens)optional
The published molecular benchmarks sit around 40.
Molecules to drawoptional
Total chargeoptional
Spin multiplicityoptional
1 for a closed-shell molecule, 2 for a radical.
You get back
Each sample as SMILES with its atom count, the model's predicted energy in eV, and the largest force on any atom.
The first sample that resolved into a molecule, in 3D.
Licence
Commercial use permitted · BSD-3-Clause code / MIT weights
Verified 2026-10-29. Code: BSD-3-Clause, github.com/Zatom-AI/nucleus/blob/main/LICENSE.md (Regents of the University of California, through Lawrence Berkeley National Laboratory), carried through to the `zatom-nucleus` and `zatom2` PyPI packages. Weights: huggingface.co/Zatom-AI/Zatom-2 declares MIT on its model card and is not gated. Its THIRD_PARTY_NOTICES is explicit that the root licence does not override dataset terms; the training sets it names are OMol25 and OMat24 (CC BY 4.0), QM9 and GEOM (CC0), LeMat-Bulk (CC BY 4.0) and SCOPe/PDB — none non-commercial — but the MIT weights carry no clause inheriting them. Dependencies: atomworks is BSD-3-Clause, rc-foundry is BSD-3-Clause covering code and weights, and NVIDIA's cuequivariance runtime kernels ship under NVIDIA's own software licence, in the same position as CUDA itself.
Questions
Can I use Zatom-2 commercially?
Commercial use permitted. Verified 2026-10-29. Code: BSD-3-Clause, github.com/Zatom-AI/nucleus/blob/main/LICENSE.md (Regents of the University of California, through Lawrence Berkeley National Laboratory), carried through to the `zatom-nucleus` and `zatom2` PyPI packages. Weights: huggingface.co/Zatom-AI/Zatom-2 declares MIT on its model card and is not gated. Its THIRD_PARTY_NOTICES is explicit that the root licence does not override dataset terms; the training sets it names are OMol25 and OMat24 (CC BY 4.0), QM9 and GEOM (CC0), LeMat-Bulk (CC BY 4.0) and SCOPe/PDB — none non-commercial — but the MIT weights carry no clause inheriting them. Dependencies: atomworks is BSD-3-Clause, rc-foundry is BSD-3-Clause covering code and weights, and NVIDIA's cuequivariance runtime kernels ship under NVIDIA's own software licence, in the same position as CUDA itself.
How do I run Zatom-2 without installing anything?
Use the form under “Run it here” on this page: fill in the inputs and press Run model. It runs on our servers on 1× A100 40GB, typically in 1.2min, and a run costs about 28 credits. Visitors without an account get 1,000 credits a month.
How are energy and force values calculated for generated molecules?
Energies and forces are predicted directly by the model's internal property heads rather than external DFT simulations. They reflect the model's estimate of interatomic potentials and are strictly comparable only across conformations sharing identical atomic compositions.
What inputs are required to generate small molecules?
Users must specify the total number of atoms including hydrogens, the desired number of samples, the net molecular charge, and the spin multiplicity.
Why do some generated samples contain separate fragments?
Because the unconditional sampling process places atoms in continuous space to satisfy a fixed atom count, it can occasionally produce detached components, such as a stray water molecule, alongside the primary structure.
Does the model guarantee valid molecular graphs for all samples?
No. Not every sampled conformation successfully resolves into a chemically bonded structure, and completion rates decline as the requested atom count grows. Generating extra samples is recommended to ensure sufficient valid structures.
How to cite it
Miruna Cretu, Alex Abrudan, Antonia Panescu, Tynan Perez, Rishabh Anand, N. Benjamin Erichson, Michael W. Mahoney, Samuel Blau, Joseph Jacobson, Rafael Gómez-Bombarelli, Rex Ying, Tuomas Knowles, Pietro Liò, Alex Morehead. Zatom-2: Multitask Pretraining on Atomistic Data for Generative Modeling across Domains. arXiv preprint arXiv:2610.11454, 2026.
Written from Zatom-2's own documentation (code, paper). Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.