Researchers have built one AI model that generates molecules, materials, and protein backbones instead of needing a separate model for each domain.
Zatom-2 is a generative model pretrained on roughly five million atomic structures pulled from the OMol25 and OMat24 datasets, which cover organic molecules and inorganic materials. It combines a multiscale Transformer with conditional flow matching, and can condition generation on force data to steer outputs toward low-force or high-force regimes. The model beats its predecessor, Zatom-1, on molecular distribution fidelity and posts strong results on existing molecule and material generation benchmarks. The same pretraining also improved protein backbone design: after finetuning on just 2,000 protein domains, designability in a length extrapolation test rose from 67.8 percent to 74.8 percent.
Biology research is chronically short on structural data compared to chemistry and materials science, and most generative models are built for one domain and don't transfer. Zatom-2's results suggest pretraining on abundant molecule and material data can carry over to data-scarce biological tasks like protein design, a cheaper path to better biology models than collecting more protein data outright.
A 7-point designability bump on one benchmark is encouraging, not proof this generalizes beyond the paper's test cases.