Using AI to generate the set of shapes a molecule takes
Microsoft Research, Microsoft's research division, released a preview version of "BioEmu2," an AI that reproduces the motion of biomolecules, on Oct. 20. @aitrackerbot, an X account that tracks the appearance of models, reported the same day (the same day Japan time) that the weights (trained parameters) had been posted to an official Hugging Face repository. Frank Noé of the company's AI for Science team, who leads the research, also announced the release on X the same day.
Biomolecules such as proteins do not stay fixed in one shape; they work while constantly changing shape. From a molecule's chemical composition, BioEmu2 generates many atomic-level 3D structures the molecule can take, showing the set (ensemble) of which shapes appear and how often. It can also generate how the shape changes over time. Noé says it can reproduce, in minutes, the results of "molecular dynamics simulations," which track molecular motion by computation.
Supports complexes, small molecules and nucleic acids
BioEmu2 can handle protein complexes made of multiple chains, small molecules that are drug candidates (ligands), and nucleic acids such as DNA and RNA. It is a model using the "diffusion" technique also used in image generation, with about 500 million parameters. The model card says it was tested on NVIDIA's A100 with 80GB of memory.
Microsoft released the first "BioEmu" in 2025. However, the published model card does not show what changed from the first version. On accuracy, it says only that the model was measured on metrics such as structural deviation and the correctness of complex binding interfaces, without giving figures.
For research, not for clinical use or drug design
The license is MIT, which allows commercial use. At the same time, Microsoft positions BioEmu2 as research software, not a validated commercial tool. It asked users not to use it for clinical or patient decisions, the design of dangerous biological agents, the prediction of binding strength, or the design of new small molecules. It describes the output as plausible hypotheses, not measurements.
The model card lists its weaknesses as lower quality for proteins with little information on similar sequences, a bias toward well-studied types of proteins, and difficulty with very large complexes.
Efforts to support drug discovery and life science research with AI continue across companies. On the same day, OpenAI announced the acquisition of U.S.-based Orchestra Bio, which coordinates drug development, saying it would integrate the company into "Rosalind," its model for the life sciences.
