Microsoft AI predicts protein conformations

The way proteins move is often what gives them biological activity: they can open, close, twist, and rearrange themselves in ways that allow them to bind to other molecules or perform complex functions. Now, an artificial intelligence tool developed by Microsoft researchers can predict the multiple conformational states of proteins in minutes with a fraction of the resources required by other techniques. The tool, BioEmu, is completely open source—meaning that its associated source code, training data, model weights have been released to the public (Science 2025, DOI: 10.1126/science.adv9817).

Frank Noé, a multidisciplinary researcher at Microsoft Research’s AI for Science lab led the work on BioEmu. He says that tools like Google DeepMind’s Nobel Prize–winning AlphaFold helped kick off the “structure revolution,” which gave scientists access to roughly 200 million computationally predicted protein structures. But those structures were static.

To understand the movement of proteins, scientists have used experiments such as cryogenic electron microscopy to capture snapshots of the movements and molecular dynamics simulations to figure out how proteins move. But these experiments can take an enormous amount of time, money, and computational power.

BioEmu is based on a deep learning neural network that Noé says was trained on such costly simulated and experimental data. He says some of that training data was produced by other labs and publicly available, and some training data was generated in-house.

The training gives BioEmu the ability to make “predictions of how proteins move and about the structures they can assume in equilibrium,” Noé says. “It is the first tool that makes quantitative predictions about the relative probabilities of different states that proteins can have that underlie their function.”

Instead of simulating how a protein moves, BioEmu emulates it, according to Noé. “Emulators behave like simulators, but they’re much faster,” he says. “They use deep learning tricks to kind of shortcut some of the very expensive numerical calculations.” BioEmu can predict the many conformational states of a protein and its free energy—a measure of a protein’s stability—with accuracy of 1 kcal/mol, which Noé describes as “experimental accuracy.”

There exists a lot of potential utility for BioEmu in drug discovery, Noé says. The tool can help researchers find cryptic binding pockets, which could be inhibited by small molecules to reduce the activity of disease-causing proteins. But it will take time to see if that hypothetical plays out.

Zhidian Zhang, a postdoctoral computational biologist at the Massachusetts Institute for Technology who wasn’t involved in the research, has lots of praise for BioEmu and the associated paper describing it. Several groups over the past couple of years have attempted to predict alternative conformational states of proteins, she says. But BioEmu is different in that it can predict “the distribution of different conformations, which is a much more difficult problem.”

Alberto Perez, a computational chemist at the University of Florida who also wasn’t involved in the research, says that he can see himself using BioEmu in his own work and is thrilled that it is being released open source. “There was a lot of blowback when AlphaFold 3 didn’t publish their code, so I’m happy to see like this code will be available and accessible,” he says.

Zhang and Perez both note that BioEmu has limitations. The algorithm works well for small proteins but struggles with larger ones, as well as with proteins that bind to ligands and membrane proteins, which are especially important as potential drug targets. It also doesn’t predict the kinetics of protein movement—how long it takes for a protein to move from one conformation to the other and in what order. The authors of the paper also acknowledge these limitations.

Yet Zhang, Perez, and Noé also think that BioEmu’s capabilities will improve over time. “When AlphaFold first came out it was big news, but it wasn’t game changing,” Perez says. “The second time around, it was game changing.” Only time will tell if the next iteration of BioEmu will be game changing as well.