Researchers present BioEmu – a new AI model that rapidly and accurately predicts the full range of shapes a protein can adopt, offering a faster, cheaper alternative to traditional molecular simulations. Proteins and their complexes are essential to nearly every biological process and are central to advances in medicine and biotechnology. While recent breakthroughs in sequencing and deep learning have made it easier to determine a protein’s sequence and structure, understanding how proteins function by shifting between different shapes in response to other molecules remains a central challenge. These dynamic shape changes underpin key biological activities. Current molecular dynamics (MD) simulation techniques provide detailed insights into protein behavior but are slow, costly, and resource-intensive. While generative machine learning models offer a faster alternative, they have yet to match experimental data reliably. Here, Sarah Lewis and colleagues developed BioEmu – a deep learning-based biomolecular emulator designed to sample the wide range of shapes a protein can adopt at equilibrium rapidly and accurately. According to Lewis et al., BioEmu can sample thousands of realistic protein conformations in just one GPU-hour, making it orders of magnitude faster and cheaper than traditional MD approaches. This is accomplished by combining training data from AlphaFold structural predictions, large-scale MD simulations, and extensive experimental measurements of protein stability, which were refined by a novel property-prediction fine-tuning (PFFT) algorithm that enables it to match experimental observations even in the absence of structural data. However, despite its efficiency and accuracy within its training domain, BioEmu has limitations. For example, it does not natively model molecular dynamics or interactions with membranes, ligands, or varying environmental conditions like temperature or pH. Nonetheless, the authors argue that BioEmu illustrates how deep learning can amortize the high cost of simulation and experimentation, paving the way toward large-scale, data-driven prediction of protein function.

More than half of people who have given birth are not using contraception two months later despite the risk that back-to-back pregnancies can pose, a new UCL study shows.

Hertz Fellow Reuben Saunders developed a powerful new CRISPRi-based screening platform to silence genes in millions of cells across living tissues. His award-winning thesis, “Pooled genetic screens with rich readouts at scale and in vivo,” is transforming how scientists study gene function in real physiological settings.

Healthcare platform moderators use strategies to manage distressing material while staying engaged enough to protect vulnerable users, finds a new study.