Microsoft(US:MSFT) 机器之心·2025-07-11 08:27Core Insights - The article discusses a groundbreaking research achievement by Microsoft's AI for Science team, which introduced a generative deep learning model named BioEmu for simulating protein conformational changes with unprecedented efficiency and accuracy [6][21]. Group 1: Research Overview - BioEmu is designed to address the limitations of existing models like AlphaFold, which typically predict static protein structures and struggle to capture dynamic conformational changes essential for protein function [8]. - The model integrates static structures from the AlphaFold database, over 200 milliseconds of molecular dynamics (MD) simulation data, and 500,000 experimental stability data points to generate thousands of independent protein structures per hour on a single GPU [8][12]. Group 2: Model Performance - BioEmu achieves a free energy prediction error of 1 kcal/mol, aligning closely with millisecond-level MD simulations and experimental data, representing several orders of magnitude acceleration compared to traditional molecular dynamics simulations [14]. - The model demonstrates excellent performance in predicting stability changes (ΔΔG) for mutants, with an average absolute error below 1 kcal/mol and a Spearman correlation coefficient exceeding 0.6 [16]. Group 3: Open Source and Future Directions - The research team has made the model parameters and code available on GitHub and HuggingFace, along with over 100 milliseconds of MD simulation data covering thousands of protein systems and tens of thousands of mutants, providing valuable resources for future research [19]. - Future plans include extending BioEmu's capabilities to more complex biological systems, such as protein complexes and protein-ligand interactions, while enhancing the model's generalization and interpretability through experimental data integration [21].