Researchers from Tokyo University of Science and partner institutions in Japan have developed a new artificial intelligence method called DeepAFM to decode protein motion from noisy experimental data. Published in the Journal of Chemical Information and Modeling, this study introduces a deep learning approach that achieves 93.4% accuracy in identifying protein shapes within high-speed atomic force microscopy (HS-AFM) images. HS-AFM is a powerful tool for observing proteins at the single-molecule level, capturing their movement as they change shape to perform biological functions. However, the technique suffers from significant limitations. The line-by-line scanning process creates temporal lags, resulting in distorted images plagued by background noise. Conventional analysis methods often struggle with these imperfections, sometimes leading to overfitting where models mistake artifacts for actual structural features. To overcome these barriers, Associate Professor Takaharu Mori and his team engineered DeepAFM. The method utilizes a synthetic dataset generated from molecular dynamics simulations of the SecA protein, which switches between closed and wide-open states. These simulations produced millions of images that included both ideal, noise-free versions and realistic variants incorporating experimental distortions and Brownian motion. By training the deep learning model on this diverse dataset, the researchers enabled it to simultaneously remove noise and classify the specific conformational state of a protein. In testing across 0.8 million images, the model successfully identified the exact state out of 19 possible conformations with 93.4% accuracy. The error margin for denoised images was as low as 0.1 nanometers. Furthermore, when applied to real-world experimental data, the AI's inferences regarding protein states aligned with independent experimental measurements, validating its practical utility. The team also demonstrated that the model can be adapted to other protein systems using transfer learning, suggesting its potential as a broadly useful tool for studying various biological molecules. This breakthrough builds upon the momentum of AI in structural biology, following the 2018 success of AlphaFold in predicting static protein structures. While AlphaFold revolutionized the understanding of protein architecture, it did not address the dynamic nature of proteins in living systems. DeepAFM fills this gap by enabling the analysis of protein movement and transitions, offering a new strategy for studying protein dynamics. The research team includes Mr. Katsuki Sato from Tokyo University of Science, along with Dr. Takayuki Uchihashi and Dr. Yui Kanaoka from Nagoya University, and Dr. Tomoya Tsukazaki from the Nara Institute of Science and Technology. Their work contributes to a broader initiative to advance AI-driven scientific research in preparation for next-generation supercomputing platforms, such as Fugaku NEXT, which is being developed by the RIKEN Center for Computational Science in collaboration with Fujitsu and NVIDIA for expected operation around 2030.