DeepAFM maps protein movements 93.4%

Tokyo University of Science researchers said this week they developed an artificial intelligence system called DeepAFM that reconstructs protein conformational states from noisy high-speed atomic force microscopy, or HS-AFM, images. The method combines molecular dynamics simulations with deep learning to denoise AFM footage and classify the shapes proteins adopt while moving, according to a paper published in the *Journal of Chemical Information and Modeling* on April 4. (t.co) The study focused on SecA, a motor protein involved in protein translocation, within the SecYAEG–nanodisc complex. (tus.ac.jp) The authors said the approach is designed to improve analysis of dynamic protein behavior from AFM movies, which often contain substantial noise and scan-related distortion. ### Why did the researchers build DeepAFM instead of using standard fitting methods? High-speed AFM can record proteins in motion at the single-molecule level, but the images are noisy and have limited spatial resolution, the paper said. Conventional analysis typically fits known three-dimensional structures to two-dimensional AFM frames, a process the authors said can be thrown off by background noise and the time lag introduced as the microscope scans line by line. Takaharu Mori, an associate professor at Tokyo University of Science and the paper's corresponding author, said existing methods can overfit image artefacts rather than true structural features. (t.co) The team built DeepAFM to address that problem by learning from simulated AFM images generated from molecular dynamics snapshots under conditions meant to mimic experimental noise, Brownian motion and scan distortions. ### How does DeepAFM turn a noisy microscope movie into protein states? The model uses simulated HS-AFM images as training data, with each image labeled to a corresponding protein conformation, the university and paper said. (t.co) Those simulations include both idealized images and more realistic ones with experimental effects added, allowing the system to learn denoising and state estimation together rather than as separate steps. GitHub materials released by the Mori laboratory describe DeepAFM as a Vision Transformer-based multitask autoencoder. (tus.ac.jp) The repository says the framework supports AFM image generation from molecular dynamics trajectories, denoising, conformational state estimation and transfer learning to other protein systems. ### What did the paper say the system achieved on SecA data? The paper reported that DeepAFM classified protein conformational states with more than 90% accuracy on simulated test images under low white-noise conditions and was more robust to image distortion than conventional rigid-body fitting. (tus.ac.jp) A Tokyo University of Science release and a Phys.org report on the paper put the headline figure at 93.4% classification accuracy. The authors said the trained model focused on image regions associated with SecA's large-scale domain motions, which improved resistance to noise-induced overfitting. (github.com) On experimental HS-AFM images, the paper said DeepAFM estimated dominant conformational states in a way that agreed with independent experimental observations. ### Which researchers and institutions were involved? The article lists Katsuki Sato, Yui Kanaoka, Tomoya Tsukazaki, Takayuki Uchihashi and Takaharu Mori as authors. Tokyo University of Science said Sato completed a master's course in 2025, while Kanaoka and Uchihashi are affiliated with Nagoya University and Tsukazaki with Nara Institute of Science and Technology. (t.co) The journal publication places the work in Volume 66, Issue 8 of the *Journal of Chemical Information and Modeling*. The university said the article was first made available online on April 4, 2026, and then appeared in the April 27 issue. (t.co) ### Where can readers find the paper and code next? The American Chemical Society journal page describes the study as an open-access article under a CC-BY 4.0 license. The Mori laboratory has also posted an official DeepAFM code repository on GitHub with setup instructions, pretrained weights and tutorial notebooks for transfer learning on additional protein systems, including MgtE and HECT. (t.co) GitHub instructions say users can download a dataset archive from Zenodo and run tutorial notebooks in JupyterLab after installing the Python requirements. The repository was public when accessed on May 14, 2026. (tus.ac.jp) (github.com) (t.co)