In 2018, an artificial intelligence (AI) program called AlphaFold achieved a major breakthrough by placing first in the critical assessment of structure prediction, a competition for predicting the three-dimensional structures of proteins. It scored close to 90 on a 100-point scale for moderately difficult targets, marking a turning point in the use of AI for understanding protein structure and highlighting its potential applications. While predicting protein structures was a major step forward, proteins in living systems are not fixed. They constantly move, change shape, and interact with other molecules, and AI is now being tasked with helping with this.
Conventionally, determining the different shapes a protein takes involves fitting a known three-dimensional structure to two-dimensional high-speed atomic force microscopy (HS-AFM) images, which capture proteins in action at the single-molecule level. However, HS-AFM images are often noisy and can be distorted due to the line-by-line scanning process, where each part of the image is recorded at slightly different times. This temporal lag, along with background noise, makes it difficult to determine the exact shape of a protein at any given moment.
Associate Professor Takaharu Mori from the Department of Chemistry, Faculty of Science, Tokyo University of Science (TUS), Japan, explains the problem with current approaches. "Because of the noise present in the images, these methods can lead to overfitting, where the models may capture artefacts or false details caused by the noise rather than true structural features of the protein."
To address this challenge, his team has developed a deep learning-based method called DeepAFM, designed to both reduce noise in HS-AFM images and accurately estimate the different shapes that proteins adopt as they move and function.
The team included Mr. Katsuki Sato of TUS, who completed the Master's course in 2025, along with Dr. Takayuki Uchihashi and Dr. Yui Kanaoka from Nagoya University, Japan; and Dr. Tomoya Tsukazaki of Nara Institute of Science and Technology, Japan. The study was made available online in the Journal of Chemical Information and Modeling on April 4, 2026, and published in Volume 66, Issue 8 on April 27, 2026.
In this approach, the team creates a dataset of synthetic HS-AFM images representing different protein shapes from molecular dynamics simulations, where each image is labeled according to the corresponding protein conformation. These simulated images include both ideal, noise-free versions and more realistic ones that incorporate experimental effects, such as background noise, scanning distortions, and Brownian motion.
The researchers trained DeepAFM on a protein called SecA, which can switch between closed and wide-open states. Using molecular dynamics simulations, they generated a wide range of possible protein shapes and used them to create millions of synthetic HS-AFM images. This dataset was then used to train a deep learning model that can both remove noise from HS-AFM images and identify the underlying protein shape.
When tested, the method produced denoised images that closely matched the ground truth, with errors as low as around 0.1 nm. In addition to improving image quality, the AI was able to accurately classify the protein's conformational state. Across 0.8 million test images, the model correctly identified the exact state out of 19 possible conformations with an accuracy of 93.4%, which increased further when allowing for small tolerances. Importantly, when applied to experimental HS-AFM images, the AI inferred protein conformational states consistent with independent experimental measurements, demonstrating its practical applicability.
"DeepAFM provides a new deep learning-assisted strategy for analyzing noisy HS-AFM data and facilitates studies of protein dynamics," says Assoc. Prof. Mori.
The team further demonstrated that the method can be extended to other protein systems using transfer learning, where knowledge gained from one system is applied to another. This suggests that DeepAFM could become a broadly useful tool for studying a wide range of biological molecules.
This work is part of a broader effort to advance AI-driven research in preparation for next-generation computing platforms such as Fugaku NEXT, being developed by the RIKEN Center for Computational Science in collaboration with Fujitsu and NVIDIA, with operations expected to begin around 2030.
