DeepAFM: Bridging the Gap Between Static Protein Structures and Dynamic Motion

The field of structural biology has been transformed by artificial intelligence, most notably by the 2018 breakthrough of AlphaFold. By predicting the three-dimensional structures of proteins with remarkable precision, AlphaFold provided a foundational map for researchers. However, living systems are defined by movement; proteins do not sit still. They constantly shift, fold and interact with other molecules in a complex dance that remains difficult to capture.

A new deep learning-based method, DeepAFM, is now addressing this challenge by decoding protein motion from high-speed atomic force microscopy (HS-AFM) images. Developed by a team of researchers from the Tokyo University of Science, Nagoya University, and the Nara Institute of Science and Technology, this approach offers a clearer lens into the dynamic behavior of biological molecules.

The Challenge of Nanoscale Imaging

Traditionally, scientists use HS-AFM to visualize proteins in action at the single-molecule level. While this technique provides high-resolution data, it is inherently noisy. The scanning process occurs line-by-line, creating a temporal lag that, combined with environmental interference, makes it difficult to distinguish true protein shapes from artifacts caused by background noise.

According to Associate Professor Takaharu Mori of the Tokyo University of Science, conventional methods are prone to overfitting. When models rely too heavily on noisy data, they often capture false details rather than the authentic structural features of the protein. This limitation has historically hindered the ability to track transitions between different conformational states accurately.

How DeepAFM Works

To overcome these hurdles, the research team—whose study was published in the Journal of Chemical Information and Modeling (Volume 66, Issue 8, April 27, 2026)—developed DeepAFM. The method utilizes a synthetic dataset derived from molecular dynamics simulations. By simulating a protein’s movement, the researchers generated millions of images that account for experimental realities, including Brownian motion, scanning distortions, and background noise.

The model was trained on the SecA protein, which is known to switch between closed and wide-open states. The results are significant:

• Denoising Accuracy: The AI produces images with errors as low as approximately 0.1 nm.
• Classification Performance: In tests involving 0.8 million images, the model identified the correct state from 19 possible conformations with 93.4% accuracy.
• Practical Application: When applied to experimental HS-AFM data, the AI inferred conformational states that aligned with independent measurements, confirming its reliability for real-world research.

Future Implications for AI-Driven Science

Beyond its immediate utility, DeepAFM demonstrates the potential of transfer learning, allowing the model to be adapted to various other protein systems. This versatility positions DeepAFM as a foundational tool for future biological studies. The development of such high-precision analysis tools is a critical component of the broader movement toward advanced AI-driven research, aligning with the development of next-generation computing platforms like Fugaku NEXT, which are expected to begin operations around 2030.

Key Takeaways

• Beyond AlphaFold: While structural prediction has been mastered, DeepAFM tackles the next frontier: observing how those structures move and change in real-time.
• Noise Reduction: The deep learning model effectively filters out scanning distortions and background noise inherent in atomic force microscopy.
• Versatility: Through transfer learning, the methodology can be extended to analyze a wide range of biological molecules, not just the protein systems used in initial training.

By effectively bridging the gap between noisy experimental data and the dynamic reality of molecular behavior, DeepAFM represents a significant step forward in our ability to understand the machinery of life at the nanoscale.