Key Takeaways
- DeepAFM, a deep learning model, improves the accuracy of protein shape determination from high-speed atomic force microscopy images.
- The method achieved 93.4% accuracy in classifying protein conformations and effectively reduces noise in imaging.
- DeepAFM shows potential for broader applications in studying various biological molecules.
Advancements in Protein Structure Analysis
In 2018, AlphaFold, an artificial intelligence (AI) program, revolutionized the prediction of protein structures by scoring nearly 90 out of 100 in a structural prediction competition. This breakthrough marked a significant milestone in utilizing AI for understanding protein architecture. However, predicting static protein structures is only part of the challenge. Proteins in living systems are dynamic, constantly changing their shapes and engaging with other molecules. To address this complexity, AI applications in protein dynamics are emerging.
Traditionally, evaluating the various conformations of proteins involves aligning known three-dimensional structures with two-dimensional high-speed atomic force microscopy (HS-AFM) images. These images capture the movement of proteins at the single-molecule level. However, the process often produces noisy images, compounded by temporal lag from line-by-line scanning, rendering accurate shape determination difficult.
Associate Professor Takaharu Mori from the Tokyo University of Science (TUS) highlights that the noise in HS-AFM images can result in overfitting, where models may inaccurately reflect protein structures due to capturing artifacts rather than genuine features.
To combat this issue, Mori’s research team developed DeepAFM, a novel deep learning method aimed at reducing noise in HS-AFM images while accurately representing the various protein shapes during their functionality. The team comprised Mr. Katsuki Sato from TUS, Dr. Takayuki Uchihashi, Dr. Yui Kanaoka from Nagoya University, and Dr. Tomoya Tsukazaki from Nara Institute of Science and Technology.
The research, published in the Journal of Chemical Information and Modeling, details a dataset created from synthetic HS-AFM images derived from molecular dynamics simulations. Each image is labeled to correspond with specific protein conformations. This dataset includes ideal, noise-free images and realistic variations incorporating experimental noise and scanning distortions.
The researchers trained DeepAFM on the SecA protein, which exists in both closed and wide-open states. Through extensive molecular dynamics simulations, they generated millions of synthetic HS-AFM images representing different protein structures. The model demonstrated superb noise reduction, with denoised images closely matching true protein shapes and achieving a mere 0.1 nm error margin.
In testing, DeepAFM correctly identified the protein’s conformational state across a vast dataset of 800,000 images, achieving a 93.4% accuracy rate. Furthermore, when applied to actual HS-AFM images, the tool inferred conformational states that aligned with independent experimental measurements, underscoring its practical utility.
Mori states, “DeepAFM provides a new deep learning-assisted strategy for analyzing noisy HS-AFM data and facilitates studies of protein dynamics.” The team also noted that DeepAFM could be adapted to analyze other protein systems through transfer learning, indicating its potential versatility in biological research.
This initiative represents a broader effort to push AI-driven research, particularly in preparation for advanced computing platforms like Fugaku NEXT, developed collaboratively by the RIKEN Center for Computational Science, Fujitsu, and NVIDIA, anticipated to launch around 2030.
The content above is a summary. For more details, see the source article.