P04-04

Analysis of HS-AFM images of proteins combining MD simulation and machine learning

Katsuki SATO *, Takaharu MORI

Department of Chemistry, Tokyo University of Science
( * E-mail: 1324572@ed.tus.ac.jp )

High-speed atomic force microscopy (HS-AFM) has been widely used for real-time, direct observation of protein conformational changes. Typical resolution of the HS-AFM images is ~0.15 nm in the vertical direction, while 2–3 nm in the lateral direction, making the identification of the protein conformation difficult. To solve this problem, we developed a new algorithm combining molecular dynamics (MD) simulation and convolutional neural network (CNN). In the method, MD simulation is first carried out to sample conformational changes of the target protein, and then pseudo-AFM images are generated from each MD snapshot as training data set of CNN. After training the CNN using the pseudo-AFM images, it is applied to the experimental AFM images. To investigate the performance of our method, we selected a protein that undergoes a large conformational change. We performed the MD simulations of the protein to sample various conformations, followed by the analysis of an “artificial” experimental AFM image of the known state. The results demonstrated that the well-trained CNN could identify a conformational state of the target protein. The detailed results will be shown at the poster presentation.