Abstract
High-speed atomic force microscopy (HS-AFM) is a powerful technique for visualizing protein dynamics in real time at the single-molecule level and has enabled direct observation of diverse biomolecular processes such as protein conformational changes, enzymatic reactions, and protein-protein interactions. Despite these advantages, HS-AFM imaging often suffers from substantial noise and limited spatial resolution, which complicates the reliable identification of detailed protein conformational states. To address these limitations, we introduce DeepAFM, a framework that integrates deep learning with molecular dynamics (MD) simulations to estimate protein conformational states while denoising AFM images. The model is trained on simulated AFM images generated from MD snapshots, incorporating realistic noise to mimic experimental conditions, including temporal lag effects between line scans. As a case study, we apply DeepAFM to the membrane protein SecYAEG-nanodisc complex, in which SecA undergoes conformational transitions between closed and wide-open states. The trained model preferentially attends to regions in the input images corresponding to large-scale domain motions of SecA, thereby increasing robustness to noise-induced overfitting compared with conventional rigid-body and flexible fitting. By effectively denoising experimental HS-AFM images, DeepAFM estimates the dominant conformational states of the protein, in agreement with independent experimental observations. DeepAFM provides a deep-learning-assisted analysis strategy for the interpretation of noisy HS-AFM data.