Extending Localization Atomic Force Micrsocopy to Flexible Membrane Proteins via Deep Learning

Atomic Force Microscopy (AFM) is an experimental technique for imaging individual biomolecules under near native, physiological conditions. For example, the molecularly sharp AFM tip continuously rastered across membrane proteins embedded in a supported lipid bilayer creates a stack of topographic images of the sample. Unfortunately, the lateral (in the plane of the membrane) resolution of the acquired AFM images is limited by the finite (several nm) AFM tip size, thus severely limiting the identification of conformational details of the imaged molecule. However, by using image reconstruction algorithms applied to peak positions in the stack of AFM images, the recently introduced Localization Atomic Force Microscopy or LAFM (Heath et al, Nature 594,385 (2021)) can lead to several-fold enhancement of the lateral resolution, thus revealing conformational details missing in the original AFM images. While the method provides impressive results for "rigid" membrane proteins with a single conformation and stable structural orientation, it is not clear that the advantages of LAFM extend to "flexible" proteins that display several conformations and variable structural orientation. Here we introduce an unsupervised artificial neural network model that makes LAFM applicable to flexible membrane proteins. Our model incorporates a recursive deep clustering algorithm that broadens the applicability of LAFM by: (1) clustering the stack of AFM images into sets corresponding to distinct protein conformations, and (2) registering the AFM images within each cluster. Finally, by applying the LAFM algorithm to each cluster independently, enhanced resolution images for each protein conformation are obtained. We demonstrate the effectiveness of the proposed method using simulated AFM images obtained from long time scale MD simulations of the SecYEG translocon, a heterotrimeric membrane protein complex embedded in fully solvated POPE lipid bilayer.