Automated detection and identification of membrane proteins in atomic force microscopy images using machine learning

Atomic force microscopy (AFM) is an effective single molecule technique for imaging membrane proteins in their native environment under physiological conditions. In spite of the high out of membrane (z-axis) spatial resolution (~1Å) of AFM images, due to the finite size of the AFM tip and other factors, the corresponding lateral resolution is usually more than an order of magnitude lower (>1nm). By employing a recently developed localization AFM (LAFM) approach from the Scheuring laboratory, the lateral resolution can be enhanced (typically by a factor of three), a major advance for the field. Yet, the processing and evaluation of the large number of images recorded in AFM experiments are time consuming and challenging. Here we present a workflow for the automated detection and identification of membrane proteins in a previously recorded stack of AFM images. The protein blobs are detected by standard image segmentation methods. After proper alignment and LAFM resolution enhancement, the protein blobs are identified by using a deep convolutional neural network (CNN). The CNN is trained by using (i) simulated AFM images of atomistic protein structures in the OPM database, and (ii) previously validated AFM images. The effectiveness of the proposed workflow is demonstrated for a system comprising membrane proteins from the general secretory system in a supported E. coli lipid bilayer.