Deep Learning Segmentation of Protein EM Maps

Protein Electron Microscopy (EM) maps are critical in determining the three-dimensional structures of bio-molecules, including proteins and their interactions. The task of identifying regions that correspond to specific proteins is challenging, but crucial for gaining insight into their function and designing drugs to enhance or suppress their processes. Conventional methods of protein EM map segmentation use algorithms that assign a voxel to a region, but can result in difficulties in obtaining a segmentation that maps each region to a single protein unit. Our approach incorporates an interactive mechanism that allows for user guidance of the segmentation process. Our deep learning model is trained on a dataset of protein EM maps and uses a U - Net architecture. Results show that our approach has potential to overcome limitations of traditional state-of-the-art approaches.