SC17 Denver, CO

P36: A Novel Feature-Preserving Spatial Mapping for Deep Learning Classification of Ras Structures


Authors: Thomas Corcoran (Lawrence Berkeley National Laboratory), Rafael Zamora-Resendiz (Lawrence Berkeley National Laboratory), Xinlian Liu (Lawrence Berkeley National Laboratory), Silvia Crivelli (Lawrence Berkeley National Laboratory)

Abstract: A protein’s 3D structure determines its functionality, and is therefore a topic of great importance. This work leverages the power of Convolutional Neural Networks (CNNs) to classify proteins and extract features directly from their 3D structures. So far, researchers have been unable to fully exploit 3D structural information with 2D CNNs, partly because it is difficult to encode 3D data into the 2D format that can be ingested by such networks. We designed and implemented a novel method that maps 3D models to 2D data grids as a preprocessing step for 2D CNN use. Our experiments focused on the Ras protein family, which has been linked to various forms of cancer. Our trained CNNs are able to distinguish between two branches within the Ras family, HRas and KRas, which are similar in sequence and structure. Analysis of saliency maps suggests classification is accomplished by detection of structurally and biologically-meaningful sites.
Award: Best Poster Finalist (BP): no

Poster: pdf
Two-page extended abstract: pdf


Poster Index