2018/02/07 by Thomas Corcoran, Rafael Zamora‐Resendiz, Corcoran, Thomas +6 · 1 citation
Biochemistry, Genetics and Molecular Biology · Computer Science · #Cell Image Analysis Techniques #Computational Drug Discovery Methods #Computer Vision and Pattern Recognition (cs.CV) #FOS: Computer and information sciences #Machine Learning (cs.LG) #Medical Image Segmentation Techniques #Protein Structure and Dynamics
paper · pdf · doi:10.48550/arxiv.1802.02532
openalex publication_date 2018/02/07 · openalex created_date 2022/10/01 · openalex updated_date 2026/07/28
Convolutional Neural Network (CNN)-based machine learning systems have made\nbreakthroughs in feature extraction and image recognition tasks in two\ndimensions (2D). Although there is significant ongoing work to apply CNN\ntechnology to domains involving complex 3D data, the success of such efforts\nhas been constrained, in part, by limitations in data representation\ntechniques. Most current approaches rely upon low-resolution 3D models,\nstrategic limitation of scope in the 3D space, or the application of lossy\nprojection techniques to allow for the use of 2D CNNs. To address this issue,\nwe present a mapping algorithm that converts 3D structures to 2D and 1D data\ngrids by mapping a traversal of a 3D space-filling curve to the traversal of\ncorresponding 2D and 1D curves. We explore the performance of 2D and 1D CNNs\ntrained on data encoded with our method versus comparable volumetric CNNs\noperating upon raw 3D data from a popular benchmarking dataset. Our experiments\ndemonstrate that both 2D and 1D representations of 3D data generated via our\nmethod preserve a significant proportion of the 3D data's features in forms\nlearnable by CNNs. Furthermore, we demonstrate that our method of encoding 3D\ndata into lower-dimensional representations allows for decreased CNN training\ntime cost, increased original 3D model rendering resolutions, and supports\nincreased numbers of data channels when compared to purely volumetric\napproaches. This demonstration is accomplished in the context of a structural\nbiology classification task wherein we train 3D, 2D, and 1D CNNs on examples of\ntwo homologous branches within the Ras protein family. The essential\ncontribution of this paper is the introduction of a dimensionality-reduction\nmethod that may ease the application of powerful deep learning tools to domains\ncharacterized by complex structural data.\n