NOVEL DEEP LEARNING COMPUTER VISION APPROACH FOR DRUG SENSITIVITY PREDICTION

You have accessJournal of UrologyBladder Cancer: Basic Research & Pathophysiology II (MP57)1 Apr 2019MP57-05 NOVEL DEEP LEARNING COMPUTER VISION APPROACH FOR DRUG SENSITIVITY PREDICTION Tom Sanford*, Reema Railkar, Stephanie Harmon, Sheng Xu, Brad Wood, Peter Choyke, Baris Turkbey, and Piyush Agarwal Tom Sanford*Tom Sanford* More articles by this author , Reema RailkarReema Railkar More articles by this author , Stephanie HarmonStephanie Harmon More articles by this author , Sheng XuSheng Xu More articles by this author , Brad WoodBrad Wood More articles by this author , Peter ChoykePeter Choyke More articles by this author , Baris TurkbeyBaris Turkbey More articles by this author , and Piyush AgarwalPiyush Agarwal More articles by this author View All Author Informationhttps://doi.org/10.1097/01.JU.0000556630.42009.04AboutPDF ToolsAdd to favoritesDownload CitationsTrack CitationsPermissionsReprints ShareFacebookLinked InTwitterEmail Abstract INTRODUCTION AND OBJECTIVES: Despite recent advances in second line therapy for patients with advanced bladder cancer, most patients eventually progress. Selection of the choice of additional therapy who have progressed despite cisplatin-based therapy and immunotherapy is difficult. Docetaxel has been tested in the 2nd line setting and is associated with a 13% response rate. In this study, we utilize a novel computer vision deep learning approach to predict sensitivity to docetaxel using gene expression microarray data. METHODS: Drug sensitivity data and gene expression data were obtained from the genomics of drug sensitivity website (www.cancerrxgene.org). Gene expression data were mapped to chromosomal bands and the expression data were averaged over each chromosomal band. Data were then displayed as chromosomal heatmaps with a total of 22 columns where each column represented an autosome in order of increasing chromosomal number from left to right. The chromosomal maps were utilized to develop a prediction model by training a convolutional neural network with residual blocks (ResNet architecture) from the fast.ai library (www.fast.ai.com). We trained a binary image classification algorithm to distinguish the most sensitive 50% of cell lines from the least sensitive 50% of cell lines. Transfer learning was performed with weights trained on ImageNet. RESULTS: A representative chromosomal heatmap based on gene expression is shown below. There were a total of 940 cell lines that were tested with Docetaxel derived from cancers derived from 30 anatomic locations. The deep neural network was trained using 150 epochs with a learning rate of 1.0x10-4. We were able to achieve 81% correct classification for sensitivity on the validation dataset. The sensitivity was 80% with 78% specificity. The positive predictive value was 86% with 69% negative predictive value. CONCLUSIONS: We employ a novel computer vision approach utilizing deep learning on a visual representation of gene expression mapped to chromosomal bands. This methodology creates a genetic "fingerprint" that is robust to changes in a small number of genes. For docetaxel, we were able to achieve a high positive predictive value for drug sensitivity, which may be helpful in determining which patients to attempt this drug as second line therapy for advanced bladder cancer. Source of Funding: NIH Intramural Research Program and Bethesda, MD© 2019 by American Urological Association Education and Research, Inc.FiguresReferencesRelatedDetails Volume 201Issue Supplement 4April 2019Page: e814-e814 Advertisement Copyright & Permissions© 2019 by American Urological Association Education and Research, Inc.MetricsAuthor Information Tom Sanford* More articles by this author Reema Railkar More articles by this author Stephanie Harmon More articles by this author Sheng Xu More articles by this author Brad Wood More articles by this author Peter Choyke More articles by this author Baris Turkbey More articles by this author Piyush Agarwal More articles by this author Expand All Advertisement PDF downloadLoading ...