Abstract 671: Evaluation of machine learning tools for cancer drug response prediction with genomic profile data and drug response data from PDX model studies

Abstract

Abstract Cancer are diseases caused by genomic aberrations which lead to abnormal cell proliferation and metastasis. Cancer treatment based on cancer mutation profiles are the focus of research and development of cancer precision medicine. Currently only a small percentage of cancer patients’ genomic information are associated with known drug responses, while others lack known cancer drivers or have cancer drivers that are not druggable. Machine learning (ML) tools have been developed recently for cancer biomarker discovery and cancer drug response prediction. However, due to a lack of sufficient amount of high quality drug response data, especially data for new drugs, the application of ML tools in this area is still in its early stage. In this study, we compared the performance of about a dozen ML methods, such as random forest classifiers, logistic regression classifiers, and support vector machines (SVM), trained on data of about 2300 treatment responses to average 12 different cancer drugs in 98 NSCLC, 55 gastric and 38 HCC PDX models with WES, CNV and RNA-seq information. Data preparation including genomic profile data selection, dimensionality feature reduction and processing was performed with considerations of known cancer gene functions, signal pathways, and pharmacodynamics information. Each dataset was randomly divided into training dataset and test dataset for ML algorithm training and evaluation. Performance of ML algorithms in testing was measured by their drug response prediction accuracy. Different regularization methods, such as LASSO and Elastic Net, were also evaluated in order to improve classification accuracy and identify relevant biomarker features. From comparing results from different algorithms or from the same algorithms but with differently prepared datasets, it seems that selection of functionally unknown gene aberrations for data preparation plays a significant role in prediction accuracy. Increasing dataset number or different data selection in different cancer types can improve prediction accuracy. Citation Format: Henry Gu, Jingjing Jiang. Evaluation of machine learning tools for cancer drug response prediction with genomic profile data and drug response data from PDX model studies [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2019; 2019 Mar 29-Apr 3; Atlanta, GA. Philadelphia (PA): AACR; Cancer Res 2019;79(13 Suppl):Abstract nr 671.