Abstract 4931: Biomarker discovery in a large panel of cell lines shows different sample size requirement for prediction of response across a set of compounds

Abstract

Abstract A panel of more than 600 cell lines from 17 tumor types has been profiled and sensitivity to a set of FDA approved compounds with different mechanisms of action has been tested. Comparison of gene expression profiles with overlapping set of publicly available profiles showed 100% accuracy of cell line identity prediction using nearest neighbor classifier. Similar analysis of CNV data had 80% accuracy due to relatively little CNV perturbation in some of the cell lines. Significant gene expression signatures have been detected for 80% of compounds. De-novo agnostic classification based on 50% train/test split and a linear classifier resulted in significant prediction on the test set for about 40% of the compounds, such as dasatinib, 5FU, paclitaxel, but failed to produce a significant prediction for others, such as doxorubicine, irinotecan, and vinblastine. For most of the compounds, the prediction of response is complex, with multiple distinct molecular features contributing to a classification algorithm. This inherent complexity requires integration of gene expression, CNV and mutation data as well as a large cell line sets for development of accurate classification algorithms. We defined functional CNV and SNV events using gene expression based modules as a functional readout. Predictive models that incorporate prior knowledge of mechanism of action of the compounds and rely on functional SNV and CNV events out perform completely agnostic methods of prediction. Citation Format: {Authors}. {Abstract title} [abstract]. In: Proceedings of the 103rd Annual Meeting of the American Association for Cancer Research; 2012 Mar 31-Apr 4; Chicago, IL. Philadelphia (PA): AACR; Cancer Res 2012;72(8 Suppl):Abstract nr 4931. doi:1538-7445.AM2012-4931