Abstract 49: Constructing pathway based predictors of cancer clinical outcome

Abstract

Abstract Current efforts to understand the mechanism of cancer involve using various whole-genome -omics measurements over large patient cohorts. Since a patient response to treatments is highly variable, the challenge then is to integrate the data in order to infer patient-specific disease mechanisms. Recent advances in the analysis of cancer (TCGA ovarian serous carcinoma and glioblastoma multiforme) has shown that a pathway interpretation of DNA copy number, DNA methylation, mRNA expression, and mutations offers a powerful framework for interpreting complex data. The hope is that a pathway-level interpretation of -omics data can identify pathway signatures to predict differences in clinical outcome, whereas traditional machine learning algorithms do not take advantage of the pathway structure of biological data. We are developing a pathway prediction method based on PARADIGM to discriminate patient outcome based on pathway signatures. Utilizing conditional random fields (CRFs) allow for formal search for a graphical model that optimizes the prediction of a particular variable of interest (VOI) defined by the given classification task, as opposed to a generative model that optimizes the model to explain the data. The method first merges pathways to build a core network around the VOI. The model then seeks to extend the pathway to include new genes and interactions which improve the model's predictive ability on the training data. Application of our method to 50 breast cancer cell-lines treated with 80 different compounds revealed general and subtype-specific signatures of response in breast cancer. We compared our CRF-based method against a compendium of standard machine-learning algorithms and found that our CRF outperformed all methods on a majority of drugs tested. We also tested the method on a cancer benchmark consisting of a dozen prediction challenges all involving the prediction of clinical outcomes on large patient cohorts using gene expression and copy number data. Again, the CRF model outperformed a majority of classifiers and performed comparably to the best classifiers on most challenges. We expect our method to generalize to a wide variety of biological systems for which high-throughput genomics and functional genomics are available. Citation Format: {Authors}. {Abstract title} [abstract]. In: Proceedings of the 102nd Annual Meeting of the American Association for Cancer Research; 2011 Apr 2-6; Orlando, FL. Philadelphia (PA): AACR; Cancer Res 2011;71(8 Suppl):Abstract nr 49. doi:10.1158/1538-7445.AM2011-49