Machine Learning Prediction of Cancer Susceptibility

Dartmouth College, Hanover NH

Investigators

Linked publications & trials

Abstract

DESCRIPTION (provided by applicant): Susceptibility to sporadic forms of cancer is determined by numerous genetic factors that interact in a nonlinear manner in the context of an individual's age and environmental exposure. This complex genetic architecture has important implications for the use of genome-wide association studies for identifying susceptibility genes. The assumption of a simple architecture supports a strategy of testing each single- nucleotide polymorphism (SNP) individually using traditional univariate statistics followed by a correction for multiple tests. However, a complex genetic architecture that is characteristic of most types of cancer requires analytical methods that specifically model combinations of SNPs and environmental exposures. While new and novel methods are available for modeling interactions, exhaustive testing of all combinations of SNPs is not feasible on a genome-wide scale because the number of comparisons is effectively infinite. Thus, it is critical that we develop intelligent strategies for selecting subsets of SNPs prior to combinatorial modeling. Our objective is to develop a research strategy for the detection, characterization, and interpretation of gene-gene and gene-environment interactions in a genome-wide association study of bladder cancer susceptibility. To accomplish this objective, we will develop and evaluate modifications and extensions to the ReliefF algorithm for selecting or filtering subsets of single-nucleotide polymorphisms (SNPs) for multifactor dimensionality reduction. (MDR) analysis of gene-gene and gene-environment interactions (AIM 1). We will develop and evaluate a stochastic wrapper or search strategy for MDR analysis of interactions that utilizes ReliefF values as a heuristic (AIM 2). The filter approach will be statisyically compared to the wrapper approach. The best ReliefF strategies will be provided as part of our open-source MDR software package (AIM 3). Finally, we will apply the best ReliefF-MDR analysis strategy to the detection, characterization, and interpretation of gene-gene and gene-environment interactions in a large genome-wide association study of bladder cancer susceptibility (AIM 4). The methods developed here will be applied to nearly 1500 haplotype tagging SNPs (tagSNPs) across approximately 300 cancer susceptibility genes measured in 542 subjects with bladder cancer and 745 healthy controls ascertained as part of a large epidemiological study from the state of New Hampshire.