Identification of gene-gene interactions for Alzheimer's disease using co-operative game theory
Vardarajan, Badri N.
MetadataShow full item record
The multifactorial nature of Alzheimer's Disease suggests that complex gene-gene interactions are present in AD pathways. Contemporary approaches to detect such interactions in genome-wide data are mathematically and computationally challenging. We investigated gene-gene interactions for AD using a novel algorithm based on cooperative game theory in 15 genome-wide association study (GWAS) datasets comprising of a total of 11,840 AD cases and 10,931 cognitively normal elderly controls from the Alzheimer Disease Genetics Consortium (ADGC). We adapted this approach, which was developed originally for solving multi-dimensional problems in economics and social sciences, to compute a Shapely value statistic to identify genetic markers that contribute most to coalitions of SNPs in predicting AD risk. Treating each GWAS dataset as independent discovery, markers were ranked according to their contribution to coalitions formed with other markers. Using a backward elimination strategy, markers with low Shapley values were eliminated and the statistic was recalculated iteratively. We tested all two-way interactions between top Shapley markers in regression models which included the two SNPs (main effects) and a term for their interaction. Models yielding a p-value<0.05 for the interaction term were evaluated in each of the other datasets and the results from all datasets were combined by meta-analysis. Statistically significant interactions were observed with multiple marker combinations in the APOE regions. My analyses also revealed statistically strong interactions between markers in 6 regions; CTNNA3-ATP11A (p=4.1E-07), CSMD1-PRKCQ (p=3.5E-08), DCC-UNC5CL (p=5.9e-8), CNTNAP2-RFC3 (p=1.16e-07), AACS-TSHZ3 (p=2.64e-07) and CAMK4-MMD (p=3.3e-07). The Shapley value algorithm outperformed Chi-Square and ReliefF in detecting known interactions between APOE and GAB2 in a previously published GWAS dataset. It was also more accurate than competing filtering methods in identifying simulated epistastic SNPs that are additive in nature, but its accuracy was low in identifying non-linear interactions. The game theory algorithm revealed strong interactions between markers in novel genes with weak main effects, which would have been overlooked if only markers with strong marginal association with AD were tested. This method will be a valuable tool for identifying gene-gene interactions for complex diseases and other traits.
Thesis (Ph.D.)--Boston University