Iterative feature selection method to discover predictive variables and interactions for high-dimensional transplant genomic data

After allogeneic hematopoietic stem cell transplantation (allo-HCT), donor-derived immune cells can trigger devastating graft-versus-host disease (GVHD). The clinical effects of GVHD are well established; however, genetic mechanisms that contribute to the condition remain unclear. Candidate gene studies and genome-wide association studies have shown promising results, but they are limited to a few functionally derived genes and those with strong main effects. Transplant-related genomic studies examine two individuals simultaneously as a single case, which adds additional analytical challenges. In this study, we propose a hybrid feature selection algorithm, iterative Relief-based algorithm followed by a random forest (iRBA-RF), to reduce the SNPs from the original donor-recipient paired genotype data and select the most predictive SNP sets in association with the phenotypic outcome in question. The proposed method does not assume any main effect of the SNPs; instead, it takes into account the SNP interactions. We applied the iRBA-RF to a cohort (n=331) of acute myeloid leukemia (AML) patients and their fully 10 of 10 (HLA-A, -B, -C, -DRB1, and -DQB1) HLA-matched healthy unrelated donors and assessed two case-control scenarios: AML patients vs healthy donor as case vs control and acute GVHD group vs non-GVHD group as case vs control, respectively. The results show that iRBA-RF can efficiently reduce the size of SNPs set down to less than 0.05%. Moreover, the literature review showed that the selected SNPs appear functionally involved in the pathologic pathways of the phenotypic diseases in question, which may potentially explain the underlying mechanisms. This proposed method can effectively and efficiently analyze ultra-high dimensional genomic data and could help provide new insights into the development of transplant-related complications from a genomic perspective.