Major interest in current epidemiology, medicine, and pharmarco-genomics is focused on identifying single nucleotide polymorphisms (SNPs) that underlie the etiology of common and complex diseases. However, due to the tremendous number of SNPs on the human genome, there is a clear need to prioritize SNPs to expedite genotyping and analysis overhead associated with disease-gene studies. Tag SNP selection and Functional SNP selection are the two main approaches for addressing the SNP selection problem. However, little was done so far to effectively combine these distinct and possibly competing approaches. Here we present a new multi-objective optimization framework for identifying SNPs that are both informative tagging and have functional significance.
Our SNP selection algorithm is based on the notion of Pareto optimality , which has been extensively used for addressing multi-objective optimization problems in game theory, economics and engineering. We describe the details of its three main steps as follows.
STEP 1. Computing Linkage Disequilibrium of SNPs
To efficiently compute the score of tagging informativeness, we calculate the pair-wise LD between all pairs of candidate SNPs in advance. As a measure of pair-wise LD, following Carlson et al. , we currently use the coefficient of determination, r2.
STEP 2. Retrieving Functional Significance of SNPs
We currently use the FS score of SNPs obtained from F-SNP , which assesses the deleterious functional effects of SNPs, using 16 bioinformatics tools, with respect to protein translation, splicing regulation, transcriptional regulation, and post-translational modification.
STEP 3. Selecting Functionally Informative Tag SNPs
Our selection algorithm is based on multi-objective simulated-annealing (SA) search. We also introduce two heuristics for generating a new neighboring solution to guide efficient search while expediting convergence. Figure 1 summarizes the proposed algorithm.
Figure 1. The Multi-Objective SA Algorithm.
We applied our system to 34 disease-susceptibility genes for lung cancer, which is one of the most extensively-studied cancer types due to its high mortality rate . Our algorithm always finds a collection of Pareto optimal SNP subsets that performs better than the subsets selected by other SNP selection approaches, with respect to both tagging informativeness and functional significance (shown in Figure 2). Moreover, we clearly show that our system improves upon general-purpose search algorithms for identifying Pareto optimal solutions (p-values are 1.37e-004, 3.11e-015, 2.43e-149 and 3.89e-179).
Figure 2. Evaluation results of three Pareto optimal search algorithms against two compared systems.