The few studies that explored functional interactions in high-order combinations are mostly based on SNP datasets that cover a small number of genes . In addition, these studies only focus on one or a few top ranked combinations discovered from a single dataset and thus only reveal disease-specific functional interactions . In this study, GDC-0449 before interpreting the top high-order SNP combinations, we first explore functional interactions in SNP combinations from a more general perspective. The aim is to exploit some common insights on functional interactions in discriminative SNP combinations consistent across multiple datasets which may provide some guidance for future studies. To measure the functional coherence of a SNP combination, we first SB203580 obtain the set of genes covered by the combination by assigning a SNP to its closest gene, and then determine the functional similarity between each unique pair of genes covered by the combination using a human functional network integrated from a comprehensive set of resources . Essentially, such an estimation decomposes the functional coherence of a set of genes covered by a SNP combination into the functional similarities of the set of unique gene pairs. We prefer this approach to a GO enrichment analysis because: 1) the former can provide more detailed functional insights on gene-gene interactions within high order combinations, and 2) the latter is usually applicable to gene sets that are of sizes larger than the high-order SNP combinations discovered in this study . With the decomposition-based approach for each SNP combination, we can get three distributions of gene-gene functional similarities for the three groups of SNP combinations GP1, GP2 and GP3 respectively, where each distribution contains the functional similarities of the union of the within-pattern gene pairs from all the patterns in one of the three groups. In addition to the three distributions, we also generate a null distribution by repeating the following procedure 100 times: we randomly sample gene pairs from the set of genes covered in the corresponding dataset as many as the number of gene pairs in GP1, while fixing the number of times each unique gene occurs with respect to GP1. Because we binarize the human functional network at 0.5 to make the size of the network efficient to manage). It is worth noting that the following results are consistent across different cutoff values for the functional network . We presented a computational framework for searching highorder SNP combinations with strong disease association from casecontrol datasets with thousands of SNPs. The framework is substantially more efficient and scalable than existing techniques that usually handle tens of or hundreds of SNPs and mostly up to size-3 combinations.