We agree with Elbers et al. (1) that there are a number of limitations to genome-wide association (GWA) study pathway-based approaches as currently implemented. Not least of which are the incomplete knowledge and characterization of human genes and pathways and the statistical difficulties in assessing the significance of a pathway. We would, however, like to address some of their specific criticisms.
First, we would like to respond to the concern that few of the type 2 diabetes loci genes are represented in the pathways tested. Our study used three different pathway datasets: Kyoto Encyclopedia of Genes and Genomes (KEGG), BioCarta, and Gene Ontology (GO; level 4 annotations in Biological Process and Molecular Function). These three datasets combined include 20 of the 23 genes from 18 known loci opposed to the 5 found in BioCarta and KEGG alone.
Second, it is true that many of the genes identified from GWA studies appear to influence β-cell function (2) but that β-cell function is not a defined pathway in any of the pathway databases. However, whereas it is accepted that β-cell dysfunction is important in type 2 diabetes development, it is a very broad category defined more by physiology than by molecular biology. More interesting would be to identify the molecular pathways that influence β-cell dysfunction.
Third, it is true that many pathway-based approaches suffer from distortion of individual pathway test statistics due to varying linkage disequilibrium patterns and gene/pathway sizes. However, this particular problem was overcome in our analysis using a case-control label-swapping permutation approach. Using the gene set enrichment analysis–implemented algorithm, each permutation cycle swaps the case-control label in the GWA dataset, finds the new single nucleotide polymorphism with the highest test statistic for each gene, and then calculates a new pathway statistic (formulae described in Wang et al. [(3)]). Over all permutations, the permuted pathway statistics are compared with the observed statistic to generate a nominal P value. Larger genes will have by chance lower observed P values in the dataset; however, each permutation will also find new lower P values through label swapping than the smaller genes. Therefore, the algorithm corrects for pathway size, gene size, and single nucleotide polymorphism coverage, and there is no bias to the overall results. Indeed, this was observed in our results with no significant association between test statistic and pathway size (Spearman's rank correlation: P = 0.85, n = 439 pathways).
Finally, we would also like to point out that pathway-based efforts may be met with varying degrees of success depending on the underlying genetic architecture of the studied disease. For example, Wang et al. recently used the same algorithm to investigate the Wellcome Trust Case Control Consortium Crohn's disease dataset (4). Here, they find and replicate association between the interleukin-12/IL-23 pathway and Crohn's disease. This 20-gene pathway contained two genes recently identified through meta-analysis of multiple studies as being associated with Crohn's disease, neither of which were genome-wide significant in any individual study. The pathway association also remained significant when these two genes were removed.
In conclusion, whereas we agree with Elbers et al. that there are limitations of pathway-based analyses, we stand by our conclusion that type 2 diabetes genes are likely to fall in many molecular pathways.
Acknowledgments
No potential conflicts of interest relevant to this article were reported.
