An integrative association method for omics data based on a modified Fisher’s method with application to childhood asthma

The development of high-throughput biotechnologies allows the collection of omics data to study the biological mechanisms underlying complex diseases at different levels, such as genomics, epigenomics, and transcriptomics. However, each technology is designed to collect a specific type of omics data. Thus, the association between a disease and one type of omics data is usually tested individually, but this strategy is suboptimal. To better articulate biological processes and increase the consistency of variant identification, omics data from various platforms need to be integrated. In this report, we introduce an approach that uses a modified Fisher’s method (denoted as Omnibus-Fisher) to combine separate p-values of association testing for a trait and SNPs, DNA methylation markers, and RNA sequencing, calculated by kernel machine regression into an overall gene-level p-value to account for correlation between omics data. To consider all possible disease models, we extend Omnibus-Fisher to an optimal test by using perturbations. In our simulations, a usual Fisher’s method has inflated type I error rates when directly applied to correlated omics data. In contrast, Omnibus-Fisher preserves the expected type I error rates. Moreover, Omnibus-Fisher has increased power compared to its optimal version when the true disease model involves all types of omics data. On the other hand, the optimal Omnibus-Fisher is more powerful than its regular version when only one type of data is causal. Finally, we illustrate our proposed method by analyzing whole-genome genotyping, DNA methylation data, and RNA sequencing data from a study of childhood asthma in Puerto Ricans.Author summary: In this research, we developed a statistical approach using a modified Fisher’s method (denoted as Omnibus-Fisher) to combine separate p-values of association testing for a trait and SNPs, DNA methylation markers, and RNA sequencing, calculated by kernel machine regression into an overall gene-level p-value to account for correlation between omics data. We further extended the method to an optimal version in order to consider all possible disease models.