Ordering of Omics Features Using Beta Distributions on Montecarlo p-Values

Abstract

The current trend in genetic research is the study of omics data as a whole, either combining studies or omics techniques. This raises the need for new robust statistical methods that can integrate and order the relevant biological information. A good way to approach the problem is to order the features studied according to the different kinds of data so a key point is to associate good values to the features that permit us a good sorting of them. These values are usually the p-values corresponding to a hypothesis which has been tested for each feature studied. The Montecarlo method is certainly one of the most robust methods for hypothesis testing. However, a large number of simulations is needed to obtain a reliable p-value, so the method becomes computationally infeasible in many situations. We propose a new way to order genes according to their differential features by using a score defined from a beta distribution fitted to the generated p-values. Our approach has been tested using simulated data and colorectal cancer datasets from Infinium methylationEPIC array, Affymetrix gene expression array and Illumina RNA-seq platforms. The results show that this approach allows a proper ordering of genes using a number of simulations much lower than with the Montecarlo method. Furthermore, the score can be interpreted as an estimated p-value and compared with Montecarlo and other approaches like the p-value of the moderated t-tests. We have also identified a new expression pattern of eighteen genes common to all colorectal cancer microarrays, i.e., 21 datasets. Thus, the proposed method is effective for obtaining biological results using different datasets. Our score shows a slightly smaller type I error for small sizes than the Montecarlo p-value. The type II error of Montecarlo p-value is lower than the one obtained with the proposed score and with a moderated p-value, but these differences are highly reduced for larger sample sizes and higher false discovery rates. Similar performances from type I and II errors and the score enable a clear ordering of the features being evaluated.