The Poisson Margin Test for Normalization-Free Significance Analysis of NGS Data

The current methods for the determination of the statistical significance of peaks and regions in next generation sequencing (NGS) data require an explicit normalization step to compensate for (global or local) imbalances in the sizes of sequenced and mapped libraries. There are no canonical methods for performing such compensations; hence, a number of different procedures serving this goal in different ways can be found in the literature. Unfortunately, the normalization has a significant impact on the final results. Different methods yield very different numbers of detected "significant peaks" even in the simplest scenario of ChIP-Seq experiments that compare the enrichment in a single sample relative to a matching control. This becomes an even more acute issue in the more general case of the comparison of multiple samples, where a number of arbitrary design choices will be required in the data analysis stage, each option resulting in possibly (significantly) different outcomes. In this article, we investigate a principled statistical procedure that eliminates the need for a normalization step. We outline its basic properties, in particular the scaling upon depth of sequencing. For the sake of illustration and comparison, we report the results of re-analyzing a ChIP-Seq experiment for transcription factor binding site detection. In order to quantify the differences between outcomes, we use a novel method based on the accuracy of in silico prediction by support vector machine (SVM) models trained on part of the genome and tested on the remainder. See Kowalczyk et al. ( 2009 ) for supplementary material.