Model based normalization improves differential expression calling in low-depth RNA-seq

Abstract

RNA-seq is a powerful tool for gene expression profiling and differential expression analysis. Its power depends on sequencing depth which limits its high-throughput potential, with 10-15 million reads considered as optimal balance between quality of differential expression calling and cost per sample. We observed, however, that some statistical features of the data, e.g. gene count distribution, are preserved well below 10-15M reads, and found that they improve differential expression analysis at low sequencing depths when distribution statistics is estimated by pooling individual samples to a combined higher-depth library. Using a novel gene-by-gene scaling technique, based on the fact that gene counts obey Pareto-like distribution1, we re-normalize samples towards bigger sequencing depth and show that this leads to significant improvement in differential expression calling, with only a marginal increase in false positive calls. This makes differential expression calling from 3-4M reads comparable to 10-15M reads, improving high-throughput of RNA-sequencing 3-4 fold.