|
Abstract:
|
High-throughput sequencing data typically harbor unwanted variation from diverse sources. Existing statistical methods for parsing the sources of unwanted variation assume that these multiple sources are uncorrelated with each other, an assumption that is frequently not met in sequencing data due to poor experimental design or technical limitations. We present a statistical framework to uncover hidden sources of variation even when these sources are correlated, namely Iteratively Adjusted Surrogate Variable Analysis (IA-SVA). IA-SVA provides a flexible methodology to i) identify a hidden factor for unwanted heterogeneity while adjusting for all known factors; ii) test the significance of the putative hidden factor for explaining the variation in the data; iii) adjust the data for the detected factor; and iv) iterate the procedure to uncover further hidden factors. Using simulated and real-world bulk and single-cell RNA-Seq data, we show that IA-SVA outperforms alternative methods in terms of statistical power, Type I error rate, and accuracy in detecting/estimating the hidden factors and proved to be an effective method in the absence of a negative control set.
|
