Addressing confounding artifacts in reconstruction of gene co-expression networks

Abstract

Gene co-expression networks capture biological relationships between genes and are important tools in predicting gene function and understanding disease mechanisms. We show that technical and biological artifacts in gene expression data confound commonly used network reconstruction algorithms. We demonstrate theoretically, in simulation, and empirically, that principal component correction of gene expression measurements prior to network inference can reduce false discoveries. Using data from the GTEx project in multiple tissues, we show that this approach reduces false discoveries beyond correcting only for known confounders. Electronic supplementary material The online version of this article (10.1186/s13059-019-1700-9) contains supplementary material, which is available to authorized users.

Co-expression analysis is biased by a mean-correlation relationship

“…An interesting question is what forces create these different patterns. Altogether, our work reinforces the message from Parsana et al (2019) that removing principal components is good practice, although exactly how many components to remove is still an open question. Together, this work shows the importance of assessing and addressing mean-correlation bias in co-expression analysis, and provides the first method for doing so.…”

“…Next, we focus on removing an increasing number of PCs. If we use the number of PCs recommended by Parsana et al (2019) (Table 1), we see similar results ( Figure 11a) compared to removing 4 PCs (Figure 7a). However, both the average and the spread of the IQRs of background distributions prior to applying spatial quantile normalization, are smaller than using 4 PCs.…”

“…Starting with 19,836 protein-coding genes and 7,036 lncRNAs, in each tissue we kept genes with a median expression above zero (on the log-RPKM scale, Methods), leaving us with 10,735-12,889 expressed genes per tissue (of which 95%-98% were proteincoding). Removing principal components from the correlation matrix has been shown to remove unwanted variation in co-expression analysis (Parsana et al, 2019). We computed the gene-gene correlation matrix of the log-RPKM values and removed unwanted variation by Increasing expression Figure 1.…”

“…The number 4 was chosen based on our previous analysis (Boukas et al, 2019), where we used positive and negative control genes to assess this. An alternative approach suggested by Parsana et al (2019) is to remove the number of PCs according to the estimated number of surrogate variables using SVA (Leek, Storey, 2007;Leek, Storey, 2008;Leek, Johnson, et al, 2012) with the number of PCs ranging from 10 to 30 in these same tissues -the impact of choosing a different number of PCs will be examined below.…”

“…While co-expression analysis is widely used, there has been less work on various sources of confounding and bias, especially in contrast to the rich literature on such issues in the related field of differential gene expression analysis. Recent work in co-expression analysis is ad-dressing this shortcoming, including work on removing unwanted variation (Freytag et al, 2015;Parsana et al, 2019) and the effect of cell type composition (Zhang et al, 2019).…”