Applying variancePartition to the GEUVADIS [6, 8] dataset illustrates how the method can decouple biological and technical variation, and further decompose biological variation into multiple components. Expression variation across individuals, ancestries and sexes is biological, variation across the labs where the samples were sequenced comprise technical artifacts, while the residual variation remains uncharacterized. Results from representative genes illustrate how variancePartition identifies genes where the majority of variation in expression is explained by a single variable such as individual or sex, while variation in other genes is driven by multiple variables (Fig. 2 a). Since the variance fractions sum to 1 for each gene, it is simple to compare results across genes and across sources of variation. Visualizing these results genome-wide illustrates that variation across individuals is the major source of expression variation and explains a median of 55.1% of variance genome-wide (Fig. 2 b). The median variance explained by laboratory (6.8%), ancestry (4.9%) and sex (0%) is substantially smaller. We note that the variance explained by individual increases to 63.8% when ancestry is removed from the analysis since ancestry is a biological property of each individual (Additional file 1: Figure S5).
