MASCARA: Coexpression analysis in data from designed experiments

When omics measurements are employed to discover novel parts of a pathway of interest, experiments are usually designed to induce variation in said pathway. Experimental treatment groups can display differences including beyond the pathway of interest. Pathway discovery by coexpression analysis is therefore hampered by the dominant variance induced by the design. However, minor changes in ambient conditions that could not be controlled for, affect the biological sample. This can lead to small, coordinated differences in gene expression or metabolite concentration. These differences can occur within pathways of interest between replicates in the same controlled experimental condition. These systematic differences are usually not directly considered, but we argue here that they can be used to improve the detection of features that co-express. We introduce a novel framework “MASCARA” which combines ANOVA simultaneous component analysis and partial least squares regression to remove the experimentally induced variance and investigate multivariate relationships in the non-designed variance. Through extensive simulations that mimic the structure of a designed transcriptome study, the benefits of MASCARA are demonstrated and tested against a selection of commonly used methods. In a coexpression analysis of a real dataset MASCARA detects several relevant novel transcripts. The concept is further demonstrated in a designed metabolomics study. Here, distinct groups of features more tightly related within treatment groups or related across experimental conditions can be described. Our results indicate that after adjusting for experimental differences, the remaining variance contains enough useful structure to study coexpression.