Abstract
BACKGROUND: Detecting the differences in gene expression data is important for understanding the underlying molecular mechanisms. Although the differentially expressed genes are a large component, differences in correlation are becoming an interesting approach to achieving deeper insights. However, diverse metrics have been used to detect differential correlation, making selection and use of a single metric difficult. In addition, available implementations are metric-specific, complicating their use in different contexts. Moreover, because the analyses in the literature have been performed on real data, there are uncertainties regarding the performance of metrics and procedures. RESULTS: In this work, we compare four novel and two previously proposed metrics to detect differential correlations. We generated well-controlled datasets into which differences in correlations were carefully introduced by controlled multivariate normal correlation networks and addition of noise. The comparisons were performed on three datasets derived from real tumor data. Our results show that metrics differ in their detection performance and computational time. No single metric was the best in all datasets, but trends show that three metrics are highly correlated and are very good candidates for real data analysis. In contrast, other metrics proposed in the literature seem to show low performance and different detections. Overall, our results suggest that metrics that do not filter correlations perform better. We also show an additional analysis of TCGA breast cancer subtypes. CONCLUSIONS: We show a methodology to generate controlled datasets for the objective evaluation of differential correlation pipelines, and compare the performance of several metrics. We implemented in R a package called DifCoNet that can provide easy-to-use functions for differential correlation analyses.