Unbiased distance correlation with sample-size-aware confidence bounds for comparative omics network analysis

Data-driven determination is a powerful approach for unbiased investigation of the functional relationships in biomolecular networks. Such networks can be inferred from omics data, where correlation analysis is a commonly used method. However, the correlation values depend strongly on sample variability and size of the sample set in the general case, leading to unstable results and possibly highly erroneous conclusions.

In this work, we show that similar to the Pearson and Spearman correlation approaches, distance correlation as a general non-linear polytonic correlation method also depends on the sample size. We show that both the p- and correlation values decrease with increasing sample sizes independent of the type of functional relationship between the features but relative to the correlation level. However, the dependence on sample size is greatly reduced in the unbiased distance correlation formulation. We validate an equation to compute the p-value and propose a threshold based on the false discovery rate (FDR) to identify significant correlations in unbiased distance correlation. Furthermore, we derive an extension of Hoeffding's inequality for estimating the error range of correlation as a function of the sample size as an additional sample-size-dependent confidence measure.

We integrated bias-corrected distance correlation with sample-size-matched bootstrapping, chi-squared p-value calculation, FDR adjustment, and empirical Bernstein/Hoeffding-type confidence bounds to support comparative correlation-network analysis in groups with unequal sample sizes. We also provide an online software solution for this application with extensive graphical presentations of the results. The use of this approach is demonstrated in the analysis of major network changes in Alzheimer's disease (AD) using previously published metabolomics data. Our analysis shows major changes in the metabolic networks, with higher connectivity in the AD group than the control group and examples of network changes for specific metabolites.

We present a method for unbiased distance correlation network derivation with permutations and comparisons between the sample groups. All approaches presented here are available through an online application (https://www.insilicobiology.ca/shiny/sidco+/), and all related code is available at https://github.com/computationalmetabolomicsca/sidco_plus.