Title: Entropy-based criteria for multivariate association and omics network models
Authors: Takoua Jendoubi - Imperial College London (United Kingdom) [presenting]
Korbinian Strimmer - University of Leipzig (Germany)
Abstract: In the last twenty years, the parallel acquisition of high-throughput omics datasets has seen a tremendous boost pushing forward deeper understanding of biological functions and molecular mechanisms. This is commonly achieved by investigating the degree of co-expression between omics variables, which is often estimated using measures of association such as correlation coefficients. Despite the widespread use of these measures, they do not convey information about the entire multivariate system. On the other hand, multivariate measures of association generalizing correlations coefficients to two random vectors, such as the RV coefficient or the distance covariance (dCov) coefficient, do not explore pairwise contributions to total association between variables which may be crucial to uncover inter-omics interactions. To address these drawbacks, we propose to use vector correlation coefficient, which seems to be largely ignored, as an alternative to the RV and dCov coefficients. By entropy derivation we show that this approach is natural in the setting of latent-variable multivariate regression and probabilistic canonical correlation analysis. In addition, we show that this measure offers a decomposition property allowing to dissect the total association in order to construct a network with pairwise contributions. We illustrate our approach by analyzing both synthetic and publicly available omics data.