Comprehensive analysis of network reconstruction approaches based on correlation in metagenomic data

Microbiome analysis is transforming our understanding of biological processes related to human health, epidemiology (antimicrobial resistance, horizontal gene transfer) environmental and agricultural studies. At the core of microbiome analysis is the description of microbial communities based on quantification of microbial taxa and dynamics. In the study of bacterial abundances, it is becoming more relevant to consider their relationship, to embed these data in the framework of network theory, allowing characterization of features like node relevance, pathway and community structure. In this work we characterize the principal biases in reconstructing networks from correlation measures, associated with the compositional character of relative abundance data, the diversity of abundances and the presence of unobserved species within a single sample, that might lead to wrong correlation estimates. We show how most of these problems can be overcome by applying typical transformations for compositional data, that allow the application of simple measures such as Pearsons correlation to correctly identify the positive and negative relationships between relative abundances, when data dimensionality is sufficiently high. Some issues remain, like the role of data sparsity, that if not properly addressed can lead to imbalances in correlation coefficient distribution.