From 16S rDNA-seq count data to microbial interactions: A comprehensive benchmark of co-occurrence network inference methods

The study of the ecological interactions within microbial communities is crucial for gaining deeper insights into community ensemble and evolution, enabling targeted modification of the community's functionality. Many inference approaches have been developed to derive ecological associations. Here, we focus on co-occurrence computational methods, in which associations are inferred from metagenomic abundance data by measuring the statistical dependence between species counts across different sampled populations. The limitations in culture-based experiments have hindered the ability to assess these methods' validity, leading to scarce application in metagenomics studies. To provide a comprehensive assessment of network inference method performance, we benchmark 18 different computational approaches across a broad spectrum of experimental scenarios. We use synthetic 16S rDNA-seq count data generated from known underlying interaction networks, enabling systematic and controlled evaluation. The benchmark spans scenarios varying in numbers of species and samples, sample richness, sequencing depth, and interaction network structure. We evaluated the performance of the methods both in terms of accuracy of the inferred network, using various metrics from binary classification theory and metrics coming from network science, and in terms of computational burden. We found that methods have valid results only in a restricted, and often unrealistic, field of application, whereas their effectiveness diminishes with more realistic simulation parameters. Our findings highlight the limits of the co-occurrence-based approach and the need for further development in microbial ecological association inference.