TestNet-PA: Sampling-Zero-Aware Inference of Microbial Presence-Absence Networks

Abstract

Microbial presence-absence networks, also referred to as co-occurrence networks, characterize patterns of coexistence and incompatibility among taxa. However, sequencing-based presence--absence analysis can be compromised by pervasive 'sampling zeros' in microbiome data, causing widely used methods such as Fisher's exact test to identify spurious co-occurrence dependence. To our knowledge, no existing method is designed to account for sampling zeros in this setting. We introduce TestNet-PA, a method for testing pairwise microbial co-occurrence dependence. TestNet-PA models each observed zero as arising from a mixture of 'structural' and 'sampling' zeros, with the mixing probabilities governed by library size and relative abundance. Because the relative abundance underlying an observed zero is unobserved, we impute it by borrowing information from other samples. We then construct a quasi-likelihood and develop an EM algorithm to infer dependence between the underlying true presence--absence statuses. Extensive simulations based on real gut, vaginal, and oral microbiome data templates showed that TestNet-PA consistently controlled the FDR while maintaining good sensitivity. By contrast, Fisher's exact test generally yielded inflated FDR. Analyses of the three real microbiome datasets further illustrated how accounting for sampling zeros can affect the inferred co-occurrence networks.