Using null models to infer microbial co-occurrence networks

Abstract

Although microbial communities are ubiquitous in nature, relatively little is known about the structural and functional roles of their constituent organisms’ underlying interactions. A common approach to study such questions begins with extracting a network of statistically significant pairwise co-occurrences from a matrix of observed operational taxonomic unit (OTU) abundances across sites. The structure of this network is assumed to encode information about ecological interactions and processes, resistance to perturbation, and the identity of keystone species. However, common methods for identifying these pairwise interactions can contaminate the network with spurious patterns that obscure true ecological signals. Here, we describe this problem in detail and develop a solution that incorporates null models to distinguish ecological signals from statistical noise. We apply these methods to the initial OTU abundance matrix and to the extracted network. We demonstrate this approach by applying it to a large soil microbiome data set and show that many previously reported patterns for these data are statistical artifacts. In contrast, we find the frequency of three-way interactions among microbial OTUs to be highly statistically significant. These results demonstrate the importance of using appropriate null models when studying observational microbiome data, and suggest that extracting and characterizing three-way interactions among OTUs is a promising direction for unraveling the structure and function of microbial ecosystems.

Incorporating phylogenetic metrics to microbial co‐occurrence networks based on amplicon sequences to discern community assembly processes

“…In turn, the resistant species closes the loop by outcompeting the toxin-producing species since it does not incur in the cost of toxin production (Hibbing et al, 2010). Despite these and other limitations that need to be taken into account when analysing co-occurrence patterns (Connor et al, 2017;Freilich et al, 2018), we detected a phylogenetic imprint that depended on the assembly mechanism and was consistent across sites.…”

“…Microbial co-occurrence networks, mainly reconstructed from amplicon sequencing data, are being increasingly used to infer significant associations between pairs of co-occurring taxa and often ascribed to biological interactions Fuhrman, Cram, & Needham, 2015;Ho et al, 2016;Pérez-Valera et al, 2017). Critical voices, however, call for caution when analysing and interpreting co-occurrence networks in order to avoid the description of ecologically meaningless interactions (Barner et al, 2018;Connor, Barberán, & Clauset, 2017;Freilich et al, 2018). Here, we use co-occurrence networks to identify associations between pairs of bacterial taxa across multiple assemblages and test their significance against a null model.…”

Relating network analyses to phylogenetic relatedness to infer protistan co-occurrences and co-exclusions in marine and terrestrial environments

“…To infer protistan co-occurrences and co-exclusions from the marine and terrestrial datasets, networks were constructed using OTUs following Connor et al (2017). This method infer positive correlations (co-occurrences), which was expanded here to also infer negative correlations (coexclusions).…”

“…First, to reduce computational load, OTUs occurring in less than 30% of marine and 10% of terrestrial samples were removed as well as samples with less than 20% of median read counts per sample in the terrestrial dataset. Low occurrence OTUs would never show any significant cooccurrence or co-exclusion using this method (Connor et al, 2017). The OTUs which passed the occurrence filter are later referred as the candidate OTUs.…”

“…This null Chung-Lu model produced networks with characteristics (e.g. modularity, diameter, clustering coefficient) more similar to natural networks compared to the most widely used null Erdős-Rényi model (Connor et al, 2017), and thus minimizes the number of parameters modified compared to the observed network. The random networks were made using the "sample_fitness" function in the R igraph package (Csardi & Nepusz, 2006).…”