Benchmark of 18 methods reveals limits of inferring microbial interactions from 16S count data

Inferring ecological interactions within microbial communities is central to understanding how these communities assemble and evolve, and to enabling targeted modification of their functions. Many computational approaches have been developed to derive such associations, and co-occurrence methods are among the most widely used: they infer associations from metagenomic abundance data by measuring the statistical dependence between species counts across different sampled populations. However, the limitations of culture-based experiments have made it difficult to assess how valid these methods actually are, which has contributed to their scarce application in metagenomics studies.

A new benchmarking study by Piero Mariotto, Matteo Baldan, Barbara Di Camillo and Giacomo Baruzzo, published in PLOS Computational Biology, set out to provide a comprehensive assessment of network inference method performance. The researchers benchmarked 18 different computational approaches across a broad spectrum of experimental scenarios. Rather than relying on real data, where the true interaction network is unknown, they used synthetic 16S rDNA-seq count data generated from known underlying interaction networks, enabling a systematic and controlled evaluation of how well each method recovers the true structure.

The benchmark spanned scenarios varying in the numbers of species and samples, sample richness, sequencing depth, and the structure of the interaction network. The team evaluated performance in two ways: the accuracy of the inferred network, assessed with various metrics drawn from binary classification theory as well as metrics from network science, and the computational burden each method imposed. This dual evaluation allowed the authors to judge not only whether the methods produce correct networks, but also whether they are practical at realistic scale.

The results were sobering. The researchers found that the methods deliver valid results only within a restricted, and often unrealistic, field of application, and that their effectiveness diminishes as the simulation parameters become more realistic. The authors emphasize that these findings highlight the limits of the co-occurrence-based approach to microbial association inference and point to a clear need for further development in this area before such methods can be applied with confidence to real metagenomic data.

Subject of Research: Benchmarking computational methods for inferring microbial interaction networks from 16S rDNA-seq count data

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

Article References: Mariotto, P., Baldan, M., Di Camillo, B., & Baruzzo, G. (2026). From 16S rDNA-seq count data to microbial interactions: A comprehensive benchmark of co-occurrence network inference methods. PLOS Computational Biology, 22(10), e1014898. https://doi.org/10.1371/journal.pcbi.1014898

Image Credits: AI Generated

DOI: 10.1371/journal.pcbi.1014898

Keywords: microbiome, co-occurrence networks, 16S rDNA-seq, network inference, benchmarking, metagenomics, microbial ecology, computational biology, synthetic data, PLOS Computational Biology, rDNA-seq, count