Nonlinear network inference reveals two independent axes of ecological organisation in the rumen microbiome

Background The rumen microbiome is shaped by numerous external factors such as diet and host genetics, but these drivers act through internal microbial dynamics, including cross-feeding, substrate competition, and threshold-driven community shifts that give rise to non-linear interactions among microbes. Standard network inference approaches detect linear co-abundance patterns but may miss nonlinear dependencies that reflect these internal ecological processes. Results We developed FANCY ( F requency A nd N onlinear C orrelation h Y brid), a network inference framework that integrates mutual information-based edge detection with distance correlation stability assessment through a multiplicative hybrid score. Applied to 2,178 metagenome-assembled genomes (MAGs) from 321 beef steers across four breeds and two diets, FANCY shared 72% of MAG association (edges) with linear methods while contributing an additional 13% of associations reflecting nonlinear dependencies. FANCY-derived nonlinear edges reorganised microbial taxa into co-abundance modules with distinct external drivers. Two modules were validated against independent biological classifications from prior studies: Module 1 captured protozoal community type structure while Module 2 captured a host-genetics link to microbial methane metabolism. The axes were independent and involved largely non-overlapping taxa. Neither structure was resolved by commonly used linear network inference methods. Conclusions FANCY reveals hidden ecological organisation in the rumen that conventional linear methods miss, resolving protozoal community type and host-genetic influences as independent axes. The framework can process 2,178 MAGs across 321 samples on a laptop and is generalisable to any high-dimensional abundance dataset. FANCY is available as an R package through Bioconductor and Github https://github.com/wala-github/Fancy.