We are pleased to share our latest preprint, now available on bioRxiv:
“Benchmarking precision matrix estimation methods for differential co-expression network analysis” Overmann M., Grabert G., Kacprowski T. bioRxiv (2026) | https://doi.org/10.64898/2026.04.13.716081
Understanding how genes interact — and how those interactions change in disease — is a central challenge in molecular biology. While classical differential expression analyses identify which genes change in abundance, they provide limited insight into the rewiring of regulatory relationships. Differential co-expression network analysis addresses this gap by modeling conditional dependencies between genes through partial correlations. A key computational step in this approach is the estimation of precision matrices (the inverse of covariance matrices), for which numerous methods have been proposed. However, a systematic and rigorous comparison of these methods under realistic conditions was lacking.
In this study, we present a comprehensive benchmarking framework for precision matrix estimation methods (PMEMs). Using simulated gene expression datasets with defined ground-truth correlation structures, we evaluated a broad panel of PMEMs across a wide range of conditions, including varying covariance structures, matrix densities, sample-size-to-dimension ratios, and sampling distributions. Our results show that method performance is strongly condition-dependent and that no single metric or simulation condition is sufficient to draw general conclusions. Among all tested approaches, GLassoElnetFast demonstrated the highest accuracy in recovering differential edges, though reliable inference still requires adequate signal-to-noise ratios and sufficient sample sizes.
This work highlights that previously published, less comprehensive evaluations may have led to misleading method recommendations. The simulation and benchmarking framework introduced here provides a reproducible foundation for the future development and evaluation of novel PMEMs.
The preprint is freely available under a CC BY license.