Heterogeneity-Aware Gaussian Graphical Models for High- Dimensional Omics: A Narrative Review

Abstract

In high-dimensional omics data, sparse precision matrices of Gaussian graphical models (GGMs) can be used to encode conditional dependence among molecular features to allow for interpretable network inference. This narrative review integrates some of the heterogeneity-aware extensions of GGMs and categorizes them based on the way in which the information on heterogeneity is incorporated into the model. The review takes into account joint estimation of networks in known groups or conditions, covariate-indexed networks or individualized networks, as well as latent-subgroup models for unobserved groups. Known-group approaches use information from other networks of the same kind of groups, covariate-indexed approaches model variation in networks as a function of observed covariates, and latent-subgroup approaches infer both the network structure and the group membership. This review shows how each representation relates to its information requirements and limitations, leading to the choice of models and prioritisation of scalable inference, uncertainty quantification and external biological validation. This review establishes a unified tripartite taxonomy based on the observability of heterogeneity drivers, which systematically compares methodological trade-offs absent from prior fragmented reviews of Gaussian graphical models.