Multivariate biological data are often modeled using networks in which nodes represent a biological variable (e.g., genes) and edges represent associations (e.g., coexpression). A Gaussian graphical model (GGM), or partial correlation network, is an undirected graphical model in which a weighted edge between two nodes represents the magnitude of their partial correlation, and the absence of an edge indicates zero partial correlation. A GGM provides a roadmap of direct dependencies between variables, providing a valuable systems-level perspective. Many methods exist for estimating GGMs; estimated GGMs are typically highly sensitive to choice of method, posing an outstanding statistical challenge. We address this challenge by developing SpiderLearner, a tool that combines a range of candidate GGM estimation methods to construct an ensemble estimate as a weighted average of results from each candidate. In simulation studies, SpiderLearner performs better than or comparably to the best of the candidate methods. We apply SpiderLearner to estimate a GGM for gene expression in a publicly available dataset of 260 ovarian cancer patients. Using the community structure of the GGM, we develop a network-based risk score which we validate in six independent datasets. The risk score requires only seven genes, each of which has important biological function. Our method is flexible, extensible, and has demonstrated potential to identify de novo biomarkers for complex diseases. An open-source implementation of our method is available at https://github.com/katehoffshutta/SpiderLearner.