SpiderLearner

SpiderLearner estimates Gaussian graphical models (GGMs) from multivariate biological data using an ensemble of candidate GGM estimation methods to produce robust partial-correlation-based networks for systems-level analysis.

Key Features:

• Ensemble integration: Integrates multiple candidate GGM estimation methods into a single ensemble estimator.
• Weighted-average ensemble: Constructs the ensemble estimate as a weighted average of results from each candidate method.
• GGM representation: Models nodes as biological variables (e.g., genes) and edges as partial correlations, where absence of an edge indicates zero partial correlation and edges can represent coexpression.
• Performance evaluation: Demonstrated superior or comparable performance to the best individual candidate methods in simulation studies.
• Application to gene expression: Applied to gene expression data from 260 ovarian cancer patients to estimate a GGM and analyze its community structure.
• Network-based risk score derivation: Facilitates derivation of a network-based risk score from community structure, demonstrated as a seven-gene score validated across six independent datasets.

Scientific Applications:

• Network inference: Inferring direct dependencies among genes and other biological variables via partial-correlation-based GGMs.
• Biomarker discovery and prediction: Identifying compact gene signatures and deriving network-based risk scores for disease risk prediction, exemplified by a seven-gene ovarian cancer score.
• Systems-level module analysis: Characterizing community structure in estimated networks to study modules of coexpression and genetic interactions.

Methodology:

Integrates multiple candidate GGM estimation methods and constructs an ensemble as a weighted average of their estimates; estimates GGMs where edges represent partial correlations; analyzes community structure of the estimated GGM; evaluates performance in simulation studies; applied to gene expression from 260 ovarian cancer patients and validated a network-based risk score across six independent datasets.