Local graph estimation with pathwise false discovery control.

Area of Science:

• Network inference
• Statistical modeling
• Systems biology

Background:

• Complex datasets often contain key target variables (e.g., biomarkers) within a larger system.
• Inferring the full network can obscure important local structures around these targets, hindering interpretability.
• Existing graph estimation methods may fail to accurately recover these localized relationships.

Purpose of the Study:

• To introduce a statistical framework for local graph estimation, focusing on inferring substructures around target variables.
• To address the limitations of traditional methods in uncovering local network patterns.
• To provide a robust method for analyzing complex systems with key variables of interest.

Main Methods:

• Developed a novel statistical framework called local graph estimation.
• Introduced pathwise feature selection (PFS) as a key method within this framework.
• PFS iteratively applies feature selection and uncertainty propagation along network paths.

Main Results:

• Demonstrated that traditional graph estimation methods often fail to recover local structure.
• PFS effectively estimates local subgraphs and provides finite-sample false discovery control.
• Applied PFS to diverse fields including public health, multiomics, and neuroimaging, recovering interpretable networks.

Conclusions:

• Local graph estimation, particularly using PFS, is a powerful approach for uncovering localized network structures.
• PFS successfully recovers established mechanisms and generates novel hypotheses across various scientific domains.
• The framework enhances interpretability in complex systems by focusing on target variable substructures.