Discovering and deciphering relationships across disparate data modalities.

Area of Science:

• Computational biology
• Data science
• Network analysis

Background:

• Assessing relationships between diverse data types (e.g., genome, connectome, disease status) is crucial.
• Existing methods for detecting data dependencies often lack interpretability and require large sample sizes.
• There is a need for efficient and interpretable methods to uncover complex relationships in high-dimensional biological data.

Purpose of the Study:

• To introduce Multiscale Graph Correlation (MGC), a novel dependence test for analyzing relationships between disparate data properties.
• To demonstrate MGC's superior statistical power and efficiency compared to existing methods.
• To showcase MGC's ability to characterize the underlying geometry of data relationships and guide experimental design.

Main Methods:

• MGC integrates k-nearest neighbors, kernel methods, and multiscale analysis.
• The approach was benchmarked against existing methods on datasets with varying dimensionality (1-1000) and nonlinear relationships.
• Statistical power and computational efficiency were key evaluation metrics.

Main Results:

• MGC achieved higher statistical power than existing methods, requiring significantly fewer samples.
• The method demonstrated effectiveness across a range of data dimensionalities and complexities.
• MGC successfully identified dependencies in real-world biological datasets, including brain imaging and cancer genetics.

Conclusions:

• MGC provides a powerful, interpretable, and computationally efficient tool for detecting dependencies between complex data properties.
• The method's ability to characterize latent geometric structures offers unique insights beyond simple correlation.
• MGC has practical applications in fields like genomics, neuroimaging, and precision medicine, guiding future research directions.