Multiple augmented reduced rank regression for pan-cancer analysis.

Area of Science:

• Statistics
• Bioinformatics
• Genomics

Background:

• Combining multiple datasets enhances statistical power and efficiency over separate analyses.
• High-dimensional data across cohorts requires methods to correctly address complex variation architectures.
• Existing methods like reduced rank regression and matrix factorization have limitations in handling multi-cohort variations.

Purpose of the Study:

• To introduce multiple augmented reduced rank regression (maRRR), a novel flexible method for analyzing high-dimensional data across multiple cohorts.
• To concurrently learn both covariate-driven and auxiliary structured variations.
• To provide a unified framework that subsumes existing methods and introduces a new single-dataset approach (aRRR).

Main Methods:

• maRRR employs a structured nuclear norm objective motivated by random matrix theory.
• The framework allows for regression or factorization terms to be shared or specific across any number of cohorts.
• It integrates regression and factorization for both single and multiple datasets.

Main Results:

• Simulations show substantial power gains by combining multiple datasets and accounting for structured variations.
• maRRR performs well in prediction and imputation tasks on held-out data.
• The method successfully applied to pan-cancer gene expression data from The Cancer Genome Atlas.

Conclusions:

• maRRR offers a powerful and flexible approach for multi-cohort high-dimensional data analysis.
• The method provides new insights into mutation-driven and auxiliary variations in cancer.
• maRRR enhances the scientific informativeness of integrated data analyses.