Identifying patterns differing between high-dimensional datasets with generalized contrastive PCA

Area of Science:

• Bioinformatics
• Computational Biology
• Data Science

Background:

• High-dimensional biological data are increasingly common.
• Comparing datasets from different conditions is crucial for biological discovery.
• Existing methods like contrastive PCA (cPCA) have limitations, including hyperparameter tuning and asymmetric treatment of conditions.

Purpose of the Study:

• To develop a novel, flexible, and hyperparameter-free dimensionality reduction technique for comparing high-dimensional biological datasets.
• To address the limitations of cPCA, enabling symmetric and robust comparison of experimental conditions.

Main Methods:

• Development of generalized contrastive PCA (gcPCA).
• Theoretical analysis explaining the hyperparameter requirement in cPCA and how gcPCA avoids it.
• Creation of an open-source gcPCA toolbox with Python and MATLAB implementations.

Main Results:

• gcPCA provides a hyperparameter-free and symmetric approach to dimensionality reduction for comparative analysis.
• Demonstrated utility in analyzing diverse high-dimensional biological data.
• Successfully detected unsupervised hippocampal replay in neurophysiological data and revealed type II diabetes heterogeneity in single-cell RNA sequencing data.

Conclusions:

• gcPCA is a fast, robust, and user-friendly method for comparative analysis of high-dimensional biological data.
• Facilitates gaining new insights into complex biological phenomena.
• Provides a valuable resource for the biological sciences.