PLIERv2: bigger, better and faster.

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Traditional gene expression analysis struggles with complex phenotypes and lacks biological knowledge integration.
• Existing methods like PCA and NMF are limited in interpretability and noise correction.
• Semi-supervised methods like PLIER improve interpretability but are computationally inefficient for large datasets.

Purpose of the Study:

• Introduce PLIERv2, a scalable and efficient algorithm for biologically informed latent variable extraction.
• Overcome the computational limitations of the original PLIER method for large-scale transcriptomic data analysis.
• Enable comprehensive analysis of modern transcriptomic compendia for deeper biological insights.

Main Methods:

• Developed PLIERv2 with a two-phase algorithm: unsupervised PLIERbase initialization and PLIERfull regression using glmnet.
• Implemented rigorous internal cross-validation for tuning regularization parameters.
• Utilized memory-mapped matrices from the bigstatsr package for efficient on-disk data handling.

Main Results:

• PLIERv2 demonstrated 7x-41x speedups compared to PLIERv1.
• Successfully modeled hundreds of thousands of samples, surpassing PLIERv1's capacity.
• Maintained or improved biological specificity of latent variables, validated by tissue-alignment and pathway enrichment analyses.

Conclusions:

• PLIERv2 provides a scalable solution for analyzing large transcriptomic datasets, overcoming previous computational bottlenecks.
• The method enhances the interpretability and biological relevance of latent variables.
• PLIERv2 facilitates deeper understanding of gene regulatory networks and supports translational genomics research.