Robust Multivariate Methods for Analyzing Pathway Interactions and Pathway-Phenotype Associations in Bulk and Single-Cell RNA-seq Studies

Abstract Gene pathway analysis is a critical tool for understanding how groups of genes within biological pathways collectively contribute to complex human traits and diseases. Widely applied in both bulk and single-cell RNA- seq (scRNA-seq) studies, these methods have provided valuable insights into molecular processes and their effects on health. Despite the development of hundreds of pathway analysis tools, recent studies have exposed significant challenges that undermine the reliability of these approaches. For instance, substantial discrepancies between gene pathway scoring methods lead to inconsistencies in pathway network construction and cell type clustering, with false-positive rates in GSEA methods reaching as high as 40%. These issues create uncer- tainty for researchers in selecting appropriate tools for real-world applications, and attempts to establish a “gold standard” have proven computationally intensive and context-dependent. This project proposes a transformative framework to address these challenges. First, we introduce a latent- factor-based top-down model for unsupervised pathway interaction analysis, designed to efficiently capture population-specific variations while remaining robust to differences in pathway activity scoring algorithms. Sec- ond, we propose a set of multivariable latent factor regression models to identify pathway-phenotype associa- tions. These models enhance statistical power by leveraging individual gene expression data and accounting for pathway correlations, reducing false discoveries. Finally, we will apply these methods to extensive bulk- and single-cell RNA-seq datasets to validate existing findings, generate new insights, and tackle complex biological questions that have previously been out of reach.