Addressing biases in gene-set enrichment analysis: a case study of Alzheimer’s Disease

Artemy Bakulin, Noam Teyssier, Martin Kampmann, Matvei Khoroshkin, Hani GoodarzibioRxiv 2023 · DOI

Abstract

Inferring driving regulatory programs from comparative gene expression analysis is a cornerstone of systems biology. However, existing approaches are susceptible to biases inherent in real-world annotations, such as "promiscuous" genes that appear in numerous pathways. We present pyPAGE, an information-theoretic framework that uses conditional mutual information to account for these biases. Applying pyPAGE to Alzheimer’s Disease (AD) datasets, we identified cell-type-specific regulatory signatures and post-transcriptional programs that predict patient survival more accurately than traditional clinical metrics.

The Challenge of Annotation Bias

In systems biology, Gene-Set Enrichment Analysis (GSEA) is used to identify active biological "programs" in disease states. A significant hurdle is annotation bias: some genes are extensively studied or multifunctional, appearing in hundreds of annotations. If these "promiscuous" genes change expression, standard tools may report dozens of false-positive pathway perturbations. Conversely, these hubs can obscure signals from more specific, less-studied genes.

Sources of bias in biological annotations and the pyPAGE workflow.

As shown in the figure above, there is a clear correlation between a gene's citation index and its gene-set membership, confirming that our biological maps are skewed by research effort.

The pyPAGE Framework

To address this, pyPAGE extends the iPAGE toolset by shifting from simple Mutual Information (MI) to Conditional Mutual Information (CMI). Instead of asking if gene expression is linked to a pathway, pyPAGE asks if that link remains significant given the total number of pathways those genes belong to.

Key features include:

  • Distribution-Free: Handles both discrete (cell clusters) and continuous (fold-changes) data without assuming a bell curve.
  • Dual Regulation Detection: Identifies factors that simultaneously activate some targets while repressing others.
  • Robustness: Maintains performance even when data is subsampled or signals are damped.

Key Insights from Alzheimer’s Disease

The authors applied pyPAGE to bulk (MSBB) and single-cell (ROSMAP) RNA-seq datasets to study AD.

Transcriptional vs. Post-Transcriptional Regulation

A critical discovery was the divergence in regulatory mechanisms across cell types. AD-related changes in neurons are primarily driven at the transcriptional level. In contrast, changes in glial cells are heavily influenced by post-transcriptional regulation, such as RNA stability and splicing.

RBPs as Survival Predictors

RNA-binding proteins (RBPs) like LIN28A and HNRNPD emerged as vital players. Their activity levels were more predictive of patient survival than standard pathological scores like Braak (tau tangles) or Cerad (amyloid plaques). Patients with low RBP activity showed a hazard ratio of 4.84 for shorter survival (p=0.012).

Benchmarking and Performance

In simulations with added "promiscuity bias," pyPAGE significantly outperformed existing tools. While the performance of the original iPAGE dropped by 30%, pyPAGE remained stable with a PR-AUC of approximately 0.94.

Performance of pyPAGE compared to iPAGE and fgsea across various simulation scenarios.

The tool identified 15 Transcription Factors (TFs), 13 RBPs, and 8 miRNAs significantly associated with AD pathology. Notably, it revealed that SOX10 is specifically deregulated in oligodendrocytes, a detail often masked in bulk tissue analysis.

Conclusion and Future Directions

pyPAGE provides a robust, unbiased method for estimating the activity of transcriptional and post-transcriptional regulons. While it corrects for membership bias, it still depends on the quality of existing annotations. Future applications could focus on clinical stratification, using molecular subtypes to predict disease trajectory or select participants for targeted clinical trials.

Summary updated

Abstract

SUMMARY Inferring the driving regulatory programs from comparative analysis of gene expression data is a cornerstone of systems biology. Many computational frameworks were developed to address this problem, including our iPAGE ( i nformation-theoretic P athway A nalysis of G ene E xpression) toolset that uses information theory to detect non-random patterns of expression associated with given pathways or regulons 1 . Our recent observations, however, indicate that existing approaches are susceptible to the biases and artifacts that are inherent to most real world annotations. To address this, we have extended our information-theoretic framework to account for specific biases in biological networks using the concept of conditional information. This novel implementation, called pyPAGE, provides an unbiased way for the estimation of the activity of transcriptional and post-transcriptional regulons. To showcase pyPAGE, we performed a comprehensive analysis of regulatory perturbations that underlie the molecular etiology of Alzheimer’s disease (AD). pyPAGE successfully recapitulated several known AD-associated gene expression programs. We also discovered several additional regulons whose differential activity is significantly associated with AD. We further explored how these regulators relate to pathological processes in AD through cell-type specific analysis of single cell gene expression datasets.
BibTeX
@misc{bakulin2023addressing,
  author = {Artemy Bakulin and Noam Teyssier and Martin Kampmann and Matvei Khoroshkin and Hani Goodarzi},
  title = {{Addressing biases in gene-set enrichment analysis: a case study of Alzheimer’s Disease}},
  year = {2023},
  doi = {10.1101/2023.08.13.553133},
  howpublished = {bioRxiv},
}
Figure