TL;DR
This paper introduces pyPAGE, a novel computational framework designed to address inherent biases in gene-set enrichment analysis (GSEA). By accounting for annotation biases and using conditional mutual information, pyPAGE offers a more robust and accurate way to identify regulatory programs underlying complex diseases, as demonstrated in a case study on Alzheimer's disease.
What Problem Does This Paper Address?
Inferring the complex regulatory networks that govern gene expression is fundamental to systems biology. Gene-set enrichment analysis (GSEA) is a popular approach for this, but existing methods often assume that all genes within a gene-set (like a pathway or a regulon) contribute equally. This assumption is flawed because gene-set annotations are often biased. For instance, genes that are more extensively studied or encode abundant proteins tend to be included in more gene-sets, creating an imbalance. These biases can lead to inaccurate identification of regulatory programs and obscure true biological signals. The authors aim to overcome these limitations by developing a method that explicitly models and corrects for these annotation biases.
Methods
The core of pyPAGE is its use of conditional mutual information to analyze the relationship between gene expression data and gene-set annotations. Unlike previous methods that might treat all genes equally, pyPAGE accounts for the "gene-set membership profile" of each gene. This means it considers how many gene-sets a gene belongs to when assessing its contribution to a specific pathway's deregulation.
The framework involves three key steps:
- Data Preprocessing: Differential gene expression profiles are discretized into bins, and gene-set annotations are converted into a binary format.
- Conditional Mutual Information Estimation: The dependency between gene expression and a gene-set is calculated, conditioned on the gene-set membership profile. This is where the bias correction happens.
- Significance Estimation: A non-parametric permutation test is used to assess the statistical significance of the inferred associations. A redundancy test is also applied to ensure that newly identified gene-sets add significant information.
pyPAGE is designed to be flexible, working with various data modalities (bulk and single-cell RNA-seq, differential transcript stability) and gene-set types (pathways, transcription factor regulons, RNA-binding protein regulons, miRNA targets).
Datasets
The study utilizes several publicly available datasets for its analysis and validation:
- MSBB (Mount Sinai Brain Bank) dataset: Bulk RNA-sequencing data from the Brodmann Area 36 region of the brain, used for initial analysis of Alzheimer's disease (AD) pathology.
- ROSMAP dataset: Single-cell RNA-sequencing data from human prefrontal cortex, used for cell-type specific analysis of AD.
- Visium platform dataset: Spatial transcriptomics data from the middle temporal gyrus, used to infer spatial regulation patterns.
- ENCODE eCLIP data: Used for characterizing RNA-binding protein (RBP) functions.
- Simulated datasets: Artificially generated datasets were created to rigorously test pyPAGE's performance, specifically evaluating its robustness to annotation biases and its ability to detect complex regulatory patterns (monotonic and non-monotonic). These simulations involved controlling for "ground truth" signals and introducing various levels of bias.
Key Insights and Ideas
- Gene-set annotations are inherently biased and form scale-free networks: The paper demonstrates that gene-set memberships, whether for pathways, TF targets, or miRNA targets, often follow a power-law distribution, characteristic of scale-free networks. This is not a normal distribution as often assumed.
- Bias in annotations significantly impacts GSEA results: The study shows that existing methods (like iPAGE and fgsea) perform poorly when faced with biased annotations, leading to a substantial drop in accuracy. pyPAGE, by design, mitigates this impact.
- Post-transcriptional regulation plays a crucial role in AD heterogeneity: The analysis reveals that post-transcriptional programs, particularly those involving RBPs, are more strongly associated with patient survival in AD than transcriptional programs. This highlights an underexplored area for therapeutic intervention and patient stratification.
- Cell-type specific and spatial patterns of deregulation are critical: AD pathology manifests differently across various brain cell types and cortical layers. pyPAGE's ability to analyze these granular datasets provides a more nuanced understanding of disease mechanisms.
Key Results
- Improved performance with biased data: In simulations, pyPAGE's performance (measured by Precision-Recall AUC) remained largely unchanged when using biased annotations, while iPAGE's performance dropped by over 30%.
- Robustness to noise and sparsity: pyPAGE demonstrated robustness to subsampling of input data (recovering over 80% of perturbed gene-sets even with only 20% of original data) and signal reduction, making it suitable for challenging datasets like single-cell RNA-seq.
- Identification of AD-associated regulators: The study identified 15 transcription factors, 16 RNA-binding proteins, and 8 miRNAs significantly associated with AD pathology.
- Clinical relevance of RBP activity: Low activity of certain RBP regulons was significantly associated with shorter survival in AD patients, suggesting their potential as biomarkers for patient stratification.
- Differential roles of transcriptional vs. post-transcriptional regulation in AD: Neuronal deregulation in AD was primarily driven by transcriptional changes, while glial deregulation was largely attributed to aberrant post-transcriptional regulation.
Limitations and Future Work
- Computational cost: While pyPAGE is implemented efficiently using numba, complex analyses involving large datasets and numerous gene-sets can still be computationally intensive.
- Annotation quality: The accuracy of pyPAGE, like any GSEA method, is dependent on the quality and completeness of the underlying gene-set annotations.
- Future Directions: The authors suggest that pyPAGE's ability to identify robust and specific signals, even in noisy or sparse data, makes it highly promising for single-cell RNA-seq analysis. Further exploration of the identified RBPs and miRNAs as therapeutic targets and biomarkers for AD is also a clear avenue for future work. The framework's generalizability suggests its application to other complex diseases.
Summary updated