Abstract
Gene expression microarray technologies provide the simultaneous measurements of a large number of genes. Typical analyses of such data focus on the individual genes, but recent work has demonstrated that evaluating changes in expression across predefined sets of genes often increases statistical power and produces more robust results. We introduce a new methodology for identifying gene sets that are differentially expressed under varying experimental conditions. Our approach uses a hierarchical Bayesian framework where a hyperparameter measures the significance of each gene set. Using simulated data, we compare our proposed method to alternative approaches, such as Gene Set Enrichment Analysis (GSEA) and Gene Set Analysis (GSA). Our approach provides the best overall performance. We also discuss the application of our method to experimental data based on p53 mutation status.
Introduction
In high-throughput biology, researchers often face a "needle in a haystack" problem: measuring the expression of tens of thousands of genes across only a handful of samples. Traditional statistical tests look at genes individually, which often leads to lists of significant genes that lack a unifying biological theme or fail to be reproducible across different laboratories.
To address this, researchers moved toward Gene Set Analysis (GSA), which groups genes into predefined sets based on known biological pathways. However, existing frequentist methods like GSEA (Gene Set Enrichment Analysis) have notable flaws. Some are overly sensitive to changes in the middle of the distribution rather than the tails, while others struggle when genes within the same pathway move in opposite directions (discordant regulation).
Methodology: The BGSA Framework
The authors propose a hierarchical Bayesian model to evaluate gene sets. This framework shifts the focus from individual gene significance to the collective behavior of gene sets.
The Core Model and Hierarchical Structure
For each gene, the observed expression is modeled as a function of an overall mean, a treatment effect ($\beta$), and random noise. Instead of treating each gene's effect independently, the model assumes that all genes in a specific set share a common variance hyperparameter, $\tau^2$.
If a gene set is not significant, the model "shrinks" the $\beta$ values toward zero, and $\tau^2$ remains small. Conversely, if a gene set is significant, $\tau^2$ becomes large to accommodate the high activity of the genes within it. This hierarchical structure allows the model to distinguish between active and inactive biological pathways effectively.
Mixture Priors and Computation
To automate the selection of significant sets, the model employs a "spike-and-slab" style mixture prior. This allows the model to categorize each gene set into either a "null" (insignificant) or "alternative" (significant) group. Due to the complexity of the posterior distribution, Markov Chain Monte Carlo (MCMC) sampling—specifically a combination of Gibbs and slice sampling—is used for estimation.
Results and Discussion
The Bayesian Gene Set Analysis (BGSA) approach was validated using six simulated scenarios and a real-world p53 mutation dataset.
Performance in Simulations
BGSA demonstrated superior performance in five out of six simulation scenarios, achieving higher Area Under the ROC Curve (AUC) values than GSEA, GSA, and standard Linear Mixed Effects (LME) models. It proved particularly robust in scenarios involving non-normal data and correlated gene expressions, where frequentist assumptions often break down.
Biological Discovery in p53 Data
In the p53 study, which involved 50 cancer cell lines and 522 curated gene sets, BGSA identified several critical pathways (such as ATM and BAD) known to interact with p53. These pathways were missed by other popular methods, demonstrating BGSA's sensitivity to biologically relevant signals.
Key Insights
- Automatic Restandardization: BGSA handles "restandardization" naturally by comparing the variance of a specific set against global background noise, preventing sets from being flagged simply due to overall experimental noise.
- Directional Agnosticism: Unlike methods that favor concordant regulation, BGSA remains effective even when pathways contain both up-regulated and down-regulated genes.
Conclusion
BGSA provides a powerful and robust alternative to frequentist gene set analysis methods. While its primary drawback is computational cost—taking approximately 20 minutes for the p53 dataset compared to seconds for permutation-based methods—the gain in statistical power and biological insight is significant. Future work will focus on optimizing sampling algorithms and extending the model to more complex experimental designs.
Summary updated