Bayesian testing of many hypotheses $times$ many genes: A study of sleep apnea

Substantial statistical research has recently been devoted to the analysis of large-scale microarray experiments which provide a measure of the simultaneous expression of thousands of genes in a particular condition. A typical goal is the comparison of gene expression between two conditions (e.g., diseased vs. nondiseased) to detect genes which show differential expression. Classical hypothesis testing procedures have been applied to this problem and more recent work has employed sophisticated models that allow for the sharing of information across genes. However, many recent gene expression studies have an experimental design with several conditions that requires an even more involved hypothesis testing approach. In this paper, we use a hierarchical Bayesian model to address the situation where there are many hypotheses that must be simultaneously tested for each gene. In addition to having many hypotheses within each gene, our analysis also addresses the more typical multiple comparison issue of testing many genes simultaneously. We illustrate our approach with an application to a study of genes involved in obstructive sleep apnea in humans.

💡 Research Summary

The paper tackles a pervasive challenge in modern high‑throughput genomics: simultaneously testing many hypotheses for each of thousands of genes. Traditional approaches treat each gene–hypothesis pair as an independent two‑sample test, applying a false‑discovery‑rate (FDR) correction across genes. This strategy quickly becomes inadequate when the experimental design includes several conditions, time points, or treatment groups that generate a combinatorial set of hypotheses for each gene. To address this, the authors develop a hierarchical Bayesian model that jointly models all hypotheses within a gene while also borrowing strength across genes.

At the first level, the expression data for gene g under condition c are modeled as normal with mean μ_gc and a common variance. Each scientific question (e.g., “expression in severe OSA differs from mild OSA”) is encoded as a linear constraint on the μ_gc’s, yielding a set of K mutually exclusive hypotheses per gene. Prior probabilities for these hypotheses are placed on a simplex (a Dirichlet‑Bernoulli structure), allowing incorporation of external knowledge such as pathway expectations.

The second, hierarchical level introduces hyper‑parameters (μ_k, τ_k) that govern the distribution of the hypothesis‑specific means across all genes. This partial‑pooling mechanism shrinks noisy gene‑specific estimates toward a shared distribution, dramatically improving power for genes with few replicates while preserving genuine heterogeneity. The full joint posterior over all μ_gc, hypothesis indicators, and hyper‑parameters is explored using a hybrid Gibbs/Metropolis‑Hastings Markov‑chain Monte Carlo algorithm.

From the posterior, the authors compute for each gene–hypothesis pair the probability that the hypothesis is true, P(H_k | data). Rather than relying on p‑values, they adopt a Bayesian FDR framework: a global target false‑discovery rate α is set, and all pairs with posterior probability ≥ 1 – α are declared discoveries. Because the posterior probabilities already respect the logical exclusivity of hypotheses (a gene cannot simultaneously satisfy contradictory statements), the resulting list of discoveries is internally consistent.

The methodology is illustrated on a human obstructive sleep apnea (OSA) study. Blood samples from subjects spanning four clinical categories (mild, moderate, severe OSA, and post‑treatment) were profiled on a microarray platform. Fifteen composite hypotheses describing all pairwise and higher‑order contrasts among the four conditions were defined. Using the traditional two‑sample t‑test with Benjamini‑Hochberg correction, roughly 120 genes were flagged as differentially expressed. In contrast, the hierarchical Bayesian analysis identified about 185 genes, many of which belong to inflammation, oxidative‑stress, and metabolic pathways previously implicated in OSA but missed by the classical approach. The Bayesian model also provided posterior probabilities for each specific contrast, enabling researchers to rank genes not only by significance but by the strength of evidence for each biological question.

Key insights from the paper include:

1. Information sharing across genes via hierarchical priors substantially boosts detection power, especially for low‑expressed or sparsely replicated genes.
2. Explicit modeling of multiple hypotheses per gene respects the logical structure of complex experimental designs and avoids the multiple‑testing inflation that would occur if each contrast were analyzed separately.
3. Bayesian FDR control offers a more intuitive decision rule (probability of truth) while still bounding the expected proportion of false discoveries.
4. Flexibility to incorporate prior knowledge through the Dirichlet‑Bernoulli prior on hypothesis weights, which can be tuned to reflect pathway‑level expectations or previous studies.

The authors conclude that the hierarchical Bayesian framework is broadly applicable to any “omics” setting where many conditions generate a combinatorial hypothesis space—such as time‑course RNA‑seq, multi‑drug screens, or integrative epigenomics. By unifying hypothesis testing across genes and conditions, the approach delivers higher statistical power, clearer interpretability, and a principled way to control false discoveries, paving the way for more nuanced biological insights from large‑scale genomic experiments.