Empirical bayes variance estimation

empirical-bayes-variance-estimation

Summary

Empirical Bayesian hyperparameter estimation stabilizes gene-wise variance estimates by fitting a prior distribution over variances, enabling computation of moderated t-statistics and B-statistics that remain robust even with small sample sizes in microarray and RNA-seq studies.

When to use

Apply this skill when you have a fitted linear model (lmFit object) from microarray or RNA-seq count data and need to compute stable variance estimates and differential expression statistics despite having few biological replicates or small numbers of arrays. Use it as a post-hoc step after linear model fitting to borrow information across genes and stabilize per-gene variance estimates before hypothesis testing.

When NOT to use

• Input expression data has not yet been fitted to a linear model; use lmFit first.
• You require raw, unadjusted per-gene variance estimates without prior regularization; empirical Bayes shrinkage is inappropriate for variance-only inference.
• Sample size is very large (e.g., >100 arrays per group); shrinkage will be minimal and standard t-statistics may be preferable for computational efficiency.

Inputs

• lmFit object (fitted linear model from limma)
• gene expression matrix (microarray intensity or RNA-seq counts)
• design matrix specifying experimental conditions

Outputs

• eBayes object containing moderated t-statistics
• B-statistics (log-odds of differential expression)
• adjusted p-values (e.g., Benjamini-Hochberg FDR)
• results table with gene identifiers, log-fold changes, moderated t-statistics, and significance metrics

How to apply

After fitting a linear model using lmFit on your expression data, apply the eBayes function to estimate hyperparameters (prior mean and degrees of freedom) of a conjugate prior distribution over gene-wise variances. This pooled variance estimate is then used to compute moderated t-statistics by combining the observed gene-specific variance with the prior estimate, weighted by degrees of freedom. The resulting posterior estimates remain stable even when individual genes have few observations. Extract the moderated t-statistics, B-statistics (log-odds of differential expression), and adjusted p-values from the fitted eBayes object. The method is most effective when array numbers are small, as the prior becomes stronger relative to observed variance and shrinkage is more pronounced.

Related tools

• limma (R package providing lmFit for linear model fitting and eBayes for empirical Bayesian variance estimation and moderated t-statistic computation) — https://github.com/bioc/limma
• R (Statistical programming environment in which limma functions are executed)

Examples

library(limma); fit <- lmFit(expressionMatrix, designMatrix); efit <- eBayes(fit); results <- topTable(efit, adjust.method='BH', number=Inf)

Evaluation signals

• eBayes object contains non-null s2.prior (prior variance estimate) and df.prior (prior degrees of freedom) fields, indicating successful hyperparameter estimation.
• Moderated t-statistics have smaller magnitudes and larger p-values than ordinary t-statistics due to variance shrinkage, reflecting increased stability.
• B-statistics (log posterior odds) are computed and finite for all genes, with ranking of genes by B-statistic reflecting combined evidence from log-fold change and variance stability.
• Adjusted p-values (e.g., using p.adjust on $p.value column) are monotonically non-decreasing when genes are sorted by significance, meeting FDR control requirements.
• Genes with large log-fold changes but high variance shrink less; genes with small fold changes but low variance shrink more, demonstrating correct weighting by both effect size and reliability.

Limitations

• Empirical Bayes assumes a common prior distribution across all genes; genes with fundamentally different variance structures may be mis-estimated.
• Method requires at least 2 arrays per group to estimate degrees of freedom; with very few replicates the prior becomes dominated by the marginal distribution and may over-regularize.
• Moderated t-statistics and B-statistics are less interpretable than classical t-statistics when the prior assumptions are violated (e.g., heavy-tailed or multi-modal variance distributions).
• eBayes does not account for gene-specific technical covariates (e.g., GC content, transcript length) that may affect variance; separate variance modeling may be needed for such structured heterogeneity.

Evidence

• [other] Empirical Bayesian methods in limma estimate hyperparameters of the prior distribution over gene-wise variances, enabling computation of moderated t-statistics and B-statistics that remain stable even with small numbers of arrays.: "Empirical Bayesian methods are used to provide stable results even when the number of arrays is small."
• [other] Moderated t-statistics and B-statistics are extracted from the eBayes fitted object.: "Extract moderated t-statistics and B-statistics (log-odds of differential expression) from the fitted eBayes object."
• [other] The linear model and differential expression functions apply to microarrays, quantitative PCR, RNA-seq and proteomics.: "The linear model and differential expression functions apply to a wide variety of gene expression technologies including microarrays (single-channel or two-color), quantitative PCR, RNA-seq or"
• [other] Limma is an R package for the analysis of gene expression data, especially the use of linear models for analysing designed experiments.: "Limma is an R package for the analysis of gene expression data, especially the use of linear models for analysing designed experiments"
• [other] A results table is generated containing gene identifiers, log-fold changes, moderated t-statistics, adjusted p-values, and B-statistics, sorted by significance.: "Generate a results table containing gene identifiers, log-fold changes, moderated t-statistics, adjusted p-values, and B-statistics, sorted by significance."