apeglm

Dependencies & Environment

Package-intrinsic requirements from the Bioconductor landing page — reproduce in any R environment.

• Version: 1.34.0 · Bioconductor: 3.23 · R: ≥ 4.6
• Depends: SummarizedExperiment, GenomicRanges, Rcpp
• Imports: emdbook, SummarizedExperiment, GenomicRanges, Rcpp
• System requirements: URL
• Install: BiocManager::install("apeglm")

When to Use

• Log Fold Change Shrinkage: Shrinking effect sizes for negative binomial GLMs in RNA-seq differential expression analysis (typically called via the DESeq2::lfcShrink wrapper).
• Beta-Binomial Modeling: Estimating and shrinking GLM coefficients for binomial or beta-binomial rate data, such as allele-specific count ratios.
• False Sign Rate Evaluation: Calculating local false sign rates (FSR) and s-values via svalue to control the probability of false-sign-or-small (FSOS) events.

When NOT to Use

• Direct Log2 Scale Requirements: If you require coefficients directly estimated on the log2 scale without manual conversion, note that apeglm natively estimates coefficients on the natural log scale (though DESeq2::lfcShrink converts these to log2 internally).
• Non-GLM Continuous Data: For standard linear modeling of continuous data (e.g., microarray log-intensities) where negative binomial or beta-binomial likelihoods do not apply.

Data Requirements

• Input Counts: A SummarizedExperiment object or a matrix of raw integer counts passed to the Y parameter.
• Design Matrix: A model matrix x specifying the experimental design (typically created using model.matrix).
• Dispersions: A numeric vector of dispersion parameters param (e.g., estimated using dispersions from DESeq2).
• MLE Estimates: A matrix mle containing maximum likelihood estimates of the coefficients and their standard errors, scaled to the natural log scale.
• Offsets: A matrix of offsets offset matching the dimensions of the count matrix (e.g., log of sizeFactors).

Key Parameters

• Y: The input data, which can be a SummarizedExperiment or a matrix of counts.
• x: The design matrix (model matrix) for the GLM.
• log.lik: The log-likelihood function to use (set to NULL when using fast C++ methods).
• param: Parameter values for the likelihood, such as dispersion estimates.
• coef: The integer index of the coefficient in the design matrix to shrink.
• threshold: A numeric value on the natural log scale specifying the threshold for false-sign-or-small (FSOS) event probability calculation.
• mle: A matrix of maximum likelihood estimates and their standard errors.
• method: The estimation method to use, such as "nbinomR", "nbinomCR", "nbinomC", "betabinCR", "betabinC", or "general".

Best Practices

• Use Integrated Wrappers: For standard RNA-seq workflows, call lfcShrink(dds, coef=2, type="apeglm") from the DESeq2 package to automatically handle scaling and parameter setup.
• Select Fast C++ Methods: Use method="nbinomCR" (which calculates posterior standard deviations) or method="nbinomC" (which returns only MAP coefficients) to achieve 10x to 100x speedups over the default "general" method.
• Scale Thresholds Correctly: When specifying a custom threshold in apeglm, multiply your log2-scale threshold by log(2) to convert it to the natural log scale required by the function.
• Iterative Beta-Binomial Fitting: When modeling ratios of counts, perform iterative estimation of MLE coefficients and beta-binomial dispersions using bbEstDisp.

Common Pitfalls

• Mismatched Coefficient Scales: Comparing raw apeglm output directly to log2 fold changes; fix this by multiplying the apeglm MAP estimates by log2(exp(1)) to convert them from the natural log scale to the log2 scale.
• Slow Execution on Large Datasets: Using the default "general" method on large count matrices; fix this by setting method="nbinomCR" or method="nbinomC" for negative binomial data.
• Line Search Failures: Encountering line search routine failures during beta-binomial dispersion estimation; fix this by ensuring your initial dispersion estimates and input counts are properly filtered and bounded.

Alternatives

• DESeq2 (normal method): Provides the original normal-prior-based shrinkage estimator via lfcShrink(type="normal").
• edgeR: Can be used as an alternative framework to estimate size factors and dispersions prior to shrinkage.
• zinbwave: Useful in combination with apeglm to model zero-inflated negative binomial data by estimating and down-weighting excess zero components.

Citations

• Zhu, A., Ibrahim, J. G., & Love, M. I. (2018). "Heavy-tailed prior distributions for sequence count data: adaptive shrinkage and its precision." Bioinformatics, 35(12), 2084-2092.
• Stephens, M. (2016). "False discovery rates: a new dataset may warrant a new look." Biostatistics, 18(2), 275–294.
• Love, M. I., Huber, W., & Anders, S. (2014). "Moderated estimation of fold change and dispersion for RNA-seq data with DESeq2." Genome Biology, 15(12), 550.

References

• Homepage: https://bioconductor.org/packages/apeglm
• Vignette: https://bioconductor.org/packages/release/bioc/vignettes/apeglm/inst/doc/apeglm.html

<!-- biomate-cta -->

---

Run this on BioMate

This skill is the knowledge layer — when, why, and how to use apeglm. To run this analysis on your own data with managed compute, automated QC, and reproducible outputs, use BioMate — free to start.

▶ Open apeglm on BioMate →