Fast effect size shrinkage software for beta-binomial models of allelic imbalance

Area of Science:

• Genomics
• Bioinformatics
• Statistical Genetics

Background:

• Allelic imbalance, differential gene expression between alleles, reveals cis-regulation and epigenetic differences.
• Accurate quantification of allelic expression proportions is vital but difficult with low read counts or small sample sizes.
• Traditional methods like filtering small-count genes may discard important data on allelic imbalance.

Purpose of the Study:

• To evaluate and compare the accuracy of different estimators for allelic expression proportions.
• To introduce and assess the performance of the `apeglm` package for analyzing allelic imbalance.

Main Methods:

• Compared four estimators: maximum likelihood, pseudocounts, approximate posterior estimation of GLM coefficients (apeglm), and adaptive shrinkage (ash).
• Developed C++ code for rapid calculation of ML and apeglm estimates, integrated into the `apeglm` R/Bioconductor package.
• Evaluated methods using two simulations and one real data set, comparing `apeglm` to other beta-binomial model packages.

Main Results:

• `apeglm` consistently outperformed maximum likelihood and generally surpassed pseudocounts in accuracy for allelic expression estimation.
• Adaptive shrinkage (ash) showed mixed performance, outperforming maximum likelihood in one simulation but not another.
• `apeglm` demonstrated superior speed and numerical reliability compared to five other beta-binomial modeling packages.

Conclusions:

• The `apeglm` package provides a fast, reliable, and accurate Bayesian approach for analyzing allelic imbalance, particularly beneficial for datasets with small read counts or sample sizes.
• This method overcomes limitations of traditional filtering approaches and offers an improvement over existing tools for allelic expression analysis.