DE Analysis

This page describes the differential-expression engines DEBrowser uses and the parameters it exposes for each.

Introduction

Differential gene expression analysis finds genes, transcripts, or regions whose difference in expression between two or more sets of samples — accounting for the variance within a condition — is higher than expected by chance. DEBrowser wraps three established Bioconductor engines, DESeq2 (Love et al., 2014), edgeR (Robinson et al., 2010), and limma (Ritchie et al., 2015), coupled with Shiny to produce real-time changes in your plot queries and interactive browsing of your DE results. You select the engine and its parameters on the Comparison step of the Data Prep wizard, under Advanced model settings.

DESeq2

See the DESeq2 user guide.

DESeq2 groups samples into conditions and computes, for every gene, the probability of differential expression using a negative binomial model, reporting both a nominal and a multiple-testing-corrected (padj) p-value.

Un-normalized counts. DESeq2 requires raw, un-normalized counts (the model corrects for library size internally). Non-integer values are converted to integers as needed. Do not feed it pre-scaled or normalized values — use edgeR or limma for those.

Parameters

  • fitType — parametric, local, or mean: how dispersions are fit to the mean intensity (see estimateDispersions).

  • betaPrior — whether to place a zero-mean normal prior on the non-intercept coefficients (see nbinomWaldTest).

  • testType — Wald (nbinomWaldTest) or LRT (likelihood-ratio test on the difference in deviance between a full and reduced model, nbinomLRT).

  • rowsum.filter — regions/genes/isoforms with total count (across all samples) below this value are filtered out.

edgeR

See the edgeR user guide.

Parameters

  • Normalization — scaling factors for raw library sizes: TMM, RLE, upperquartile, or none.

  • Dispersion — a numeric dispersion, or a string telling edgeR to estimate it from the data.

  • testType — exactTest (per-gene exact test between two groups) or glmLRT (negative-binomial GLM with genewise likelihood-ratio tests).

  • rowsum.filter — low-total-count features are removed.

limma

See the limma user guide. limma is ideal when data are already normalized (e.g. spike-in or another scaling), in which case prefer it over DESeq2 or edgeR.

Parameters

  • Normalization — TMM, RLE, upperquartile, or none.

  • Fit Type — ls (least squares) or robust (robust regression).

  • Norm. Bet. Arrays — normalize between arrays so intensities/log-ratios share similar distributions across samples.

  • rowsum.filter — low-total-count features are removed.

Batch-effect correction

On the Batch effect step DEBrowser can correct known technical batch effects before DE, using ComBat or ComBat-Seq (from the SVA package) or Harman. ComBat is applied with SVA’s defaults; ComBat-Seq operates directly on counts. Choose a normalization method (MRN, TMM, RLE, upperquartile) alongside the correction method, and confirm the effect with the inline PCA / IQR / Density plots.

References

  1. Anders S. et al. (2014) HTSeq — a Python framework to work with high-throughput sequencing data.

  2. Chang W. et al. (2016) shiny: Web Application Framework for R.

  3. Johnson W.E. et al. (2007) Adjusting batch effects in microarray expression data using empirical Bayes methods. Biostatistics 8, 118–127.

  4. Love M.I. et al. (2014) Moderated estimation of fold change and dispersion for RNA-seq data with DESeq2. Genome Biol. 15, 550.

  5. Murtagh F. and Legendre P. (2014) Ward’s hierarchical agglomerative clustering method. J. Classification 31.

  6. Ritchie M.E. et al. (2015) limma powers differential expression analyses for RNA-sequencing and microarray studies. Nucleic Acids Res. 43, e47.

  7. Robinson M.D. et al. (2010) edgeR: a Bioconductor package for differential expression analysis of digital gene expression data. Bioinformatics 26, 139–140.

  8. Vernia S. et al. (2014) The PPARα-FGF21 hormone axis contributes to metabolic regulation by the hepatic JNK signaling pathway. Cell Metab. 20, 512–525.