Moderated estimation of fold change and dispersion for RNA-seq data with DESeq2

Key results

1. ShownShrinking log fold changes (LFCs) toward a data-derived zero-centered prior made effect-size rankings far more reproducible between independent halves of one experiment than unshrunken maximum-likelihood LFCs.

   Bottomly et al. mouse strain samples split into two balanced halves: of the 100 genes with largest LFCs in half I, 21 reappeared in half II's top 100 using unshrunken estimates versus 81 of 100 using shrunken maximum a posteriori (MAP) estimates, though maximum-likelihood LFCs after adding a pseudocount of 1 gave nearly equal concordance. · Empirical Bayes shrinkage for fold-change…; Figure legends

2. ShownDESeq2 combined false discovery rate (FDR) control with competitive power: in simulations it, like edgeR, was often the most sensitive tool among those keeping actual FDR at or below 0.1; on real data it ranked high in precision, with sensitivity similar to edgeR and voom but below DSS.

   Simulation: 10,000 genes, total sample sizes 6, 8, 10, 20, with 20% differentially expressed at fold changes 2, 3, 4; DESeq2's edge was clearest at fold change 2 or 3. Bottomly et al.: 30 random 3 vs 3 evaluation splits checked against 7 vs 8 verification calls; median sensitivity typically 0.2 to 0.4 for all tools; DESeq2 often second in median precision after old DESeq. · Sensitivity and precision; Sensitivity; Precision

3. ShownOn real null comparisons with no true condition difference, DESeq2 controlled type-I error, keeping its median false positive rate just under the nominal level, whereas the original DESeq and Cuffdiff 2 were overly conservative.

   Pickrell et al. lymphoblastoid cell line RNA-seq from unrelated Nigerian individuals: 26 male samples with 46 base pair reads; 10 drawn at random for 5 vs 5 mock comparisons, repeated 30 times; false positive rate estimated as the fraction of p-values below 0.01. All tools generally controlled false positives; EBSeq was excluded because it outputs posterior probabilities. · False positive rate; Precision; Figure legends

4. SuggestedIn simulations with known true LFCs, DESeq2's shrunken estimates kept error consistently low across sample sizes and effect-size distributions, while GFOLD and edgeR's pseudocount-based predictive LFCs each lost accuracy in some regimes.

   Simulations of 1000 genes under four true-LFC models and varying sample size (10 simulation replicates). Over all genes, edgeR's default predictive LFCs (prior.count 0.125) had higher root mean squared error; on differentially expressed genes GFOLD erred more at large sample sizes. Under spike models, GFOLD and edgeR with prior.count 10 had lowest all-gene mean absolute error. · Precision of fold change estimates; Figure legends

5. SuggestedThe regularized logarithm (rlog) transform stabilized variance across the expression range and recovered sample groupings better than ordinary log counts, generally outperforming alternatives including the variance stabilizing transformation (VST) when size factors differed.

   Hammer et al. rat nerve ligation vs control (2 weeks, 2 months): rlog stabilized per-gene spread and grouped samples by treatment and time, unlike log2 counts plus pseudocount 1. Simulations (4 clusters of 4 samples, 2000 genes, 20 replicates): rlog had the best adjusted Rand index with size factors [1,1,1,3]; with equal size factors, Poisson distance and VST were also top. · Regularized logarithm transformation; Clustering; Figure legends