Figure 1.
Ref seq db is the reference sequence used to align reads, e.g. whole genome, miRBase. Pre-alignment filter refers to filters used to trim 5′ and 3′ linkers. Alignment to reference refers to the step in which a specific algorithm is used to align each of the reads to the reference sequence. This alignment can be done with/without considering the quality score associated with each base. Post-alignment filters are those used to remove low quality reads, alignments characterized by sequencing errors or multiple mismatches. Peaks segmentation refers to the definition of genomic regions characterized by enrichment of reads mapping, i.e. clusters of reads. Differential expression detection is the part of the analysis in which digital data are used to identify differentially expressed genes. Each of the workflow steps can be done using a variety of bioinformatics and statistical tools.
Table 1.
Spike-in experiment.
Figure 2.
Discrepancy in short reads counts detection using whole genome (wg-set) and miRNA precursor set (mir-set) as reference.
We expect that, if reference is not playing any specific role in the alignment procedure, then the same number of counts should be detected independently from the reference set in use. A higher number of miRNAs are shown to be underestimated when the wg-set is used as reference for the mapping (36 miRNAs) with respect to the mir-set (4 miRNAs). Red and black dots refer respectively to miRNAs detected in experiment A and B without significant variation between mir-set and wg-set. Green and blue dots refer respectively to miRNAs detected in experiment A and B with significant variation between mir-set and wg-set.
Table 2.
Discrepancies in mapping between mir-set and wg-set.
Table 3.
Primary mapping tools evaluated in this paper.
Table 4.
True positive and negative miRNAs set.
Figure 3.
Efficacy of detecting differentially expressed miRNAs.
The ability of edgeR, DEGseq, DESeq, baySeq and RankProd to detect differential expression in presence of absolute log2 fold change >3 folds was evaluated by mean of ROC analysis.
Figure 4.
ROC curves describing differential expression for baySeq (A), RankProd (B) and DESeq (C). D-F as A-C but zooming above 75% sensitivity and below 10% 1-specificity.
The legend shows the number of expected differentially expressed miRNAs associated to each of the 10 groups of spike-in and the corresponding expected log2 fold change variation range.
Figure 5.
ROC curve of the sample size effect for RankProd.
A) Four replicates for each experimental condition. B) Two replicates for each experimental condition, using background bk0. C) Two replicates for each experimental condition, using background bk1.
Figure 6.
ROC curves describing the effect of an increasing number of differentially expressed miRNAs: baySeq (A), RankProd (B) and DESeq (C).
Legend shows the ratio between expected differentially expressed miRNA and the full set of mapped miRNAs.
Figure 7.
Optimized microRNA differential expression analysis workflow for digital data.
a) reference sequence, b) post-processing filter, c) alignment tool, d) post-processing filter, e) differential expression statistics.