Fig 1.
Read counts for each exonic nucleotide (nt) position in CisGenome Browser [19] along gene ENSG00000197746 (PSAP) in different tissues of the Human Body Map dataset (a) brain, (b) breast, (c) liver and (d) lung.
Counts for reads starting at each exonic nucleotide position are shown.
Fig 2.
Normalized read counts for each exon along gene ENSG00000197746 (PSAP) in different tissues of the Human Body Map dataset (a) brain, (b) breast, (c) liver and (d) lung.
Reads mapped to each exon are counted and normalized according to the exon length. Overlapping exons are segmented into multiple non-overlapping ones.
Fig 3.
Transcript expression deconvolution by PGseq for gene ENSG00000130816.
(a) and (b) show the normalized read counts for samples HBR and UHR, respectively, for each exon. (c) shows the estimated gene-specific read distribution. The 3rd panel shows the obtained NB distributions for both gene and isoforms for the two samples. The last panel presents the approximated Gaussian distributions of logged gene and isoform expression. This gene contains four isoforms and the total normalized read counts for both samples are 601 and 2341, respectively.
Fig 4.
Transcript expression deconvolution by PGseq for gene ENSG00000152291.
(a) and (b) show the normalized read counts for samples HBR and UHR, respectively, for each exon. (c) shows the estimated gene-specific bias distribution. The 3rd panel shows the obtained NB distributions for both gene and isoforms for the two samples. The last panel presents the approximated Gaussian distributions of logged gene and isoform expression. This gene contains six isoforms and the total normalized read counts for the two samples are both 1321.
Table 1.
Comparison of expression estimation accuracy at gene level using various datasets.
Table 2.
Comparison of gene expression estimation accuracy for groups with different level of expression.
Table 3.
Comparison of expression estimation accuracy at transcript level.
Fig 5.
Usefulness of measurement error obtained by PGseq.
The scatter plots (left column) show the standard deviation vs. the logarithm of gene expression for the PCR-validated genes in the MAQC (upper panel) and HCC (lower panel) datasets. As the expression level increases the associated measurement error decreases. The ROC curves (right column) indicate the difference between accounting for and ignoring measurement uncertainty in the DE analysis. The DE analysis employs MMDiff which considers expression measurement error. The solid curves show the performance of the DE analysis considering expression measurement error, and the dashed lines ignore measurement error by setting zero measurement error.
Fig 6.
ROC curves of DE analysis for the selected PCR-validated genes in the MAQC (left) and HCC (right) datasets.
Cufflinks and MMSEQ are combined with the corresponding embedded DE analysis methods, Cuffdiff and MMDiff, respectively. RSEM is combined with DESeq. PGseq is combined with MMDiff for propagating measurement error in the DE analysis.
Table 4.
Area under ROC curves for detection of DE genes.
Fig 7.
Venn diagram of the significant DE genes for the HDB dataset.
The big ovals represent the number of the significant DE genes found by the four methods: Cufflinks, MMSEQ, PGseq and RSEM, which combined with CuffDiff, MMDiff, MMDiff and DESeq, respectively. The overlap of the four ovals in the middle of the diagram is 729, which is the number of the DE genes found by all of the four approaches.
Fig 8.
Scatter plot of the average logged RPKM estimation for the HDB dataset.
There are 23,402 expressed genes (RPKM>0 for both conditions), among which the 729 common DE genes found by all the four DE methods are represented by red triangles and others by green crosses.
Fig 9.
ROC curves of DE analysis for the lowly expressed genes in the HBD datasets.
The 348 DE genes found by all the four methods are taken as the “true” DE genes and the rest as non-DE genes.