by DGE (x-axis) against those obtained by qPCR (y-axis). All data and primer sequences can be found in Supplementary Table 3.
With DGE we found a much wider distribution of fold-changes between the closely correlated groups of mice than for the microarray platforms, where
the highest fold-change measured was 2. ByDGE, we observed 1491
significantly differentially expressed tags (error rate <0.05) with an absolute fold change >2 (Figure 2). The only three genes that were significant on all microarray platforms and confirmed by qRT-PCR, Plac9, D14Ertd449e and Gabra2, were also significant in DGE (Bayesian error rates of 2.0210–48, 3.5210–47 and 3.9210–12, respectively). For the comparison between DGE and qPCR, we selected 29 significant genes from the DGE experiments (randomly chosen and covering the entire range of
significance (Bayesian error rates between error rates between 1210–47 and 0.05) and fold-changes), and 33 genes significant genes from the microarray analyses (9). Results are given in Supplementary Table 3 and displayed in Figure 4B. The fold-changes obtained by DGE were generally also more extreme than those obtained by quantitative PCR, as is evident from the slopes of the curve. Out of 62 genes assayed, 43 demonstrated a concordant direction of change for DGE and qPCR, but only five were significant according to both technologies.
We made a more general comparison of the lists of differentiallyexpressed genes from the DGE and microarray experiments. Differential gene
expression for DGE was established with Vencio's algorithm as described above (estimated FDR 8.5%) and for microarrays with the Empirical Bayes model LIMMA (32) (estimated FDR 10%). Complete results on correspondence between DGE tag counts and microarrays are reported in Table 5. The biggest overlap was found with the Affymetrix platform (P = 1.2210–5; chi-square test): 31 transcripts were significant on both platforms with expression changes in the same direction. Also when assessing the correlation of the expression between transgenic and wild-type mice, Affymetrix chips were found to correlate better with DGE (Pearson correlation: 0.25) than the other microarray platforms (Supplementary Figure 7). The number of differentially expressed transcripts by DGE was closest to the number detected with the Agilent platform (2414 and 2710). However, the overlap between these transcripts was hardly greater than would be expected by chance and there was little correspondence in the direction of change.
View this Table 5. Overlap between DGE and microarrays in
table: differentially expressed transcripts [in this window] [in a new window]
TOP ABSTRACT INTRODUCTION MATERIALS AND METHODS RESULTS DISCUSSION SUPPLEMENTARY DATA FUNDING REFERENCES DISCUSSION Deep sequencing is a powerful technique for the identification of
differentially expressed transcripts. The large sequencing depth clearly boosts the detection of differential expression of low-abundant transcripts that are well beyond the reach of classical SAGE. The sequencing depth of the Solexa/Illumina DGE technology compares favorably
5
to the earlier MPSS systemfrom Lynx Therapeutics [7210 sequences per run (7)] and Roche [454 sequencer, 32105 sequences per run (33)], and is comparable to the polony multiplex analysis of gene expression (34). Instead of sequencing SAGE tags, some recently published papers now describe the use of random shotgun RNA sequencing (RNASeq)
(27–29,35–38). This overcomes the limitations of tag-based methods in the detection of transcripts alternative splicing in regions remote from the 3'-end, and enables detection of allele-specific transcription. With the continuously increasing number of reads at reduced costs, RNASeq will become affordable for standard differential gene expression analysis. However, at the present throughput it is favorable to use methods that provide a specific tag for each transcript, when the aim is to detect
subtle expression differences in larger group of samples: We demonstrate that 2 million tags are required to reliably detect low abundant genes with DGE, whereas RNASeq requires at least 20 million tags per sample to obtain reasonable coverage of most transcripts (29,36).
We have implemented a dedicated Bayesian method to identify genes that are significantly differentially expressed between two groups of
biological replicates. In most previously published reports analyzing differential gene expression in count-based data, the statistical tests applied did not account for within-group variation (28,34). We illustrated the importance of proper estimation of within- and
between-group variation by showing that classical tests lead to the
identification of false-positive genes dueto the presence of a single blood contaminated sample. In an earlier deep sequencing report (27), in analogy to microarray data analysis, quantile normalization and a
moderated t-statistic as implemented in the R package Limma were used to find differentially expressed genes. We believe that our method is better suited for the comparison of independent sequence libraries because one of the intrinsic properties of the test is that it puts more weight on samples which were sequenced at greater sequencing depth.
The availability of biological replicate measurements allowed us to use the global test (11), which takes into account the expression levels in individual samples, for the detection of disturbances in several biological pathways. Several of the identified pathways were highly relevant given the function of the DCLK1 protein (8,23–25). These pathways had not been identified by any of the microarrays using the same statistical test (9).
Our results demonstrate many advantages of DGE over expression microarray technology: (i) DGE gives an unbiased view of the transcriptome, not limited by predictions of expressed transcripts used to determine array content; (ii) DGE detects high levels of differential polyadenylation and antisense transcription, which are not detectable with standard
microarrays; (iii) DGE data are more precise than microarray data; (iv)
DGE data analysisrequires a lower number of preprocessing steps (like background correction and normalization), which facilitates
interlaboratory comparisons; (v) interlaboratory comparability of DGE data is high, probably due to the avoidance of hybridization processes, which are notoriously difficult to standardize (1); and (vi) DGE is more sensitive in the detection of low-abundant transcripts and of small changes in gene expression. This is probably due to the absence of background signal and saturation effects, which are major causes of ratio compression on microarrays (39). Some of these advantages have already been discussed in older literature comparing tag-based methods (SAGE,
MPSS) and microarray data (2,26,30,31,40–45). The higher sequencing depth of DGE and the avoidance of laborious cloning steps add to the presumably superior precision and accuracy of DGE over these older methods,
in particular when low-abundant transcripts areconsidered, and makes DGE
a much more practical technique.
The correlation between DGE and microarrays and between DGE and qPCR assays was clear but modest. In accordance to what has been previously reported in comparisons between SAGE or MPSS and microarrays (31,40), the correlation between tag-based methods and microarrays was particularly poor for low-abundant transcripts. An important reason for the relatively low correlation across different technologies is the great similarity
betweenour two sample sets. The resulting small differences in geneexpression are difficult to pick up with microarrays, as also shown in the inter-microarray comparison of the same samples published recently (9), and also with qPCR assays. In samples with larger differences in gene expression, like the samples analyzed by the MAQC consortium (46), the correlation is likely higher. We believe that, apart from differences in sensitivity, an important reason is that the different platforms detect different transcripts. The microarray probes and qPCR assays detect, in many cases, a mix of different transcripts (1), where DGE can discriminate between transcripts with different 3'-ends; standard qPCR assays will detect cumulative presence of sense and antisense transcripts. Indeed, when all DGE tags behave similarly, as with the Gabra2 gene where we find 6 tags with an 2.5-fold decrease in the DCLK mice (four from the sense and two from the antisense strand, see Supplementary Table 4), DGE results
are consistentwith all microarray platforms and qPCR (see Supplementary Figure 8). In many other cases, there will be no co-regulation between alternatively spliced transcripts or sense and antisense transcripts, which, especially in low-abundance situations, will result in poor correlation with microarrays and qPCR. In addition to the limited overlap in transcripts detected by both DGE and microarrays, many transcripts are detected only by one or a few of the platforms. For DGE, missing data for some transcripts are likely attributable to the absence of a CATG or a unique tag sequence (estimated frequency: 1% of murine RefSeq RNAs); for
microarrays this isdue to inadequate probe design. We also noted that there was a higher consistency between the fold-changes obtained by qPCR and microarrays when compared to those obtained by DGE. Apart from the explanations mentioned above, this is likely attributable to the fact that
DGE measures absolute expression levels andDGE data are Poisson
distributed (47), while qPCR and microarrays provide relative expression levels, which are log normal distributed.
Our finding that DGE results were more consistent with Affymetrix results than with other microarray platforms is consistent with an earlier study (31,42), in which MPSS results correlated better with Affymetrix than with
other arrays. We think this lies inthe use of multiple probes per gene, which should even out most probe-specific effects. Sequence biases in the different technologies have been described before. Comparative analysis of SAGE and microarrays shows that the GC content of microarray probes is important for detection sensitivity and for the correlation across technologies (26,30,41,43–45). We investigated GC bias in the DGE tags. The overall GC percentage observed in our tags is 42%. This is lower than for classical SAGE or MPSS (44) and better reflects the relatively low GC content of 3'-UTRs (48). By ranking the tags from high to low abundance,
we find a higher percentage of Ts in the higher abundant tags
(Supplementary Figure 9). This supports an earlier observation that highly expressed genes contain more T-rich 3'-UTRs than lowly expressed genes (48). Thus, the GC bias in DGE seems to be limited, but needs further
investigation, also in the lightof a recently published study where considerable overrepresentation of GC-rich sequences was observed in Solexa/Illumina-based resequencing experiments (49).
We foresee that further enhancements in sequencing depth will yet improve accuracy, in particular for low-abundant transcripts. Whole transcript sequencing (RNAseq) is another step forward. These advances, in
combination with the currently achieved improvements in sensitivity, resolution and, notably, interlaboratory consistency, will tremendously boost the field of expression profiling. Multicenter biobanking and rare disease studies, where biological materials are scarce and widely spread and legal and logistical limitations may impede sharing of source materials, would gain enormously from better possibilities for robust post hoc integration of results. Also basic research and comparative genomics fields, which have been held back by extensive and lengthy standardization issues, will greatly benefit from the major improvement of data portability.
SUPPLEMENTARY DATA