The transcriptome can be profiled by high throughput techniques including SAGE [1], microarray [2, 3], and sequencing of clones from cDNA libraries [4, 5]. For more than a decade, oligo-nucleotide microarrays have been the method of choice providing high throughput and affordable costs. However, microarray technology suffers from well-known limitations including insufficient sensitivity for quantifying lower abundant transcripts, narrow dynamic range and non-specific hybridizations. Additionally, microarrays are limited to only measuring known/annotated transcripts and often suffer from inaccurate annotations [6]. Sequencing-based methods such as SAGE rely upon cloning and sequencing cDNA fragments. This approach allows quantification of mRNA abundance by counting the number of times cDNA fragments from a corresponding transcript are represented in a given sample, assuming that cDNA fragments sequenced contain sufficient information to identify a transcript. Sequencing-based approaches have a number of significant technical advantages over hybridization-based microarray methods. The output from sequence-based protocols is digital, rather than analog, obviating the need for complex algorithms for data normalization and summarization while allowing for more precise quantification and greater ease of comparison between results obtained from different samples. Consequently the dynamic range is essentially infinite, if one accumulates enough sequence tags. Sequence based approaches do not require prior knowledge of the transcriptome and are therefore useful for discovery and annotation of novel transcripts as well as for analysis of poorly annotated genomes. However, until recently the application of sequencing technology in transcriptome profiling has been limited by high cost, by the need to amplify DNA through bacterial cloning, and by the traditional Sanger approach of sequencing by chain termination [7].

The next-generation sequencing (NGS) technology [8–10] eliminates some of these barriers, enabling massive parallel sequencing at a high but reasonable cost for small studies. The technology essentially reduces the transcriptome to a series of randomly fragmented segments of a few hundred nucleotides in length. These are amplified by a process that retains spatial clustering of the PCR produces, and individual clusters were sequenced in parallel by one of several technologies [7]. Current NGS platforms include the Roche 454 Genome Sequencer, Illumina's Genome Analyzer, and Applied Biosystems' SOLiD. These platforms can analyze tens to hundreds of millions of DNA fragments simultaneously, generate giga-bases of sequence information from a single run, and have revolutionized SAGE and cDNA sequencing technology [11, 12]. For example, the 3' tag Digital Gene Expression (DGE) uses oligo-dT priming for first strand cDNA synthesis, generates libraries that are enriched in the 3' untranslated regions of polyadenylated mRNAs, and produces 20-21 base cDNA tags. It's reported that 99.8% of 21-base tags occur only once in human genome [13], while analyses based on actual sequence information from approximately 16,000 known genes suggest that >75% of 21-base pair tags are expected to occur only once in the human genome, with the remaining tags matching duplicated genes or repeat sequences [13]. The work reported in this manuscript effectively generates 20-base pair tags and the % of tags mapped to multiple locations on genome would be slightly higher. The 3' tag DGE profiling using Illumina Genome Analyzer II generates up to 10-12 million reads from individual libraries of unique, positionally known 20- or 21- base pair 3' cDNA tags. DGE technology generates such extensive sequencing depth-of-coverage that single copy resolution of gene expression quantification should be possible. For example, it's generally accepted that on average there are 350,000 transcripts expressed per cell (Illumina White Paper: mRNA Expression Analysis). Therefore a 10 million tag sequencing experiment by the Illumina GA II would detect 30 tags per transcript expressed at one copy per cell. This technology makes it possible, for the first time, to interrogate low abundance transcripts, which may comprise as much as half of the non-structural RNA within a cell. The alternative to 3' tag DGE sequencing is full-length RNA sequencing (RNA-Seq), which collects both quantitative and qualitative information about the entire transcriptome. RNA-Seq is a powerful tool for identifying expressed polymorphisms as well as differentially expressed splice variants, fusion genes, transcriptional start sites, and polyadenylylation sites [14]. However, the number of reads needed to characterize whole transcript sequences, using RNA-Seq, increases the cost of measuring rare transcripts and may be more economically quantitated with 3' tag DGE profiling. Assessment of the 3' DGE method has been limited. One study used 3' tag DGE approach in mouse hippocampal expression profiling [15], and another paper compared 3' tag DGE with SAGE [15]. More research is needed for evaluation of the sensitivity, dynamic range, reproducibility, and measurement accuracy of 3' tag DGE.

Our work focuses on exploring and validating the applicability of 3' tag DGE to the quantification of transcripts. To this end, we have designed a set of experiments to assess the use of 3' tag DGE as an alternative to microarrays or to RNA-Seq. We used two RNA samples described in the MicroArray Quality Control (MAQC) project [16]: Human Brain Reference RNA (HBRR) and Universal Human Reference RNA (UHRR). The samples, from the exact batches used by the MAQC project, are commercially available to the research community and have been well-characterized by microarray (repeatedly assayed by 7 microarray platforms) and quantitative real time PCR (qPCR, 3 quantitative PCR platforms). Therefore, comparing the profiles of these two RNA samples generated by the 3' tag DGE, by microarray, and by qPCR will provide an excellent frame of reference for evaluation of the sensitivity, dynamic range, reproducibility, and accuracy of 3' tag DGE sequencing.

Massively parallel sequencing for transcriptome profiling generates digital counts of gene expression levels compared to the "analog" hybridization signals from the traditional microarray or quantitative PCR methods. Our analysis highlights the precise nature of DGE profiling, as demonstrated by the high reproducibility between technical replicates of the same library run in different lanes, and between biological replicates of different libraries on the same or different flow cells runs. The correlations between technical replicates were >0.97. Although variability between biological replicates of two different libraries (L6 and L8) was less than 0.95, it is likely that the variance came from library construction and not the sequencing process, since the correlations between biological replicates of the libraries constructed at one lab (Florida) were excellent (> 0.95, Additional File 2). We also demonstrated that DGE profiling is accurate. The measurement of fold changes of genes between HBRR and UHRR was highly correlated with data obtained from qPCR. This correlation is similar to that between microarrays and qPCR.

With 20 million tags each from HBRR and UHRR library, which is the current sequencing throughput of one lane of sequencing on GA II, DGE detected 10-20% more transcripts than microarrays, a majority of which were expressed at levels below the sensitivity threshold of microarray platforms. The detection of the lower-expressed genes and the wider dynamic range have been shown to be the main advantages by DGE compared to microarray analysis, since other parameters evaluated between the two platform were mostly comparable. It has been suggested that lower expressed transcripts may account for nearly half of all transcripts in a cell, and play critical but currently undefined roles in pathology and physiology. DGE's ability to quantify these transcripts may open new horizons for the application of genomic profiling to translational research. In addition, DGE may lead to the discovery of new functional genomic regions. For example, 13-15% of the reads from our DGE libraries aligned to intergenic regions which may be related to novel transcripts. However, since no DNase treatment was performed during RNA extraction, these un-mapped reads could also be from the contamination of the genomic DNA although this scenario is less likely because of the selection of PolyA+ RNAs during library construction.

One limitation of 3' DGE is that using a particular enzyme such as DpnII for library preparation requires the presence of DpnII restriction site(s) in the mRNA. Some transcripts, even though highly expressed, may lack the DpnII site(s) and therefore not represented in the libraries. Among all 43,569 transcripts in Human RefSeq RNA database (version of Aug 24, 2009), 2,912 (6.68%) don't have a DpnII site. Among 28,061 mature RNAs with accession numbers starting with NM_ (excluding XR_, NR_ and XM_ accession numbers), 571 (2.04%) don't have a DpnII site. Additional File 5 lists all Human RefSeq RNAs and the number of DpnII sites within each molecule. In addition, the non-polyadenylated transcripts are not represented in the current libraries. In an effort to measure both mRNA and non-polyadenylated RNAs (data not shown), we tried to treat total RNAs with RiboMinus™ Transcriptome Isolation Kit (Invitrogen) prior to library preparation to deplete 18S and 28S rRNAs and to enrich polyadenylated mRNA, non-polyadenylated RNA, pre-processed RNA, tRNA, and small rRNAs (5S rRNA, 5.8S rRNA). However, the sequencing of RiboMinus libraries revealed that ~70% of the reads were rRNAs (data not shown). Therefore, we made a decision to prepare polyadenylated mRNA libraries for the current study to avoid wasting sequencing depth on rRNAs.

On the other hand, some of the genes that were detected by Affymetrix only may only have DpnII tags that were mapped to multiple genes. These tags were excluded from our analysis. We point the readers to the web site of the Stowers Institute where these redundantly mapped tags were recorded http://research.stowers-institute.org/microarray/tag_tables/downloads.html.

Several challenges currently limit the adoption of 3' DGE in replacement of microarrays. The higher cost per sample and low sample throughput per run limit the sample size in a study. Currently 3' tag DGE analysis costs ~$1200 per lane of sequencing including library construction. It also has long individual run times, sequencing only ~12 bases per day, and sample preparation is significantly more difficult and time consuming than that of microarray. Finally, the bioinformatics challenges associated with 3'DGE analysis are significant, including storage, archiving, and retrieval of the vast volume of data, and development of algorithms to assemble and align sequence reads as short as 35-40 nucleotides [19]. However, advances in these new sequencing technologies will substantially increase sample throughput, leveraging techniques such as multiplexing that could reduce the cost of sequencing per sample. Bioinformatics challenges are currently being addressed, faster and more powerful alignment tools being developed.

We are aware that RNA-Seq enables a more extensive profiling of the transcriptome by facilitating the discovery of fusion genes [20], detection and quantification of alternative splice forms, and characterization of expressed mutations and polymorphisms [14]. However, in order to comprehensively sequence the full length of all transcripts, the sequencing depth of RNA-seq needs to be significantly greater compared to that of 3' tag DGE profiling. It was estimated that at least 40 million reads (compared to < 5 million reads required for DGE) needs to be sequenced from a single library to achieve 90% coverage of the transcriptome [10] making RNA-seq even more financially demanding that 3' tag DGE. In addition, the increased complexity of the data poses even greater analytic challenges. Therefore, the advantages of 3' tag DGE over microarray in profiling lower-expressed genes and measuring with greater dynamic range make it arguably attractive for use in today's medical and biological research.