LongSAGE libraries were constructed using total RNA purified from nine different human ESC lines cultured as undifferentiated cells by serial passaging on mouse embryonic fibroblast (MEF) feeder layers 30 (Table 1). To enable detection of the majority of the moderately to abundantly expressed transcripts, we sequenced most libraries to a depth of approximately 200,000 tags. However, in one case (the library prepared from WA09 cells), we generated 468,252 tags. To ensure that tags included in the libraries were not contaminated with transcripts expressed from the MEF feeder layers, all tags matching the mouse reference genome sequence were excluded from further analysis (Additional data file 1). SAGE libraries were analyzed individually, and also as an electronically pooled 'meta-library' containing 2.5 million tags representing 379,645 different tag sequences. Of these, 73% were observed only once ('singletons'). Our previous experience indicated that singletons are enriched for experimental artifacts (sequencing errors, reverse transcriptase artifacts, and so on) as well as rare transcripts 30. To reduce the artifacts, we assigned confidence values to each tag sequence and selected for analysis only high quality tags as described 30. This filtering reduced the total number of different tag sequences to 268,515 (Additional data file 2). Of these, 40% of the singletons and 87% of the non-singletons could be mapped to publicly available gene expression resources.

To investigate the similarities and differences between the libraries, we performed hierarchical clustering using Pearson correlation coefficients 31. For this comparison, we included data from four LongSAGE libraries generated from terminally differentiated cells (available from the Cancer Genome Anatomy Project 32) to provide an 'out-group'. Figure 1 shows that the libraries for all nine human ESC lines form a cluster distinct from the libraries for the four terminally differentiated cell preparations, as expected. The ESC libraries also do not cluster together based on obvious commonalities between the lines, such as the MEF feeder lines used, sex chromosome karyotype or passage number.

To assess the representation of known genes in the nine human ESC transcriptomes, we compared our data to other human sequence tag-based resources 15161718. Highly expressed genes in each of the human ESC libraries showed significant overlap with previously published ESC SAGE 15 and MPSS datasets 18, but the diversity of genes identified by our LongSAGE data was significantly greater (Figure 2). To explore the functions encoded by transcripts detected in the LongSAGE libraries, we divided the genes (identified by uniquely mapping tags) into their respective Gene Ontology (GO) slim categories 33. Pair-wise comparisons of individual human ESC libraries showed little difference in the relative proportions of each of the GO slim categories (Additional data file 3). In contrast, a similar comparison of individual or pooled ESC libraries to the differentiated cell lines showed a statistically significant increase, in the ESC libraries, in the proportion of transcripts encoding RNA binding proteins and mitochondrial proteins (P = 1.8 × 10^-7 and 1.0 × 10^-6, respectively, by one-sided t-tests).

To investigate the potential functional significance of increased expression of transcripts for RNA binding proteins, we compared the global splicing profile of the ESCs and the four libraries of terminally differentiated cells. This was done by performing pair-wise comparisons across all transcripts in both the ESC and the terminally differentiated cell meta-libraries with the position of each uniquely mapped LongSAGE tag for which a transcript was known. These analyses did not reveal any difference in global transcript splicing patterns between the two meta-libraries, although differences in the relative abundance of specific transcript isoforms were identified. Of a total of 70 transcript isoforms found to be differentially expressed between the ESC and differentiated cell meta-libraries, 8 demonstrated statistical significance (P < 3.0 × 10^-5; Additional data file 4). The most significantly affected transcript (lowest P value) encoded Secreted frizzled-related protein-1 (Sfrp1), a well characterized antagonist of WNT signaling. Our analysis suggested that the two isoforms of Sfrp1 we identified either retained or lost the 3' untranslated region (UTR; Figure 3). Only the transcript isoform lacking the 3' UTR was found exclusively in the ESCs. Closer examination of the 3' UTR region revealed putative miRNA target sites for two evolutionarily conserved miRNAs 34, the mouse homologues of which were found previously to be expressed in murine ESCs 35 (Figure 3). Given that activation of the canonical WNT signaling pathway induces differentiation and cell proliferation 36, we speculate that the expression of Sfrp1 may be regulated through miRNA-directed translational repression and that this regulation is bypassed through alternative 3' end formation in pluripotent ESCs.

We next examined the expression of transcripts that encode previously identified markers of undifferentiated ESCs. These include transcription factors such as Oct 4 37, Nanog 45, the cell surface proteins tdgf-1 38 and thy-1 39, Lck 13, connexin cx43 40, Rex1 41 and Lefty-A and Lefty-B 42. In addition, we looked for transcripts from six genes associated with early stages of ESC differentiation 14. Table 2 shows the normalized gene expression levels across all cell lines. A similar pattern of expression is observed across all lines, with the exception of HSF-6, which exhibited a decrease in expression of ESC marker genes and a concomitant increase in expression of genes associated with differentiation, including alpha-fetoprotein. Notably, expression of Nanog, a divergent homeodomain protein that directs propagation of undifferentiated mouse ESCs 5, was not detected in the HSF-6 library. These features are consistent with the closer relationship of the HSF-6 library to libraries from differentiated tissues than to other ESC libraries (Figure 1). We therefore excluded the HSF-6 library from further analysis.

A previous analysis of SAGE data generated using ES03 and ES04 cells showed that Rex1 was within the top 25 differentially expressed transcripts, with no Rex1 tags detected in the ES04 line and an absence of Rex1 expression in ES04 cells confirmed by quantitative and semi-quantitative real time (RT)-PCR 15. Interestingly, in our LongSAGE libraries, tags for Rex1 were present in all nine ESC libraries, including the library prepared from ES04 cells and there was less than a three-fold difference in Rex1 expression between ES03 and ES04 (Table 2).

To generate a list of transcripts common to all libraries (excluding the HSF-6 library because of the differentiation markers found therein), we first identified tags from each library that uniquely mapped to transcripts within RefSeq 43 and the Mammalian Gene Collection (MGC) 44. This analysis identified a set of 4,337 LongSAGE tags present in all libraries (Additional data file 5). Comparison of this list to those generated by previous MPSS and SAGE approaches revealed extensive (80%) concordance between the SAGE-based transcriptomes. In contrast, 52% of genes identified by MPSS were not found in either of the SAGE common gene lists. Some of this lack of concordance may be explained by differences in the tagging restriction enzyme used by the two protocols (NlaIII for SAGE and Dpn1 for MPSS) and the fact that different mRNA preparations were used in each study. To further explore this lack of concordance, we compared the longSAGE and MPSS-derived gene lists to a common gene list derived from Affymetrix expression arrays generated from the same RNAs used to construct our LongSAGE libraries 45. The Affymetrix common gene set contained more than 80% of the LongSAGE common gene list (Additional data file 5) while MPSS contained only 68% of the genes on this list.

LongSAGE offers opportunities for discovering novel transcripts. These can be identified as tags that map uniquely to the genome but not to any available transcript resources. To look for these, we used the 2.5 million tag meta-library, which contained 379,645 unique tag sequences. Grouping LongSAGE tags that mapped to genomic locations in close proximity to one another 30 resulted in the identification of 24,593 transcription units. Of these, 14,588 did not overlap with known genes and were classified as novel. Most tags were expressed at low levels with 46% (6,672) identified by a single LongSAGE tag. Even though singletons are enriched for artifacts, many of these are likely to represent real transcripts, for two reasons: first, they map to the genome; and second, we 30 and others 46 have shown previously that at least 70% of novel, singleton, high quality LongSAGE tags identify rare transcripts whose expression can be confirmed in RNA-dependent RT-PCR experiments.

To further characterize these putative novel, low-abundance ESC library specific transcripts, we compared the ESC meta-library to publicly available data derived from 247 non-ESC SAGE libraries that together contained 654,491 unique tag sequences. This comparison identified 20,047 tag sequences found only in the human ESC meta-library (Additional data file 6). For subsequent analyses, we focused on those tags that uniquely mapped at least 2 kb away from any known gene. This analysis reduced the number of tags to 634 (Additional data file 7), of which 301 were found within genomic regions exhibiting sequence conservation between human and mouse or rat (Additional data file 8). We used rapid amplification of cDNA ends (RACE) 4748 to clone the 5' ends of 52 of these (Additional data file 9). Alignment of the resulting sequences to the human genome revealed that 22 (40%) were spliced. An open reading frame (ORF) scan of the 52 RACE clone sequences using Bioperl 49 tools and custom scripts identified 6 transcripts that encoded peptides longer than 100 amino acids in length. However, with the exception of one transcript (HA_003333) that overlapped the 3' end of the MAPK2 gene, none of the identified ORFs demonstrated Ka/Ks ratios suggestive of purifying selection 50. Hence, these transcripts may not encode proteins but may instead represent non-coding RNAs (ncRNAs).

Four RACE clones were found to have genomic coordinates that overlapped with those of known transcripts (Additional data file 9). One of these (HA_003240; Figure 4) is of particular interest because it contains the entire coding sequence of the Foxb1 gene within its first intron. Foxb1 encodes a winged helix transcription factor involved in the development of the vertebrate central nervous system and Foxb1^-/- mice display phenotypes consistent with a requirement for this gene in both embryonic and postnatal stages of development 515253. Interestingly, the ESC meta-library did not contain any tags corresponding to known Foxb1 transcripts except for a single Foxb1 tag in the HSF-6 library. This general lack of Foxb1 expression in ESCs and the genomic location of the Foxb1 gene within the first intron of HA_003240 are consistent with the notion that Foxb1 expression is repressed by expression of HA_003240, possibly by steric inhibition of the transcription initiation complex 54. The HA_003240 sequence overlaps partially with an EST obtained from an undifferentiated human ESC line (CD049816), as well as with ESTs from an embryonic carcinoma line, a kidney carcinoma line and hypothalamus tissue (for example, DA713666, DB173211 and BI458015, respectively). Examination of the promoter region of HA_003240 revealed the presence of highly conserved sequences containing an Oct/Sox binding element, suggesting that HA_003240 expression may be maintained in pluripotent ESCs through the recruitment of an Oct4/Sox2 complex (Figure 4). Oct4 encodes a transcription factor that regulates a number of key human ESC markers, including Nanog, through co-operative binding with a Sox family member 55. Given the documented role for Foxb1 in controlling the differentiation of neuronal cell types, the genomic organization of the Foxb1 locus is intriguing and suggests an interesting mechanism for negatively regulating Foxb1 expression in Oct4-expressing cells.

Many pseudogenes have been identified in the human genome using homology-based approaches 565758. Pseudogenes are generally not transcribed due to their lack of functional promoters 5960. However, there are examples of pseudogenes that have retained or acquired functional promoters, leading to their transcription 61. Because of the low levels of expression of the 52 novel transcripts (on average, only 3 tags per million) we asked whether the 5' RACE clones were derived from expressed pseudogenes. Comparison of the RACE clone sequences to three computationally generated lists of known human pseudogenes 565758 revealed only one clone (HA_003350) with a predicted pseudogene contained within its exon. Furthermore, with the exception of HA_003333, none of the novel transcript sequences showed significant sequence similarity to any known ORF (using a 70% ORF threshold 58). Taken together, these analyses do not support the notion that the novel genes identified by our analysis are enriched for expressed pseudogenes.

To more fully characterize a transcript identified by a singleton tag (Additional data file 9), we attempted to recover a full length transcript using 5' and 3' RACE and primers annealing within the terminal exon of the putative transcript. Alignment of the resulting candidate full length sequence to the human genome revealed a transcript that contained two introns (Figure 5). Examination of the genomic region surrounding this transcript showed that it resides in a region of the long arm of chromosome 3 (chr3:110,539,351-110,584,565) lacking annotated transcripts. The putative transcriptional start is located 266 bp from the transcriptional start site of Dppa4, a gene known to have an expression pattern in ESCs that is similar to that of Oct4 62 (Figure 5). To investigate the possibility that this promoter region is regulated directly by Oct4, we looked for the presence of conserved Octamer and Sox (high mobility group (HMG)) elements. A single 20 bp region of cross-species sequence conservation was found that contains a consensus binding element for an Octamer/Sox dimmer, suggesting that the novel gene is regulated by Oct4/Sox2 (Figure 5; chr3: 110,539,180-111,539,200). In support of this finding, the conserved region was found to reside within a probe identified by chromatin immunoprecipitation (ChIP)/CHIP 63 as a target of Oct4 and Sox2 (Probe spans chr3: 110,539,028-110,539,588). Taken together, these analyses suggest that both Dppa4 and the novel transcript are divergently transcribed from a common promoter bound by an Oct4/Sox complex. Based on its proximity to the Dppa4 gene we have named this novel transcript Spd4 (for 'shares promoter with Dppa4').

Comparison of the 5' RACE clone sequences to publicly available ESTs revealed 36 (69%) with matches to other ESTs, of which 7 were found only in data derived from pluripotent human ESC lines. One RACE clone that overlapped an EST derived from pluripotent human ESC lines (HA_003152) was also found to be expressed in all nine ESC lines studied here. BLAT 64 alignment of the 5' RACE clone sequence to the human reference genome sequence revealed that HA_003152 contained two introns and resided within a genomic region that exhibited sequence similarity to long interspersed nuclear elements. An ORF scan revealed a 129 amino acid peptide encoded in the second exon with homology to the carboxyl terminus of the LINE p40 ORF.

To explore the expression pattern of the HA_003152 transcript we used quantitative RT-PCR (qPCR) to compare transcript levels in RNA purified from human ESCs maintained under conditions that promote their maintenance in an undifferentiated state to RNA extracts obtained from human ESCs that had been stimulated to differentiate into embryoid bodies. To provide a comparative dataset we selected five additional novel transcripts for qPCR. In all cases, qPCR amplicons were designed to cross exon-exon boundaries. As controls we also monitored expression of Oct4, Lin28 and Msx1 in the same RNA preparations. Figure 6 shows the expected expression pattern for the control gene set, with a reduction in expression of Oct4 and Lin28 in the human ESCs stimulated to differentiate into embryoid bodies and an up-regulation of expression of the early differentiation marker Msx1. Significant reduction of expression was observed in four of the six transcripts tested, including HA_003152, whose expression was undetectable at d30 (Figure 6). These transcripts are hence potential markers of pluripotency.

Detailed information regarding the human ESC lines used in this study can be found at the NIH Stem Cell Information website 65. The passage numbers of the cells analyzed in this study are presented in Table 1. Total RNA was prepared using Trizol reagent (Invitrogen, Burlington, ON, USA) following the manufacturer's protocol and was assayed for quality and quantified using an Agilent 2100 Bioanalyzer (Agilent Technologies) and RNA 6000 Nano LabChip kit (Caliper Technologies, Hopkinton, MA, USA).

Nine LongSAGE 29 libraries were constructed from 5-20 μg of DNase I-treated total RNA as described 30 (DNase I from Invitrogen). LongSAGE data generated for this study are available through our embryonic stem cell transcriptomes website 45 and through the CGAP web portal 32.

LongSAGE tags of at least 99.9% accuracy (calculated using Phred 6667 quality scores) from the meta-library were compared to 247 publicly available human SAGE libraries (GEO 68, Discovery db 69). To allow direct comparison of the LongSAGE data to the 14 bp SAGE tags available in the public libraries, the 3' ends of the 21 bp tags were truncated in silico to form 14 bp tags. A total of 2,508,608 tags corresponding to 222,337 unique 14 bp tag sequences (379,465; 21 bp parental sequences) were utilized in this analysis. These tags were directly compared to all unique tags from the human SAGE libraries to generate a list of tags found solely in the ESC meta-library.

Tag-to-gene mapping was performed using the comprehensive mapping of SAGE tags (CMOST) software 69 as follows. Tags were mapped to various publicly available transcript databases in a hierarchical fashion with the highest quality transcript databases used first. As tags were mapped to a known transcript in a higher quality database, they were excluded from further analysis with subsequent lower quality databases to mitigate redundancies arising from lower quality DNA sequence resources. The following databases were used for CMOST tag-to-gene mapping in this order: MGC 70, RefSeq 71, Ensembl transcripts 72 (exon sequences only), Genbank Human Mitochondrial Sequence (accession AY289102.1), Genbank Non-coding sequences 73, Ensembl genes 72 (1,000 bp UTR and intron sequences included), Ensembl ESTs 72, and Golden path genomic contigs (Genbank Human Genome Assembly Contigs build 34, January 2004 73). In addition to allowing perfect matches, the CMOST approach attempts to account for single base permutations, insertions and deletions, improving the rate of tag-to-gene mapping.

LongSAGE tags were mapped to known and computationally predicted transcripts using versions of the following databases available as of March, 2005: RefSeq 71, RefSeqX 71, Mammalian Gene Collection 70, and RefSeqGS 71. Tags were also mapped to human genomic sequence using the NCBI Reference Sequence Genome database 71, release 35, August 2004. From the genome sequence, a table was generated containing all 27.4 million potential SAGE tags adjacent to genomic NlaIII restriction sites (CATG). Of these, our analysis defined a subset of 19.4 million genomic tag sequences that were unique within the genome.

A second table was generated that stored information about exons: genome sequence contig, transcript orientation, exon number, exon boundary type and nucleotide positions of exon boundaries for all approximately 267,000 exons annotated on release 35 of the Reference Sequence genome. The LongSAGE tag sequences were compared to the unique genomic tag table, yielding sets of genomic positions for all tags in the library. These in turn were compared to the table of exon information, producing a mapping for each tag relative to annotated exons.

For the GO category comparisons, a standard t-test comparing two samples was used. The null hypothesis was that the two samples arose from populations with the same mean and standard deviation. The values within each sample were the number of GO categories represented in each library of the set, nine in the ESC set and four in the normal set. To account for variation due to library size, only the transcripts with the top 1,000 expression values were included. A one-sided p value was reported. Microsoft Excel was used to perform the computation.

To select differentially expressed LongSAGE tags, the ESC and CGN meta-libraries were compared on a tag per tag basis to obtain a p value for the null hypothesis that the two tag frequencies arose from Poisson distributions with the same mean. This was derived using a normal approximation to the Poisson as described by Kal et al. 74. All transcripts that showed differences with a significance of p < 0.05 were selected. Tag counts were converted to tags per million, and transcripts that differed by less than three-fold were eliminated. All pairs of tags existing within the same transcript were then listed if the differential expression for the two tags was in the opposite direction.

First strand 5' and 3' RACE ready cDNA was synthesized from 2.0 μg of DNase I (DNA-free™ kit; Ambion, Austin, TX, USA) treated RNA using the BD SMART RACE cDNA Amplification kit following the manufacturer's recommended protocol (BD Biosciences Clontech, Mountain View, CA, USA). Gene specific 5' RACE primers were designed using custom scripts and Primer 3 75 to lie downstream of the target LongSAGE tag with an optimal Tm of 68°C (Additional data file 10). For 3' RACE reactions a series of primers were designed manually based on the 5' RACE clone sequence (Additional data file 10). The cDNA was amplified using the Phusion™ High-Fidelity PCR Kit (MJ Research, Inc., Waltham, MA, USA) following the manufacturer's recommended protocol with the addition of DMSO to a final concentration of 3%. The cycling conditions consisted of an initial denaturation at 98°C for 30 seconds followed by 10 touchdown PCR cycles starting with 98°C for 10 seconds, 72°C (decreased by 1°C in each subsequent cycle) for 15 seconds, 72°C for 30 seconds; then 29 cycles of 98°C for 10 seconds, 62°C for 15 seconds, 72°C for 30 seconds; followed by an extension at 72°C for 10 minutes. PCR product for each sample (10 μl) was loaded on a 1.2% agarose gel and subjected to electrophoresis for 3.5 hours at 110 mA in 1× TBE buffer (Tris/Boric Acid/EDTA). The gel was stained with SYBR Green (Mandel, Guelph, ON, Canada) and visualized using a Typhoon 9400 Variable Mode Imager (Amersham, Baie d'Urfe, PQ, Canada). Amplicons were extracted from the gel, purified and cloned into the pCR4^®-TOPO^® vector using the TOPO TA Cloning^® Kit for Sequencing (Invitrogen). Plasmid vectors were electroporated into bacterial cells, and recombinant clones were selected on agar plates containing appropriate antibiotics as described 76. Glycerol stocks were prepared from 12 individual clone isolates per amplicon and stored in 384-well plates. Clone inserts were sequenced on an ABI PRISM 3730 XL DNA Analyzer using BigDye primer cycle sequencing reagents (Applied Biosystems, Foster City, CA, USA).

RNA was obtained from H9 cells before and after induction of differentiation using a 30-day embryoid body protocol. Undifferentiated H9 cells maintained for 7 days on matrigel (BD Biosciences, San Jose, CA, USA) in media conditioned by mouse embryonic fibroblasts and supplemented with 4 ng/ml fibroblast growth factor (bFGF-2) were harvested for embryoid body formation. Briefly, the cells were incubated with TrypLE (Invitrogen) for 10 minutes at 37°C and then collected by scraping. Resultant cell aggregates were subsequently cultured in non-adherent dishes using KOSR-based media without FGF2, for 15 to 30 days. At appropriate time-points RNA was extracted into Trizol. cDNA was synthesized from 2.0 ug of DNase I (DNA-free™ kit, Ambion) treated total RNA using the SuperScript Choice System following the manufacturer's recommended protocol (Invitrogen). Gene specific primer pairs were designed using custom scripts and Primer 3 75 to amplify approximately 150 bp of the target gene with an optimal Tm of 68°C (Additional data file 10). Whenever possible amplicons were designed to cross exon/intron boundaries. Amplification was performed in a 10 μl reaction mixture containing 5 μl of 2× SYBR Green PCR Master Mix (Applied Biosystems), 2 μl of template cDNA, and 250 pmol of the forward and reverse primer pair. After preparation of the reaction mixtures in 96-well plates, the plates were centrifuged at 800 rpm for 1 minute in an Eppendorf 5810 swing rotor centrifuge (Eppendorf, Westbury, NY, USA). Amplification and detection were performed on an ABI Prism 7600 Sequence Detection System (Applied Biosystems). The PCR protocol consisted of the following: a single cycle of 10 minute at 95°C and 40 two-step cycles, with one cycle consisting of 15 seconds at 95°C and 60 seconds at 60°C. Results were analyzed as described 77 using a GAPDH probe for normalization.