We have studied the transcripts produced from annotated gene regions, novel TARs previously identified by high-density oligonucleotide tiling arrays, and regions that were not previously shown to be transcribed (nonTx regions) using 5'-RACE and 3'-RACE and DNA sequencing [15,18,20]. The chromosomal regions for our analysis are primarily from the ENCODE regions of chromosome 22, which is particularly well annotated, as well as additional ENCODE regions on chromosomes 11 and 21. The RNAs analyzed were from NB4 acute promyelocytic leukemia cells, HeLa cells, and placental tissue. Both polyA+ and total RNA were used. A summary of the experiments performed is presented in Table 1.

In total, 420 regions were analyzed; primers to each strand were designed and subjected to 5'-RACE and 3'-RACE reactions for a total of 1,680 reactions. Approximately 80% of the reactions generated products that were detected by gel electrophoresis (see Additional data file 1 for examples); 25% of these reactions yielded heterogeneous products (smears). The entire PCR reaction was subjected to DNA sequence analysis, and approximately 40% of the sequence reads mapped to the expected locations of the genome and were therefore deemed as products derived for the intended locus (see Materials and methods, below, for details regarding mapping of RACE sequences to the genome and the fitness score assignment). The average length of these sequence reads is 516 base pairs (bp). As expected, primers designed in known exons gave the highest proportion of valid RACE products. This is followed by the primers designed to the novel TARs. The nonTx regions gave the fewest RACE products (Figure 1). Similar results were observed with both polyA+ and total RNAs, as well as from human cell lines or tissue.

We first analyzed the RACE sequences from eight known gene loci. For six of these loci we analyzed RNA from cells in which the gene was known to be expressed. For two genes, 5'-RACE and 3'-RACE reactions were performed using primers designed to the forward and reverse strand of each exon. For an additional four genes we analyzed a subset (1 to 8) of exons in the gene. As shown in Figure 2, the sequences of the known loci mostly matched the known annotations. For example, analysis of the DRG1 and FBXO7 genes, which are known to be expressed in NB4 cells, revealed cDNA sequences that matched the expected transcripts described in Refseq. In addition to detecting known transcripts, we also found novel isoforms. Some of these isoforms contained new exons whereas others contained different combinations of the known exons. An example is shown in Figure 2b for FBXO7. A novel exon was found for one of the RACE products and a novel combination was observed for another product. For the six genes analyzed we found evidence for 16 novel isoforms.

We also analyzed expression of two gene loci, namely SYN3 and TIMP3, in cells in which their expression was not detected by tiling microarray analysis. SYN3 and TIMP3 are encoded on opposite strands from one another on chromosome 22. SYN3 (Homo sapiens synapsin III mRNA) encodes a neuronal phosphoprotein that is involved in synaptogenesis and in the modulation of neurotransmitter release, and it is implicated in several neuropsychiatric diseases such as schizophrenia [21,22]. TIMP3 encodes tissue inhibitor of metalloproteinase 3. Mutations in this gene have been associated with the autosomal dominant disorder Sorsby's fundus dystrophy [23]. NB4 RNA hybridization to high-density oligonucleotide tiling arrays did not produce signal above background in the SYN3/TIMP3 region. With RACE sequencing a number of products were observed. Most RACE sequences (eight) matched that of the annotated RefSeq isoforms for SYN3 (NM_003490.2). RACE sequences also revealed three other novel isoforms with exon skipping and intron inclusion (Figure 3a). Similar results were found for TIMP3. The presence of additional RNA isoforms suggests that additional messages are probably produced from each gene locus.

To gain a better understanding of why the SYN3 and TIMP3 genes were not detected by microarray analysis, we examined their expression level by real-time quantitative PCR. As shown in Figure 3b, the expression levels of SYN3 and TIMP3 are 1 × 10⁴ and 1 × 10⁵ times lower than that of the HPRT1 transcript. HPRT1 is expressed at low levels in various cell lines and tissue types, with fewer than 8 to 15 serial analysis of gene expression (SAGE) tags per 200,000 (<10^-5), according to the SAGE Anatomic Viewer [24]. Thus, the transcripts produced by Syn3 and TIMP3 in NB4 cells are present at an extremely low level.

The novel RNA isoforms from annotated genes were examined for their ability to produce novel protein isoforms. The 16 novel RNAs identified in this study can produce five novel protein isoforms.

Antisense transcription plays diverse and important biologic roles, and recent studies using reverse transcription based approaches have reported a large amount of antisense transcription in the human genome [16,19]. Our study employed primers to analyze transcription from both DNA strands and thus examined antisense transcription. In addition to detecting transcription from the expected DNA strand, the RACE experiments produced sequences from the complementary strand for five of eight known gene loci. These sequences were revealed in experiments using both natural tissue and cell lines. However, careful inspection revealed that in most cases (29 out of 35) the splice junctions of most of the antisense products are not consistent with the GT-AG, GC-AG, or AT-AC pattern. Instead, they merely mirror (reverse complement) the splice junctions of the sense products. Two examples are shown in Figure 2 for the DRG1 region and the FBX07 regions (known genes on the plus strand). In these regions large numbers of antisense products (14 and 21, respectively) were detected on the opposite strand from both 5'-RACE and 3'-RACE reactions. Most of the antisense products lack the (GT-AG, GC-AG, or AT-AC) consensus splice sequences. It therefore appears likely that many of these antisense products are derived from the in vitro reverse transcription reaction, where double strand cDNAs might have formed [25], or from a complementary RNA in vivo [16].

To investigate further whether the antisnese transcript may be an artifact due to reverse transcription, we employed a novel strategy, namely direct chemical labeling of RNA followed by strand-specific oligonucleotide tiling microarray analysis. As shown in Figure 2c for the DRG1 locus, hybridization of cDNA prepared from NB4 cells using reverse transcriptase to the strand-specific microarray produced both sense and antisense signals. However, hybridization of RNA that has been labeled directly by chemical means, thus omitting the use of reverse transcriptase, usually yielded signals only from the annotated strand. This experiment indicates the much (but not all) antisense signal is directly tied to the use of reverse transcriptase and not likely to be present in vivo.

In addition to examining annotated genome regions, we analyzed a large number of novel TARs by RACE sequencing in order to gain a better understanding of their structure, their connectivity to known genes, and whether they might encode proteins of significant length. In all, 856 RACE reactions were generated to 214 TARs of the ENCODE regions [18]. End sequencing of the 5'-RACE and 3'-RACE products on both strands of the genome revealed overlapping sense and antisense transcripts (Figure 4a). This is consistent with recent work by Kapronov and coworkers [5,11,26] using RACE microarray experiments, although they did not analyze the transcript structure. In addition to analyzing TARs, we designed primers to 140 regions not known to produce transcripts; 17% of the primers were able to generate a RACE product whose sequence mapped to the expected region (Figure 1). Control experiments lacking reverse transcriptase do not produce products, indicating that the products are derived from RNA and not contaminating DNA. This frequency of successful RACE products from nonTX regions is lower than in known exon or novel TAR regions, but it nonetheless indicates that a substantial fraction of the human genome produces RNA. The transcripts from the nonTx regions exhibited an interleaved distribution similar to those from the novel TARs (Figure 4b).

The majority (85%) of the RACE sequences from the TARs and nonTX regions map contiguously (without introns) to the genomic sequence. Products from primers that lie close together on the genome often overlap one another or known exons, suggesting extensive transcription throughout the entire region. In addition, whereas the RACE sequences derived from known exons are mostly connected with known exons, the sequences from nonTx regions are rarely connected to others (Figure 5a). Although some of the regions yield results consistent with discrete transcripts, many do not.

Approximately 16% and 11% of the products produced from TARs and nonTx regions, respectively, produce transcripts that are spliced with consensus GT-AG, GC-AG, or AT-AC splice sequences (see Consensus splice site analyses [under Materials and methods, below]; Figure 5b). This is in contrast to products produced from exons in which approximately 50% of the messages are spliced. Moreover, further analysis of the novel TARs revealed that the RNA sequences with consensus splice sites originated from regions with higher microarray signal intensity on average than the unspliced ones (Figure 5c). Microarray signal intensity of the nonTx regions is close to background for both spliced and unspliced RACE sequences.

In order to determine better whether the novel transcripts may be functional, we examined their ability to encode protein. The sequences of RACE products were analyzed with respect to whether they contain open reading frames (ORFs) and/or whether the potential protein coding sequences are homologous to those in the nonredundant protein database. For two spliced sequences and 25 unspliced sequences, potential ORFs were found that have at least 50 codons, and the predicted protein sequence was homologous to that of a known protein present in the nonredundant database with a BLASTX threshold score of 1 × e^-9 [27]. The 27 transcripts contain 20 unique proteins, and nine out of 20 protein encoding ORFs have a translational start and stop codon (11 of the 27 transcripts).

One example of a potential protein coding transcript is shown in Figure 6. The novel transcript 5NGSP2F8 detected by RACE end sequencing was properly spliced with a consensus pattern. It encodes a potential ORF that is 142 codons in length. Evidence for the transcript is also supported by a spliced expressed sequence tag (EST), although for 5NGSP2F8 the EST sequence contains a shorter ORF, presumably through DNA sequencing errors.

We examined the expression level of novel transcript 5NGSP2F8 using real-time quantitative PCR. The 5NGSP2F8 expression level is more than 1,000-fold lower than that of the HPRT1 transcript, indicating that the gene is expressed at a low level (Figure 6b).

The regions of our analysis are selected mainly from the chromosome 22 ENCODE region, with additional targets in chromosome 11 and 21 ENCODE regions. Except for a few regions for test purposes, we selected most of the exon and novel TAR primer regions from among those expressed (cell type specific) regions in known exons and novel TAR regions detected by transcriptional tiling array experiments. The nontranscribed primer regions are selected in a tiled manner from among those regions that are neither known exons nor novel TARs.

We designed four primers for each targeted region, which can be exons of known gene, TAR, or previously identified untranscribed regions. Two gene-specific primers (GSP1 and GSP2) and two nested GSPs (NGSP1 and NGSP2) on both plus and minus strand were selected for each targeted region using a modified Primer3 program. The primers are 23 to 28 nucleotides long, with GC content of 50% to 70% and with Tm (melting temperature) above 70°C (optimally 73°C to 74°C). Self-complementary primers that could form hairpin were avoided. We also voided complementarity between GSPs and UPM (universal primer A in the SMART RACE™ kit [Clontech, Mountain View, CA, USA]), particularly in their 3' Ends (UPM long: 5'-CTAATACGACTCACTATAGGGCAAGCAGTGGTATCAACGCAGAGT-3'; UPM short: 5'-CTAATACGACTCACTATAGGGC-3'). Complementarity between NGSPs and NUP (nested universal primer A), particularly in their 3' ends, was avoided (NUP: 5'-AAGCAGTGGTATCAACGCAGAGT-3'). The primers were mapped against the genome to ensure that they mapped to only one location (with identity <80% to other locations).

Human NB4 cell line total RNA, Hela S3 polyA+ RNA, placenta total RNA, and polyA+ RNA (Ambion, Austin, TX, USA) were used in cDNA amplification by SMART RACE™ kit (Clontech), in accordance with the manufacturer's instructions [43]. 5'-RACE-Ready cDNA and 3'-RACE-Ready cDNA were synthesized using PowerScript (Clontech, CA, USA) or Superscript II (Invitrogen, CA, Carlsbad, USA) moloney murine leukemia virus reverse transcriptase and SMARTII A oligo (5'-AAGCAGTGGTATCAACGCAGAGTACGCGGG-3'), 5'-CDS primer A ([5'-(T)25V N-3' (N = A, C, G, or T; V = A, G, or C)]), or 3'-CDS primer A (5'-AAGCAGTGGTATCAACGCAGAGTAC[T]30V N-3' [N = A, C, G, or T; V = A, G, or C]). A total of 1 μg RNA was used in a final volume of 10 μl reverse transcription reaction (100 ng/ul). A reverse transcription reaction without reverse transcriptase was used as negative control to distinguish genomic DNA contamination. RACE was followed by PCR amplification using UPM and GSP1 or GSP2 on both strands of the genome. A 0.5 μl reverse transcription reaction from the above was used in 50 μl of PCR reaction in the Advantage™ 2 PCR Enzyme System (Clontech). Nested PCRs were performed using NUP (5'-AAGCAGTGGTATCAACGCAGAGT-3') and NGSP1 or NGSP2. One microliter of RACE PCR product was used in a 50 μl reaction. The PCR program was 94°C for 30 seconds and 72°C for 3 minutes, five cycles; then 94°C for 30 seconds, 70°C for 30 seconds and 72°C for 3 minutes, five cycles; followed by 25 cycles of 94°C for 30 seconds and 68°C for 30 seconds; and concluding with by an extension cycle of 72°C for 3 minutes. Nested PCR products were end sequenced using NGSP1 or NGSP2. The RACE sequences have been submitted to GenBank EST database (accession numbers from EW712308 to EW712635).

We first use the BLAT alignment tool [44] to compare all the RACE sequence reads to the human genome assembly (hg17, May 2004), and then evaluated the 'fitness scores' of the BLAT output matches using the following formulas:

sizeDif = abs([tEnd - tStart] - [qEnd - qStart]) + abs(qSize - [qEnd - qStart])

insertFactor = qNumInsert + tNumInsert

total = matches + repMatches + misMatches

badness = (1,000 × misMatches + insertFactor + 3 × log[1 + sizeDif])/total

fitness = 100 - badness × 0.1

Where parameters such as tEnd have the same meanings as those defined in the BLAT documentation. The fitness score is based on the 'percent identity score' in the University of California at Santa Cruz Genome Browser [45], and it includes additional penalty on small overall matches. Once these fitness scores have been computed for one RACE experiment, a distribution of these scores was derived based on the characteristics of those BLAT matches that are located on the 'correct' chromosomes, and only those 'correct' matches with scores above a certain threshold were kept as 'valid' products and correspondingly as 'valid' transcripts. See Additional data file 2 for further details.

For those BLAT matches with multiple blocks, the corresponding splice sites in the transcripts were further examined in the following way. A splice site is defined as a consensus one if and only if a 'GT-AG' (or 'GC-AG'/'AT-AC', which appear much less often) pattern can be observed within windows of eight nucleotides on the two ends of it. (For example, for a splice site starting at chromosome position i and ending at j, the windows are [i - 3, i + 5) and [j - 5, j + 3].) An overall consensus score was then assigned to each transcript according the proportion of consensus splice sites in all its splicing events. We also used this consensus splice site criteria to filter out mirroring antisense transcripts caused by experimental artifacts. See Additional data files 3 for discussions on the choice of window size and the resulting transcripts after this filtering.

Normalized signal intensities from across tiling array experiments were extracted for those primer regions and correspondingly assigned to the transcripts. These signal intensities were correlated with different transcript characteristics such as splicing events in our analysis.

The 'valid' transcripts were also compared against RefSeq gene annotation [46] and the union of all of the novel TARs for connectivity information. For each transcript, its connectivity to known RefSeq genes is the number of unique known exons that overlap with this transcript, and the connectivity to novel TARs is the number of novel TARs that overlap with this transcript.