A Transcript Finishing Initiative for Closing Gaps in the Human Transcriptome

Abstract

We report the results of a transcript finishing initiative, undertaken for the purpose of identifying and characterizing novel human transcripts, in which RT-PCR was used to bridge gaps between paired EST clusters, mapped against the genomic sequence. Each pair of EST clusters selected for experimental validation was designated a transcript finishing unit (TFU). A total of 489 TFUs were selected for validation, and an overall efficiency of 43.1% was achieved. We generated a total of 59,975 bp of transcribed sequences organized into 432 exons, contributing to the definition of the structure of 211 human transcripts. The structure of several transcripts reported here was confirmed during the course of this project, through the generation of their corresponding full-length cDNA sequences. Nevertheless, for 21% of the validated TFUs, a full-length cDNA sequence is not yet available in public databases, and the structure of 69.2% of these TFUs was not correctly predicted by computer programs. The TF strategy provides a significant contribution to the definition of the complete catalog of human genes and transcripts, because it appears to be particularly useful for identification of low abundance transcripts expressed in a restricted set of tissues as well as for the delineation of gene boundaries and alternatively spliced isoforms.

A primary objective of the Human Genome Project has been the identification of the complete set of human genes and their derived transcripts. A major step towards this goal was achieved at the beginning of 2001 with the publication of two independent draft versions of the human genome sequence and the identification of >30,000 genes (Lander et al. 2001; Venter et al. 2001). However, it became apparent that extracting exonic sequences directly from the human genome is not straightforward and that a variety of complementary strategies are required for gene identification and characterization.

In this context, microarrays (Penn et al. 2000; Dennis 2001; Schoemaker et al. 2001; Kapranov et al. 2002) and sequence comparisons with other organisms at an appropriate evolutionary distance (Batzoglou et al. 2000; Roest et al. 2000) constitute powerful preliminary approaches to identifying transcribed regions within the genome sequence. Nevertheless, transcript sequencing is necessary both for the final proof of the existence of an expressed gene and for the precise identification of intron/exon boundaries and alternatively spliced forms (Camargo et al. 2002).

A full-length cDNA sequence, ideally including a transcription initiation site and a polyadenylation site, is the gold standard for transcript definition. Considerable progress has been made in the generation of representative full-length cDNA sequences (Strausberg et al. 1999, 2002; Wiemann et al. 2001; Kikuno et al. 2002; Nakajima et al. 2002), especially following the development of sophisticated protocols for obtaining full-length transcript molecules and to correct for transcript expression bias (Bonaldo et al. 1996, Carninci et al. 2000).

Expressed sequence tags (ESTs) are another major source of transcript sequence. ESTs either are single-pass, partial sequences derived either from the 5′ and 3′ extremities of cDNA clones (Adams et al. 1991) or are specifically directed towards the central coding regions of transcripts, in the case of open reading frame ESTs (ORESTES; Dias et al. 2000; Camargo et al. 2001). Initially, ESTs were exploited as a source for gene discovery (Adams et al. 1992, 1993), but they have also been widely used to build tissue-specific transcript profiles (Bortoluzzi et al. 2000a,b,c; Huminiecki and Bicknell, 2000; Phillips et al. 2000; Yu et al. 2001; Katsanis et al. 2002; Megy et al. 2003), to construct gene-based physical maps (Hudson et al. 1994), to compare genomes of different organisms (Tugendreich et al. 1994; Lee et al. 2002), to accurately identify transcripts in genomic sequences (Bailey et al. 1998; Jiang and Jacob 1998; Kan et al. 2001), and to study aspects of mRNA structure, such as splicing variants (Hide et al. 2001; Modrek et al. 2001; Clark and Thanaraj 2002; Kan et al. 2002; Xie et al. 2002; Xu et al. 2002; Wang et al. 2003), alternative polyadenylation (Gautheret et al. 1998; Beaudoing and Gautheret 2001; Iseli et al. 2002), and single nucleotide polymorphisms (Garg et al. 1999; Picoult-Newberg et al. 1999; Clifford et al. 2000; Irizarry et al. 2000; Hu et al. 2002).

To date, >5,200,000 human ESTs have been generated from different organs and tissues, deriving mainly from the Merck Gene Index Project (Williamson 1999), the Cancer Genome Anatomy Project (CGAP; Strausberg et al. 2000), and the Human Cancer Genome Project Ludwig/FAPESP (HCGP; Dias et al. 2000; Camargo et al. 2001). Nevertheless, it is widely recognized that EST databases are subjected to artifacts related to the partial, low-quality nature of the sequences and the presence of various kinds of contamination (Sorek and Safer 2003). In addition, because of the large differences in abundance between RNA species, the coverage of individual transcripts by ESTs is highly variable. Despite that, it is believed that the vast majority of transcripts have been sampled at least once by either a full-length cDNA or EST sequence (Ewing and Green 2000; Liang et al. 2000).

Although the amount of transcript data currently available is not sufficient to identify all human genes, the judicious use of this data set, in conjunction with the draft sequences of the human genome, has been highly informative in the characterization of new human genes (Reymond et al. 2002; Silva et al. 2003). Here we describe the utilization of a transcriptome database to guide the generation of novel human transcript sequences on a genome-wide basis. By using the genomic sequence as a scaffold for EST mapping and clustering, we have used RT-PCR to bridge gaps between EST clusters that we judged as likely to be derived from the same genes. The resulting novel sequence confirms that the ESTs from different clusters are, in fact, derived from a common transcript and defines the intervening region between them.

Because this process is very similar to the finishing phase of genome projects, we called it transcript finishing (TF). This powerful, albeit laborious, approach allows the characterization of novel human transcripts and splicing isoforms, which appear to be generally expressed at a low abundance level and/or in a restricted set of tissues and avoids the necessity of a full-length cDNA clone in order to confirm the structure of a gene.

RESULTS

Generation of the Transcriptome Database and EST Cluster Selection for Experimental Validation

We have used the publicly available human genome and transcript sequences to identify and experimentally validate additional transcribed regions in the human genome. The two data sets were integrated into the transcriptome database by using the BLASTN program to map all transcript sequences onto the assembled version of the human genome available from the National Center for Biotechonology Information (NCBI). We have also mapped to the genome, using the raw data generated by EST sequencing projects, a set of trusted 3′tags that provide unique identifiers for transcript 3′ ends (Iseli et al. 2002). The tags were used for positional orientation and as a start point for transcript reconstruction. To facilitate visualization of the alignments and the access to information such as project and tissue source of the sequences, alignment scores, and the position of 3′tags, a graphical interface was also developed (Fig. 1).

Figure 1.

TFI graphical interface. The TFI graphical interface displays a region of the human genome sequence as a yellow line, with a scale in base pairs (bp). Expressed sequence tags (ESTs) that align with the genome sequence are shown in different colors, depending on the project of origin: ORESTES from the FAPESP/LICR Human Cancer Genome Project in purple; CGAP in green, MGC in blue, and TFI in yellow, with splicing structures represented as gray lines. The interface shows an experimentally validated TFU (number 171) joining two EST clusters. The TFI interface also provides information on the tissue of origin of the transcript sequences, the percentage of similarity of each exon with the human genome sequence, and the presence of 3′ tags represented as green triangles.

We identified 244,148 human transcript clusters, of which 14,598 contained at least one full-length cDNA sequence, and 229,550 clusters that were composed exclusively of partial transcript sequences. Of the set of 14,598 clusters containing full-length sequences, 13,149 (90%) had at least one corresponding EST, and the remaining 1449 (10%) were composed only of full-length cDNA sequences. These data demonstrate that, despite the fact that >5 million EST sequences are available, they do not fully cover the human transcriptome and that the generation of additional transcribed sequences is still required.

It is noteworthy that clusters composed exclusively by ESTs have a reduced number of sequences (average, 5.9 sequences) derived from fewer different tissues (average, 3.0), compared with clusters containing a full-length cDNA, which have an average of 65.5 sequences derived from eight different tissues. Based on these observations, we conclude that the human transcripts that remain to be defined are expressed at low levels in a restricted set of tissues and that their characterization will benefit from a direct approach such as the TF.

Because EST databases contain a significant fraction of artifactual and contaminant sequences, we selected, for experimental validation, pairs of clusters that consist of ESTs that align noncontiguously to the genome, consistent with the presence of a splicing structure. We also restricted our validation to pairs of clusters that map at a maximum distance of 10 kb from each other, in order to increase the probability that these clusters belong to the same transcript. By using these criteria, a total of 2373 pairs of clusters (∼2% of the total number of clusters composed of partial sequences) were initially selected and subjected to manual inspection using our graphical interface.

Manual inspection allowed the assessment of similarity and extension of the alignments, as well as the position of the selected pair of clusters relative to the 3'tags. Following this procedure, a subset of 489 pairs of clusters was initially selected for experimental validation. The number of clusters eliminated by manual inspection was very low; therefore, the 489 pairs of clusters selected for experimental validation can be considered an unbiased sample of the 2373 initially selected clusters. Clusters selected for validation were separated from each other by an average 2879 bp of intervening genomic sequence and were composed by an average of 5.92 EST sequences derived from an average of three distinct tissues. Each pair of EST clusters selected for experimental validation was designated as a single TF unit (TFU). Information related to the 489 TFUs selected for validation can be accessed at http://200.18.51.201/viewtfi.

Experimental Validation and the Generation of New Transcribed Sequences

A general overview of the computational and experimental validation strategies is presented in Figure 2. A total of two coordination groups, four bioinformatics groups, and 29 validation laboratories, linked through the Internet, participated in the computational and experimental phase of the project (http://200.18.51.201/transcript/Participants.html). Following cluster selection and manual inspection, primers for RT-PCR validation of each TFU were designed automatically. The genomic sequence was chosen as a template for primer design because it is generally of a higher quality than are EST sequences. cDNA preparation was also a critical issue, because both the quality and the representation of different tissues directly influence the validation efficiency. As an indicator of genomic DNA contamination, the total RNA preparations were subjected to PCR amplification by using primers within intronic sequences flanking the exon 12 of the MLH1 gene and found to be negative. The quality of the cDNA product was demonstrated by PCR amplification of sequences located at the 5′ extremity of the NOTCH2 transcript (a long transcript of 11.4 Kb). A total of 22 cDNA preparations, derived from a number of cell lines and representing 18 distinct tissues, were used.

Figure 2.

General scheme of the TFI strategy. Schematic outline of the strategy used for computational and experimental validation of TFU sequences. Following the development of bioinformatics tools, the generation of the transcriptome database, and automatic cluster selection, the project tasks were divided between the coordination and the validation laboratories.

The total of 3019 sequences, generated during the project, was subjected to an automated cleaning protocol. High-quality sequences were aligned against the genomic sequence, and the alignment coordinates and scores for validated sequences were loaded into the transcriptome database and displayed on the graphical interface (Fig. 1). We successfully validated 211 of the 489 TFUs that were distributed, yielding an overall validation efficiency of 43.1%.

A single pair of primers was tested for each TFU, and experimental validation was undertaken in a high-throughput single-pass format. Few modifications were adopted when a positive amplification was not achieved (see Methods). To estimate the false-negative amplification rate of the TF strategy, we determined the number of the nonvalidated TFUs for which a full-length cDNA sequence had been made available by other sequencing projects during the course of our project. For 40 of the nonvalidated TFUs, we were able to identify a full-length sequence linking the two EST clusters initially selected for validation. Thus, these cases can be considered to be false-negative amplifications. For 118 of the nonvalidated TFUs, the existence of a full-length sequence matching just one of the two selected clusters allowed us to conclude that the two clusters were in fact part of different transcripts. For these cases, the absence of an RT-PCR product thus reflects true negatives. For the 120 of the remaining nonvalidated TFUs, a conclusive result could not identify any corresponding full-length sequence. Therefore, based on these results, we can estimate that the rate of false-negative amplifications in the TF strategy is ∼25% (40/158 nonvalidated TFUs).

In addition, to identify variables related to the expression pattern of the novel transcripts that influence the efficiency of validation, two sets composed of 174 validated TFUs and 208 nonvalidated TFUs were compared. As shown in Table 1, the validated TFUs had, on average, more ESTs in each cluster derived from a larger number of different tissues. Both of these differences were statistically significant according to Mann-Whitney tests, indicating that a higher expression level and a broader expression pattern of the selected transcripts favored validation. The presence of ESTs derived from the same tissue in both clusters did not influence the likelihood of validation according to χ² tests.

Table 1.

Comparison Between Validated and Nonvalidated TFUs

	Validated TFUs (Std deviation)	Nonvalidated TFUs (Std deviation)	P value
Average distance between clusters	2609 (3202)	3105 (2942)	0.008
Average no. of ESTs in each cluster	6.10 (8.91)	5.77 (13.23)	0.010
Average no. of distinct tissues in each cluster	3.45 (4.27)	2.85 (4.54)	0.002
Presence of a common tissue in both clusters	Yes 63	Yes 62	0.223
	No 111	No 146

A total of 59,975 bp of transcribed sequence, organized into 432 exons, were generated, contributing to the definition of the structure of 211 distinct human transcripts. Each validated TFU had a mean of 281.6 bp and a median of 207 bp of novel sequence not represented by the original EST clusters and a mean of 2.03 and a median of two newly defined exons. The validated TFU sequences have been submitted to GenBank under the accession numbers CF272536 to CF272733, which are provided as Supplemental Table 1.

Consensus Sequences Generation and Annotation of the Validated Human Transcripts

Consensus sequences produced by assembling the sequences derived from the validation fragment and the sequences from all ESTs in both clusters were obtained for 186 of the 211 validated TFUs. Assembly of a consensus sequence was not possible for 25 TFUs, due mainly to the presence of repetitive sequences and alternative splicing forms. Consensus sequences, with an average of 1240 bp, can be accessed at (http://200.18.51.201/viewconsensus/).

Consensus sequences derived from the validated TFUs were aligned to the July 2003 version of human genome sequence assembly provided by the University of California, Santa Cruz (UCSC), using the BLAT search tool (http://genome.ucsc.edu/cgibin/hgBlat) to compare the validated consensus sequences with known genes and gene predictions (Table 2). A significant fraction (68.8%) of the validated transcripts completely overlapped with the alignment coordinates of a known gene or full-length human mRNA submitted to the GenBank during the course of our project (Fig. 3A), and a smaller fraction (10.2%) represented extensions (mostly 5′) to partial cDNA sequences deposited in public databases (Fig. 3B). However, for 21% of the validated TFUs, a full-length cDNA sequence was not available in public databases as of July 2003. The structure of the majority (69.2%) of the validated TFUS without a corresponding full-length cDNA sequence had not been correctly predicted by ab initio gene prediction programs such as Fgenesh++, Geneid, and GenScan. These TFUs can thus be considered as new human transcripts.

Table 2.

Annotation of Validated Consensus

Categories	Number of consensus sequences	Percentage (%)
Known gene	128	68.8
Extension of a known gene	19	10.2
New transcript w/total prediction	12	6.5
New transcript w/partial prediction	15	8.0
New transcript w/o prediction	12	6.5
Total	186	100

Figure 3.

Characterization and annotation of validated TFUs. Alignment of four consensus sequences, derived from the validated TFUs, to the July 2003 version of the UCSC human genome sequence assembly, using the BLAT search tool. (A) TFU00023 corresponds to YourSeq (black) completely overlapping with known genes based on SWISS-PROT, TrEMBL, mRNA, and RefSeq (dark blue). (B) TFU01102 represents a 5′ extension of a partial cDNA (FLJ23834). (C) TFU01013 represents a new human transcript structure that was correctly predicted by ab initio gene predition transcripts, such as Fgenesh++ (green). (D) TFU00125 represents a new human transcript with no predicted transcripts described by gene prediction programs.

The consensus sequences corresponding to new human transcripts were further characterized by BLASTX analysis, and protein domains were predicted by using the Pfam and Prosite databases. Of the 39 consensus sequences representing new human transcripts, 27 (69.2%) contained an ORF of at least 100 amino acids, and eight (20.5%) contained a clearly defined protein domain including three IG-like domains and a protein kinase. Complete information on the characterization of the validated TFUs, including consensus size, annotation, chromosomal location, and expression pattern based on ESTs distribution, are provided as Supplemental Table 2.

The validated transcripts that completely overlapped with the alignment coordinates of a known gene containing a defined ORF were used to estimate the percentage of consensus sequences that represent complete transcripts. Only a small fraction (9.7%) of the 93 validated TFUs analyzed contained a complete ORF. The low percentage obtained was expected because, in the TF strategy, RT-PCR is used to brigde gaps between partial transcript sequences.

Identification and Experimental Validation of Alternatively Spliced Isoforms

Several reports have suggested that at least 30% to 35% of human genes undergo alternative splicing (Brett et al. 2000; Modrek et al. 2001); nevertheless, this value is probably underestimated because many cell types have not yet been fully explored by cDNA sequencing. The use of different cDNA sources during the experimental validation phase of the new human transcripts allowed us to identify many new splicing variants. We explored the degree of sequence variability due to alternative splicing in the set of 186 consensus sequences that we generated and found evidence for alternative splicing in 22 (12%) cases (Table 3). Intron retention was observed in 11 TFUs, and alternative exon usage was detected in 11 of the 22 TFUs with alternative splicing. Conserved GT-AG splice junctions were present in all TFUs with alternative exon usage. The possibility of genomic DNA contamination, in those cases in which we have observed the retention of an intron, was excluded due to the presence of processed introns in the same cDNA molecule containing the retained intron. Moreover, the RNAs used for experimental validation of the alternatively spliced forms have been treated with DNase and tested for the absence of intronic sequences, as described in Methods.

Table 3.

Alternative Splicing Forms Within Validated TFs

Validated consensus	Type of alternative splicing	Presence of conserved acceptor and donor sites	No. of alternative isoforms	No. of validated isoforms
TFU0118	Exon usage	Yes	2	1
TFU0200	Exon usage	Yes	4	4
TFU0274	Exon usage	Yes	2	2
TFU0351	Exon usage	Yes	2	2
TFU1004	Exon usage	Yes	2	1
TFU1058	Exon usage	Yes	3	0
TFU0155	Exon usage	Yes	2	nd
TFU0238	Exon usage	Yes	2	nd
TFU0308	Exon usage	Yes	2	nd
TFU0003	Intron retention	nd	nd	nd
TFU0019	Intron retention	nd	nd	nd
TFU0035	Intron retention	nd	nd	nd
TFU0052	Intron retention	nd	nd	nd
TFU0099	Intron retention	nd	nd	nd
TFU0112	Intron retention	nd	nd	nd
TFU0125	Intron retention	nd	nd	nd
TFU0131	Intron retention	nd	nd	nd
TFU0209	Intron retention	nd	nd	nd
TFU0285	Intron retention	nd	nd	nd
TFU0371	Intron retention	nd	nd	nd
TFU0148	Exon skipping	nd	nd	nd
TFU1061	Exon skipping	nd	nd	nd

nd = not done.

We selected six TFUs with alternative exon usage, representing a total of 14 splicing isoforms, for further experimental validation. Each pair of primers used for experimental validation of the alternatively spliced forms was assayed against all 22 cDNA sources, without pooling. Touchdown PCR confirmed 10 (83%) of the putative investigated isoforms. No PCR amplification was achieved for one TFU. Some splicing isoforms were expressed in a restricted pattern, being detected in one or a few of the tissues analyzed by RT-PCR (data not shown). None of these splicing isoforms had been previously identified, highlighting the potential use of the TF strategy for uncovering the genetic variability generated at the transcriptome level.

A typical example of this experimental validation is illustrated in Figure 4. In this case, we were able to identify two alternative exons, one of which presents an extra exon of 138 bp and the other a 21-bp extension of an exon already represented by EST sequences. The possible combination of these variants results in four splicing isoforms. Figure 4 shows a 388-bp product (obtained with primers P1 and P2) corresponding to the proto-type isoform, a 370-bp product (primers P2 and P3) corresponding to the isoform containing the additional exon, a 314-bp product (primers P1 and P4) corresponding to the isoform with the extended exon, and a 452-bp product corresponding to the isoform containing both the additional exon and the extended exon.

Figure 4.

Experimental validation of MGC5601 gene alternative splicing isoforms. (A) Gene structure for exons XVI-XIX (boxes) of the MGC5601 gene located on chromosome 12. Introns are represented by lines. Two alternative exons are shown on TFU reads, and a hypothetical combination of these two exons is also shown. Sequence F07R has an extra exon between exon XVII and XVIII. Sequence A01R has an extended exon XIX. Four primers were designed for validation tests, as indicated in the figure (P1–P4), and each pair of primers were assayed against all 22 cDNA preparations without pooling. (B) We detected all four of these alternative splicing isoforms in MGC5601. Numbers one through four indicate the tissues from which the cDNA was obtained (1, multiform glioblastoma; 2, glioblastoma; 3, prostate carcinoma; and 4, primary kidney cell culture). The sizes of the bands obtained are indicated. L indicates 100-bp ladder.

DISCUSSION

Currently, intense activity is directed toward defining the complete set of genes and their derived transcripts in the human genome. This information will have a profound impact in diverse areas of biology such as human evolution, structural genomics, and medicine. However, because of the highly dispersed and complex structure of human genes, it is extremely difficult to correctly identify transcribed regions within the genome (Camargo et al. 2002).

Estimates based on gene prediction both within individual finished chromosomes (Dunham et al. 1999; Hattori et al. 2000), as well as in the draft human genome sequences (Lander et al. 2001; Venter et al. 2001), have uniformly concluded that the human genome possesses <35,000 genes. This number has been supported by a preliminary analysis of EST coverage of known genes (Ewing and Green 2000) as well as comparative genomics analysis (Roest et al. 2000). Most of these 35,000 genes are already represented by a full-length cDNA sequence in transcript databases. In UniGene (http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?db=unigene), for example, there are currently 28,412 transcript clusters represented by at least one full-length cDNA sequence.

Here we have proposed and validated the use of the TF strategy for characterization of new human transcripts that are only partially represented by ESTs. Because EST databases contain a significant fraction of artifactual and contaminant sequences, we selected pairs of clusters for experimental validation that exhibited a clear splicing structure when aligned to the genome. By requiring the occurrence of splicing, the level of contamination in the EST databases is significantly reduced, although at the expense of eliminating many genuine 3′ ESTs. The selection criteria used in our initial analysis are very restrictive, and the adoption of less stringent criteria (including clusters without a splicing structure) will certainly be required to complete the catalog of human genes using the strategy we described. Given the 2373 initially selected clusters, of which 489 were subjected to experimental validation, 1884 pairs of clusters remain to be validated. If we assume an overall validation efficiency of 43%, we can estimate that the TF strategy might contribute to the definition of at least 791 additional genes in the human genome.

Several factors may have influenced our validation efficiency, including experimental limitations related to primer and cDNA synthesis, the particular characteristics of human transcripts such as low expression level, and the existence of a significant proportion of sense-antisense transcriptional units on opposite DNA strands of the same genomic locus (Yelin et al 2003). A 25% false-negative amplification rate was estimated for the TF strategy and is probably related to the high-throughput single-pass format adopted for the experimental validation. In this context, the use of additional primer pairs and modifications of cycling parameters that would favor the amplification of difficult targets could be added to the process to reduce the negative amplification rate.

We found that validation efficiency was enhanced by implementation of quality controls for cDNA synthesis, the use of polyA+-derived cDNA, a combination of both oligo dT and random primers for cDNA synthesis, and also the use of nested RT-PCR. We also observed that validated pairs of clusters had a higher average number of ESTs per cluster and a higher number of different tissues represented by the clusters compared with pairs of clusters that we were not able to validate. Validated TFUs had, on average, 6.1 ESTs in each cluster derived, on average, from 3.45 distinct tissues. Noteworthy, in 41% of the validated TFUs, one of the two EST clusters was composed of a single EST, and in 13% of the cases, both clusters corresponded to singleton ESTs, indicating the often overlooked importance of this kind of data.

For a reasonable fraction (21%) of the validated TFUs, a full-length cDNA sequence was not yet available in public databases. The structure of the majority (69.2%) of these validated TFUs had not been correctly predicted by ab initio gene prediction programs and, consequently, was not annotated in the human genome. In addition, the use of different cDNA sources in the validation process allowed us to identify many splicing variants that were further validated by RT-PCR. As for 21% of validated sequences, none of these splicing variants had been previously identified.

We conclude that the TF strategy provides a convenient and unique means for delineating gene boundaries and new transcribed sequences. The TF strategy permits the characterization of new human transcripts and splicing isoforms expressed at a low level and in a restricted set of tissues and will certainly continue to contribute to the definition of the complete catalog of human genes and transcripts.

METHODS

Cell Culture

Human cell lines were obtained from the American Type Culture Collection (ATCC) and cultured as recommended (http://www.atcc.org). The following cell lines were used in order to generate a cDNA panel representing different tissues: A172 glioblastoma; T98G multiform glioblastoma; FaDu squamous cell carcinoma; SW480 colorectal adenocarcinoma; Skmel-25 malignant melanoma; DU145 prostate carcinoma; HeLa cervix adenocarcinoma; XP Xeroderma pigmentosum fibroblasts; ZR-75-1, MCF-7, and Hs578T breast ductal carcinoma; IM9 B transformed lymphoblasts; TT thyroid carcinoma; U937 histyocytic lymphoma; Hs1.Tes normal testis; Hs732.PL normal placenta; Hep G2 hepatocarcinoma; NCI-H1155 and H358 lung carcinoma; SCaBER urinary bladder carcinoma; SAOS 2 osteosarcoma; and Tu-rim primary culture of a kidney tumor.

RNA Extraction and cDNA Synthesis

Total RNA was prepared from cultured cells seeded in four 150-mm-diameter (P150) plates by using the cesium chloride cushion technique (Chirgwin et al. 1979). Poly A⁺ RNA was isolated from 200 μg total RNA with the PolyAttract mRNA isolation kit (Promega), and the total yield of this purification was used for cDNA synthesis. For cDNA synthesis, 100 to 200 μg total RNA or the corresponding purified mRNA were treated with 100 U DNAse I (FPLC-pure, Amersham) and reverse-transcribed by using oligo(dT)12-18, random primer and SuperScript II (Invitrogen), following the manufacturer's instructions. The resulting cDNA was then subjected to RNase H treatment and distributed among the 31 validation laboratories involved in the project. The quality of the cDNA synthesis and the absence of genomic DNA contamination were evaluated for each preparation. Total RNA was subjected to PCR amplification by using primers within intronic sequences flanking exon 12 of the MLH1 gene (forward, 5′-TGGTGTCTCTAGTTCTGG-3′; reverse, 5′-CATTGTTGTAG TAGCTCTGC-3′). The quality of the cDNA product was also tested by PCR amplification of sequences located at the 5′ extremity of the NOTCH2 transcript (a long transcript of 11.4Kb; forward, 5′-ACTGTGGCCAACCAGTTCTC-3′; reverse, 5-′ CTCT CACAGGTGCTCCCTTC-3′).

RT-PCR and Sequencing

RT-PCR was carried out in 25 μL reaction mixtures containing 1 μL cDNA, 10× Taq DNA polymerase buffer, 200 μM dNTP, 6 pmoles of primers, 1.5 mM MgCl₂, and 1 U Taq DNA polymerase (GIBCO BRL). Standard PCR conditions were as follows: 4 min at 94°C (initial denaturation), 40 sec at 94°C, 40 sec at 55°C, and 1 min at 72°C for 35 cycles and a final extension step of 10 min at 72°C. Modifications of the standard protocol included annealing temperature, MgCl₂ concentration, addition of PCR enhancers such as betaine, and the use of polymerases with hot start activity. PCR products were directly sequenced with the same primers used for RT-PCR or cloned before sequencing. If more than one fragment was obtained for the same TFU using different cDNA sources, all fragments were sequenced. This was also the case if multiple bands were obtained in PCR amplifications using a single cDNA source. Sequencing different fragments obtained for a specific TFU allowed us to characterize a number of alternatively spliced transcripts. Sequencing reactions were carried out by using the DYEnamic ET terminator Cycle Sequencing Kit (Amersham Pharmacia) and separated by electrophoresis using an ABI 377 Prism Sequencer (Applied Biosystems) according to supplier's recommendations.

Transcriptome Database and Graphical Interface

BLASTN was used to identify pair-wise similarities between all known transcript sequences and the draft genome sequence deposited in release 66 (March 2001) of the European Molecular Biology Laboratory (EMBL) database. Transcribed sequence data were extracted from several sources: (1) the human EST section of EMBL release 66, (2) human mRNA documented in the human section of EMBL release 66, (3) ORESTES sequences from the Ludwig Institute for Cancer Research (LICR)/FAPESP Human Cancer Genome project, and (4) human mRNAs documented in the NCBI curated RefSeq database (http://www.ncbi.nlm.nih.gov/refseq). For genomic sequence, we used contigs of at least 10 kb deposited in the HUM and HTG sections. Those HTG entries that had not been fully assembled were split into individual components. Therefore, the human genome data set used is highly redundant but can easily be reduced to one of the available assemblies.

The transcript sequences were filtered for contaminants, and repetitive elements were masked out by using the PFP software package (Paracel). For each pair of matching transcribed and genomic sequences, local alignments were generated by using Sim4 (Florea et al. 1998), with parameters W = 15, R = 0, A = 4, and P = 1. The output of Sim4 was filtered to eliminate all alignments that did not contain at least one matching region within the genome with at least 95% identity over 30 nt. The alignment coordinates and related information were uploaded into a MySQL relational database. We used the data stored in the relational database to create clusters of transcribed sequences, based on their position within individual genomic contigs. The coordinates of the putative exons on the genome sequence were used to determine membership in a cluster. If coordinates of at least one exon were common to two transcripts, then these were considered to be part of the same cluster.

The 3′ tags were generated as previously described (Iseli et al. 2002). Briefly, poly(A) or poly(T) were identified from original sequence trace files, and the 50 nucleotides immediately adjacent to it were recorded as a candidate tag (after obtaining the reverse complement for poly(T) tracts). Duplicate tags were eliminated, as were the tags matching LINE and Alu repetitive elements, ribosomal or mitochondrial sequences, and those containing simple repeats. Matches for the remaining tags were mapped to the genome, and the 50 nucleotides found downstream of the match were also recorded. Individual tags were incorporated into the MySQL database. A graphical interface was developed in TCL/TK language in order to visualize the 3′ tags, EST alignments and related information, such as tissue origin and project source of the sequences.

By querying the transcriptome database, we were able to select EST clusters that do not correspond to known full-length mRNA for validation. These were at a maximum of 10 kb apart from each other and exhibited a clear splicing structure when aligned to the genome. Clusters selected for validation were visually inspected before ordering primers. All systems used in this work were developed by using PERL and PHP programming languages on a Linux-based server running the MySQL database management system and the Apache Web server.

Cluster Selection and Primers Design

The automated primer protocol received a fixed format file containing the accession number of the genomic clones and the genomic interval where the two noncontiguous EST clusters map and where the system searched for primers. A single pair of primers was designed for each TFU, which usually targeted the two exons flanking the putative gap. In a few cases, in which the presence of repetitive sequences or atypical base composition prevented the design of primers, adjacent exons were used. The Primer3 program (version 0.9) developed by the Whitehead Institute for Biomedical Research was used for primer design, adopting the following parameters: primer size of a minimum of 17 bp, optimal 18 bp and maximum 21 bp; melting temperature of a minimum of 55°C, optimal 60°C and maximum 65°C; and GC clamp set to one. The output of Primer3 was processed in order to filter primers that had alternate annealing sites in the given genomic sequence. The system uses a Web-based interface that allows submission of files containing information on primer design, retrieval of primers found, and the modification of default parameters for primer picking.

Sequence Analysis and Database Update

Sequences were subjected to an automated protocol to (1) assess sequence quality, (2) trim vector sequences, (3) mask repetitive elements, and (4) remove undesirable sequences such as bacterial, mitochondrial, and fungi sequences. The sequence quality was determined by Phred analysis using a trimCutOff of 0.06171 (Ewing and Green 1998; Ewing et al. 1998). Sequences with <100 bases were excluded. Mitochondrial, bacterial, and fungi sequences were identified by BLAST searches against the GenBank entry corresponding to the human mitochondrial complete genome sequence and against a locally developed bacterial and fungal database, respectively. Significant hits were determined by using an E value of 10^–5 for searches against mitochondrial genome and an E value of 10^–30 for searches against bacterial databases. Masking of repetitive elements was undertaken by using the RepeatMasker (http://www.repeatmasker.org) under default parameters. The remaining high-quality sequences were aligned against the original genomic clone by using the BLASTN program, and alignment coordinates and scores were loaded into the MySQL database on a daily basis.

Consensus Assembly

The reads corresponding to validated TFs were assembled into a contig by using the PhredPhrap. The contig sequence was aligned with both EST clusters by using the BLASTN program, and alignment coordinates were used for consensus generation. A Web-based interface was developed to monitor the assembly and access the consensus sequences (http://200.18.51.201/viewconsensus/).

Characterization of Validated Transcripts

Characterization of validated transcripts was pursued by using the UCSC Genome Browser (Kent et al. 2002), which is available at http://genome.ucsc.edu. This allowed determination of sequence overlap between the validated consensus sequences, known genes, and gene predictions. Consensus sequences derived from the validated TFUs were aligned to the July 2003 version of the human genome sequence assembly provided by UCSC using the BLAT search tool. The annotation tracks used for comparison to already known genes were known genes, RefSeq genes, and human mRNAs from the GenBank. A validated transcript was considered a new gene if its alignment coordinates did not match the coordinates of any other sequence available through the known genes, RefSeq genes, or human mRNA annotation tracks. For comparison to gene predictions, the following tracks were used: Fgenesh++, Geneid, and GenScan predictions. The prediction of individual exons instead of the full transcript prediction was considered. A validated exon was considered as predicted if it aligned within the coordinates defined by any of the three gene prediction programs (not necessarily sharing borders) and a new validated transcript was considered not predicted if all exons were not predicted by the computer programs. The consensus sequences corresponding to new validated transcripts were further characterized by BLASTX analysis, and protein domains were determined by using the Pfam and Prosite databases.

Characterization and Validation of Alternatively Splicing Forms

The individual sequences generated during the process of validation of each TFU were aligned to the human genome assembly by using the BLAT search tool, together with the final consensus sequence and representative sequences derived from both EST clusters. Alternatively spliced isoforms were visually identified by using the UCSC browser. To eliminate alignment artifacts caused by sequencing errors and problems in the genome assembly, we have considered as alternatively spliced forms only exons defined by conserved acceptor and donor splicing sites (GT/AG). Primers for validation of predicted alternative splicing isoforms were designed by using Primer3 with default parameters. The presence of alternative isoforms was analyzed by using a cDNA panel composed of 20 different normal and tumor tissues. GAPD amplification was used as a control for integrity and quantification of the RNA used for cDNA synthesis. RT-PCR products obtained in touchdown reactions were analyzed on 1.5% agarose gels.

Complete List of Authors

Coordination Group Ludwig Institute

Fabiana Bettoni,3 Dirce Maria Carraro,3 Lilian C. Pires,3 Raphael B. Parmigiani,3 Elisa N. Ferreira,3 Eloísa de Sá Moreira,3,32 Maria do Rosário D. de O. Latorre,4 Andrew J.G. Simpson,3 and Anamaria A. Camargo3

Coordination Group University of São Paulo Chemistry Institute

Luciana Oliveira Cruz,5 Theri Leica Degaki,5 Fernanda Festa,5 Katlin B. Massirer,5 and Mari C. Sogayar5

Bioinformatics Groups

Fernando Camargo Filho,6 Luiz Paulo Camargo,6 Marco A.V. Cunha,7 Sandro J. De Souza,8 Milton Faria Junior,6 Silvana Giuliatti,6 Leonardo Kopp,9 Paulo S.L. de Oliveira,9 Paulo B. Paiva,10 Anderson A. Pereira,6 Daniel G. Pinheiro,7 Renato D. Puga,6 and Jorge Estefano S. de Souza8

Validation Groups

Dulcineia M. Albuquerque,11 Luís E.C. Andrade,12 Gilson S. Baia,13 Marcelo R.S. Briones,14 Ana M.S. Cavaleiro–Luna,15 Janete M. Cerutti,16 Fernando F. Costa,¹¹ Eugenia Costanzi-Strauss,17 Enilza M. Espreafico,18 Adriana C. Ferrasi,19 Emer S. Ferro,¹³ Maria A.H.Z. Fortes,¹⁵ Joelma R.F. Furchi,20 Daniel Giannella-Neto,¹⁵ Gustavo H. Goldman,21 Maria H.S. Goldman,22 Arthur Gruber,23 Gustavo S. Guimarães,16 Christine Hackel,24 Flavio Henrique-Silva,20 Edna T. Kimura,¹³ Suzana G. Leoni,¹¹ Cláudia Macedo,25 Bettina Malnic,26 Carina V. Manzini B.,26 Suely K.N. Marie,27 Nilce M. Martinez-Rossi,25 Marcelo Menossi,28,29 Elisabete C. Miracca,30 Maria A. Nagai,30 Francisco G. Nobrega,31 Marina P. Nobrega,31 Sueli M. Oba-Shinjo,27 Márika K. Oliveira,18 Guilherme M. Orabona,32 Audrey Y. Otsuka,33 Maria L. Paço-Larson,18 Beatriz M.C. Paixão,⁷ Jose R.C. Pandolfi,34 Maria I.M.C. Pardini,19 Maria R. Passos Bueno,32 Geraldo A.S. Passos,35 Joao B. Pesquero,36 Juliana G. Pessoa,36 Paula Rahal,37 Cláudia A. Rainho,38 Caroline P. Reis,28 Tatiana I. Ricca,¹⁴ Vanderlei Rodrigues,39 Silvia R. Rogatto,38 Camila M. Romano,23 Janaína G. Romeiro,37 Antonio Rossi,39 Renata G. Sá,39 Magaly M. Sales,19 Simone C. Sant'Anna,24 Patrícia L. Santarosa,40 Fernando Segato,25 Wilson A. Silva Jr.,7,25 Ismael D.C.G. Silva,33 Neusa P. Silva,¹² Andrea Soares-Costa,20 Maria F. Sonati,41 Bryan E. Strauss,42 Eloiza H. Tajara,37 Sandro R. Valentini,34 Fabiola E. Villanova,33 Laura S. Ward,40 and Dalila L. Zanette⁷