Poly (A)+ Transcriptome Assessment of ERBB2-Induced Alterations in Breast Cell Lines

We report the first quantitative and qualitative analysis of the poly (A)+ transcriptome of two human mammary cell lines, differentially expressing (human epidermal growth factor receptor) an oncogene over-expressed in approximately 25% of human breast tumors. Full-length cDNA populations from the two cell lines were digested enzymatically, individually tagged according to a customized method for library construction, and simultaneously sequenced by the use of the Titanium 454-Roche-platform. Comprehensive bioinformatics analysis followed by experimental validation confirmed novel genes, splicing variants, single nucleotide polymorphisms, and gene fusions indicated by RNA-seq data from both samples. Moreover, comparative analysis showed enrichment in alternative events, especially in the exon usage category, in ERBB2 over-expressing cells, data indicating regulation of alternative splicing mediated by the oncogene. Alterations in expression levels of genes, such as LOX, ATP5L, GALNT3, and MME revealed by large-scale sequencing were confirmed between cell lines as well as in tumor specimens with different ERBB2 backgrounds. This approach was shown to be suitable for structural, quantitative, and qualitative assessment of complex transcriptomes and revealed new events mediated by ERBB2 overexpression, in addition to potential molecular targets for breast cancer that are driven by this oncogene.


Introduction
Global comparative analysis of transcriptomes is the most effective approach for definition of alterations in gene expression profiles and has led to the identification of key defective elements involved in complex diseases such as cancer. Different aspects of quantitative gene expression have been investigated in breast cancer by microarray-based analysis [1][2][3][4][5][6], with important implications for prognosis [7,8]. At present, the management of breast cancer patients takes into consideration a combination of clinical and histopathological characteristics, together with the measurement of estrogen (ER) and progesterone (PR) hormone receptors and Her2/ERBB2 overexpression/amplification. ERBB2 (human epidermal growth factor receptor) is a member of the tyrosine kinase receptor family, and its amplification has long been considered to play a crucial role in the tumorigenic process [9,10]. ERBB2, overexpressed in 25 to 30% of human breast cancers [11,12], is associated with metastasis [13], and ERBB2-overexpressing cells are self-sufficient with respect to, anchorage-independent growth and efficient in invasion [14].
Patients bearing ERBB2-overexpressing tumors are usually treated with Trastuzumab (HerceptinH), a therapeutic monoclonal antibody against ERBB2. However, a significant fraction (,60%) of patients with metastatic breast tumors does not respond to the treatment [15], highlighting the necessity for continued investigation of ERBB2-mediated modifications in breast cells.
In light of the importance of ERBB2 in breast cancer, HB4a [16] and HB4aC5.2 [17], the parental and the ERBB2 overexpressing cell lines, respectively, have been used to investigate quantitative transcriptional alterations in mammary cells mediated by ERBB2 overexpression [18,19]. However, given the high complexity of the mammalian transcriptome [20,21] the use of more sensitive approaches that enable assessment of not only quantitative but also qualitative aspects of the transcriptome has presently been intensified. In this sense, clear consensus has emerged that next-generation sequencing (NGS), which provides digital-counting of the transcriptome, is more advantageous than other solely quantitative methodologies [22][23][24][25][26].
Here we present the first NGS-based qualitative and quantitative evaluation of the mammary cell transcriptome modulated by ERBB2 over-expression. By combining DpnII-restriction and parallel-tagged sequencing, we performed analysis of the poly (A) + transcriptomes of two human mammary cell lines: HB4a [17] and its ERBB2-overexpressing clone, C5.2 [16]. Whereas no qualitative aspects were correlated with ERBB2 over-expression, significant enrichment of alternative splicing events was shown to be mediated by the overexpression of this oncogene. Additionally, novel ERBB2-driven genes and transcript variants were revealed in these cell lines and were also validated in tumor specimens with high ERBB2 expression. Moreover, novel genes, splicing variants, single nucleotide polymorphisms (SNPs) and gene fusions were detected in sequences from both cell lines, data contributing further to the definition of the human transcriptome.

Results
Whole transcriptome assessment of HB4a and C5.2 cell lines A customized method for simultaneous sequencing of the poly (A) + transcriptome of multiple samples was established. Doublestranded cDNA converted from purified poly A + mRNA from HB4a and C5.2 cell lines was prepared and digested with the frequent cutter DpnII. Adapters containing specific 4nt-barcode were designed and added to DpnII digested cDNAs from both cell lines, and samples were pooled before sequencing ( Figure 1). cDNA samples were sequenced on the Titanium 454-Roche platform generating 802,214 reads, with a total of 160,223,981 high-quality nucleotides (Phred $20) (Figure 2). The mean size of reads was 199.7666.72. Sequence reads were submitted to The Sequence Read Archive (SRA) (accession number SRA012436.2).
Sequences were filtered to exclude mitochondrial DNA and ribosomal RNAs, as well as sequences without adapters, yielding 731,628 reads (91.2%) (Figure 2A). From this, two subsets were separated according to genome alignment parameters that differed in percentage of coverage and identity: a subset of completely aligning reads and another subset of reads with partial alignment to the human genome, respectively containing 651,058 reads (89%) (coverage $70% and identity $96%) and 80,570 reads (11%) (coverage $20% and #80% and identity $99.9%). From the former, 614,434 (94.4%) that present single-hit matches ( Figure 2B) in the genome were used for gene coverage and identification of novel genes, splicing variants and SNPs. From the latter subset, reads that present perfect matches in two genomic regions were used to assess gene-fusion events.

Gene coverage and novel genes
A small fraction of 1,009 (0.16%) reads mapping out of any gene-track position was explored to find novel human genes ( Figure 2C). Most of these sequences (960 reads) were single continuous hits, and only a minority (49 sequences) was derived from multiple-exon hits. Six out of the 49 multi-exon genes exhibited canonical splice sites at their putative introns and were selected for validation by the use of cDNA from these cell lines, and three (50%) new human transcripts were thereby confirmed. The minimal distance between the novel transcribed regions and the known genes in the vicinity was 8 Kb (Figure 3), suggesting that the three confirmed transcripts are not extensions of the known genes.
From the 613,425 reads (98.83%) that could be mapped against at least one gene track ( Figure 2B), 476,337 reads (73.16%) were aligned to RefSeq (Figure 2C) that demonstrated a expression of 38.9% (17,887 entries) of the complete RefSeq transcript repertoire (build 36) in these cell lines. Given that 17,887 fulllength mRNAs correspond to 60,500,115 nt, the base-pair representation was 23.48% (14,208,089 nt covered).
To estimate transcript representation we verified the read distribution throughout the entire full-length mRNAs by calculation of their relative position as described [27]. A slightly higher concentration of sequences was seen in the central portion of the transcripts ( Figure S1A). The 59 end of full transcripts was wellrepresented, irrespective of the original transcript lengths ( Figure  S1B). This fact shows high-quality RNA templates, efficient cDNA synthesis and library construction methodology with no biased representation of short transcripts.

Single nucleotide polymorphism
The 614,434 reads mapping at a single genome position were used for assessment of novel SNPs ( Figure 2B). According to a set of parameters and criteria (see Methods), 3,532 known SNPs and 138 (3.8%) potentially novel SNPs were thereby revealed.
Eighteen putative SNPs were subjected to validation by Sangersequencing of genomic DNA from both cell lines. A high validation rate (89%) was obtained and revealed 16 new SNPs (12 heterozygous and 4 homozygous SNPs) (Table S1 and Figure  S2). Nine of the 16 new SNPs were located in coding regions, and four resulted in non-synonymous amino acid substitutions, all predicted as benign substitutions by PolyPhen [28]. Intriguingly, the different chromatogram peak height of DKK1 heterozygous SNP observed in C5.2 cells was suggestive of an apparent allelic expression imbalance in the ERBB2-overexpressing cell line.

Novel splicing variants
Putative novel splicing variants could be identified from the 597,565 reads that aligned against mRNA databases ( Figure 2C). From them a total of 2,865 were potentially novel alternative splicing (AS) events and were distributed in distinct categories ( Figure 4). For validation, we focused on the exon inclusion set ( Figure 4C), because the feasibility of designing primers in the newly-included exons leads to more reliable results with regard to the expression of specific variants. From the subset of 89 events containing two known flanking exons ( Figure 4C) 20 exoninclusion events were selected for validation. From them, 18 (90%) new bona fide exon inclusion AS-variants ( Figure S3) were confirmed.

Gene fusion
From the partially aligning subset ( Figure 2B), reads that presented perfect match in two distinct genomic regions were considered as potential gene fusion events. After removing 26 false-positive candidates detected by the presence of DpnII restriction sites at the fusion junction ( Figure S4), which indicates chimeric artifacts produced during library construction, 330 reads remained. Next, chimeras within the same gene and between a gene and its corresponding pseudogene were discarded. Finally, reads presenting partial alignment to unique exonic regions were selected resulting in a total of 39 gene fusion candidates, 33 inter and 6 intra-chromosomal events ( Figure 2B). From the 39 gene fusion reads, 32 events presented micro-homology (27 inter-and 5 intra-chromosomal), and only one (inter-chromosomal) presented 2 bases that did not align to any of the regions involved in the fusion, named short non-templated sequences (Table S2).
Ninety five percent of gene fusion events (37 out of 39) were reported by more than 2 reads and 72% (28 out of 39) were reported by both cell lines, suggesting that they are bona-fide gene fusion events ( Figure 5A). However, only 3 out of 14 events randomly selected for validation (11 inter-and 3 intra-chromosomal) were confirmed by qRT-PCR assays using cDNA and genomic DNA from both cells ( Figure 5B; Table S2). The two inter-chromosomal gene fusions (FTH1/chromosome 11 -EIF5A/ chromosome 17; VAMP8/chromosome 2 -SCAF1/chromosome 19) were validated in both cDNA and genomic DNA from HB4a and C5.2 cells, a result that demonstrates their genomic rearrangement origin. Conversely, the intra-chromosomal gene fusion reported (CDH13/chromosome 16 -MLYCD/chromosome 16) was validated only in cDNA from C5.2 cells. The absence of amplification in genomic DNA suggests that this fusion might result from trans-splicing event or from genomic rearrangement, whose fragment size precludes its amplification by standard PCR conditions. To check whether ERBB2 overexpression could augment genome instability that could be reflected by gene fusion events in the transcriptome of both mammary cell lines, we compared the number of gene fusion events between ERBB2-basal (HB4a) and ERBB2-high expression (C5.2) cell lines normalized by the total number of reads of each respective cell. The normalized number obtained from each cell line was highly similar: 34.2 events in Hb4a cells and 38 events in C5.2 cells (p = 0.24). From the three validated gene fusions, VAMP8/SCAF1 and CDH13/MLYCD were reported in RNAseq from C5.2 cells exclusively (Table S2). However VAMP8/SCAF1 was detected in both cDNA and genomic DNA form HB4a and C5.2 cells, whereas CDH13/ MLYCD confirmed to be exclusively expressed in cDNA of C5.2 cells. Based on our results, no conclusive evidence was found towards the influence of ERBB2 overexpression on genomic instability.
To explore the influence of ERBB2 overexpression on the regulation of alternative splicing, we compared the number of AS events of each category in each cell line (Table S3), normalized by the total number of reads obtained. An enrichment of alternative splicing events was observed in the C5.2 cells represented by the categories of exon skipping (p = 1.35E-6), exon inclusion (p = 0.005), and alternative acceptor/donor splice sites (p = 2.4E-7).
To confirm this enrichment, the expression of eight events out of 18 that were reported only by reads obtained with C5.2 specific tag, indicative of a specifically or highly expressed splicing variant in C5.2 cells, was evaluated by quantitative RT-PCR in both cell Completely Aligning to Human Genome corresponds to reads that aligned to the genome sequence using the criteria of coverage $70% and identity $96%. Sequences aligning to more than one genome region following these criteria were discarded. Single-hit high coverage genome alignment sequences were used for discovery of novel SNPs. Partially Aligning to Human Genome corresponds to reads that aligned to genome sequence using the criteria of coverage $20% and #80% and identity $99.9%. These reads were used for discovery of gene fusion events. (C) Transcript databases alignment. For transcriptome analysis, reads from completely aligning sub-set were further aligned to known gene databases for the discovery of novel splicing variants. Additionally, reads were aligned to RefSeq databases for obtaining the number of transcripts (mRNAs) and genes identified. doi:10.1371/journal.pone.0021022.g002 lines. Although none of these AS events was shown to be specific for C5.2 cells, since amplification was detected in both cell lines, six out of eight (75%) confirmed a higher expression in C5.2 cells (fold .2), a result indicating that the enrichment resulted from an increase in the expression level of the AS transcripts possibly influenced by ERBB2 overexpression. Further evidence that this oncogene influenced specific AS variants of a gene is that for five out of six events, only the splicing variants evaluated seem to be influenced by ERBB2 overexpression ( Figure S3 and Table S4), i.e., a comparison between the number of entire gene-related sequences obtained from each cell line showed no difference in expression levels (or even higher expression in HB4a cells).
Five out of six splicing variants overexpressed in C5.2 are positioned within the coding sequence (CDS). For three variants the exon insertion resulted in premature stop codons (KIAA1033, CSRP2BP and PRCC) that might produce truncated proteins. The exon insertion in the variant of the CLTC gene generated an inframe insertion of 7 amino acids, with no alteration in protein domains. For the NR2C1 gene, the inserted exon resulted in a smaller protein isoform (177 amino acids are removed) that resulted in an alteration in the N-terminal portion with the loss of two domains, (the vitamin D and zinc-finger nuclear hormone receptor) and gain of the retinoid X receptor domain. If the putative non-functional protein plays a role in ERBB2-driven breast tumor remains to be addressed.
Gene expression analysis was carried out by counting the number of reads representing each gene, independently of their relative position within the full-length mRNA. Indeed, as expected, we have found more tags representing the ERBB2 gene in C5.2 than in HB4 cells, with a 15-fold expression difference, a result indicating that RNA-Seq can provide informative and confident quantitative results.
To check if the transcript level is preserved by our RNA-seq we compared our RNAseq data to Unigene cluster-size, a measure of transcriptional abundance. We have selected Unigene clusters and grouped them according to the total number of ESTs for each class of abundance (top or bottom 25%, 10% and 5%). Then we calculated the average number of sequences obtained in RNAseq experiments for the respective set of genes, for each of these abundance-classes. RNAseq reads showed a positive correlation, where genes with larger Unigene-cluster sizes were sequenced much more often than genes with lower Unigene-cluster sizes ( Figure 6A). Thus, we can conclude that transcriptional levels are indeed preserved by our RNAseq data.
A total of 436 potentially differentially expressed genes, 192 upregulated and 244 down-regulated, was identified in C5.2 cells ( Figure 6C). Eighty-eight of these were evaluated by qRT-PCR, and the differential expression of 46 genes (52.3%) was validated (Table S5).
To evaluate the correlation between gene expression quantification based on RNA-seq and qRT-PCR, a comparison between numbers of reads and Cycle threshold (CT) values was performed for 88 genes using TaqMan assays and p,0.01 was considered as statistically significant. This comparison showed a negative significant correlation when all genes (p,0.0001) or only the validated ones (p,0.0001) were analyzed ( Figure 6B). Even for the non-validated genes a marginal significance was seen (p = 0,0263). This result reinforces the effectiveness of this methodology for parallel sequencing of two different and barcoded cDNA populations and the feasibility of our approach for the determination of gene expression profiles.
To investigate the biological properties of genes modulated by ERBB2-mediated expression, we classified the 46 genes with confirmed differential expression according to Biological Processes in Gene Ontology (GO) and also within KEGG database pathways. Using FunNet [29] we identified 17 enriched GO and 8 KEGG categories in both cell lines ( Figure S5).
To determine whether these quantitative gene expression profiles could also be associated with human tumors over-expressing ERBB2, we investigated the expression levels of the 46 validated genes in 14 human breast ductal carcinoma samples containing high (7 samples) or basal levels (7 samples) of ERBB2 expression ( Figure 6D). Despite tumor and patient heterogeneity, as well as the gap between cell line models and clinical samples, 4 genes (8.6%) were also modulated in breast tumor samples with distinct ERBB2 backgrounds: ATP5L was increased in the ERBB2-positive samples, whereas LOX (ENSG00000113083), GALNT3 (ENSG00000115339), and MME (ENSG00000196549) showed reduced expression when ERBB2 was elevated ( Figure 6E).
One of the most important signaling pathways driven by ERBB2 involves PI3K/AKT that activates the protein kinase mammalian target of rapamycin (mTOR), an important regulator of mRNA translation that controls cell proliferation [30]. To evaluate whether the genes modulated by ERBB2-mediated expression identified in this study were under mTOR transcriptional control, we treated both cell lines with rapamycin and investigated their expression after treatment. From the 46 validated genes, 19 (41.3%) showed reduction or inversion in relative fold-difference between C5.2/HB4a cells (Table S6). Some of these genes have already been reported to be modulated rapamycin treatment [31][32][33][34], strengthening the possibility that the additional genes, not previously reported, are also mTOR downstream regulated.

Discussion
The approach presented here has provided data for quantitative and architectural aspects of two mammary cell-line poly (A) + transcriptomes, and has shown the potential to fully represent transcripts. Restriction enzyme digestion revealed advantage over physical methods for cDNA fragmentation, such as prompt  High-throughput transcriptome sequencing has been used to identify genomic rearrangements resulting in gene fusion events [23,35,36,37] and superior sensitivity was achieved when pairedend sequencing was applied [36,37,38]. Single-end sequencing of long or short reads has led to low validation rates, which seem to be increased when single-end long and short sequencing are used in combination [35]. The low validation rate obtained here (21%) reinforces the difficulty of confirming these events using PCRbased approaches. Nonetheless, despite using single-end long reads sequencing we identified three bona fide gene fusions, reported here for the first time to the best of our knowledge.
One of the validated gene fusions reported 10 nt of microhomology sequence, also detected in 82% of the gene fusion candidates, probably resulting from a replication mechanism known as Fork Stalling and Template Swichting Model (FoSTeS). These replication disorders arises due to nucleotide similarities between DNA strands [39], and have been detected in breast cancer samples [40].
For exploring SNPs, we used stringent bioinformatics and manual inspection that resulted in a high rate of confirmation (89%), including validated SNPs with apparent allelic dosage imbalance in C5.2 cells. Whether ERBB2 overexpression can mediate allelic dosage imbalance during transcriptional process remains to be addressed.
The detection of alternative splicing by our method is enhanced by the longer fragments produced by the 454-platform, compared to other next-generation sequencing technologies. Here we showed a 90% validation rate of the exon inclusion splicing variant class. An extrapolation of this value over the 1,704 novel AS events in multi-exon splicing variants, with conserved splice sites, identified by RNAseq would result in 1,533 bona-fide AS events. Our approach therefore demonstrates a high capacity for identification of novel splicing variants and, consequently, for the definition of the mammary transcriptome.
Amplification of the ERBB2 oncogene is considered an important tumor driver [41], and has been reported in approximately 25% of breast cancers [11]. The quantitative transcriptional aspect of overexpression of the oncogene has been previously assessed by 39 end sequence methodology [18,19]. However, only whole transcriptome sequencing enables the assessment of some relevant structural aspects. The influence of ERBB2 was observed in quantitative aspects of breast cell line transcriptomes, not only on gene expression but also on specific splicing variants. Enrichment of exon skipping/inclusion and alternative splice site selection by ERBB2 overexpression in C5.2 cells, observed in the current study, indicated a potential influence of the oncogene in the regulation of the splicing process. In this sense, evidence has already been presented by us concerning expression level regulation of specific AS variants mediated by ERBB2 [42]. Additionally, it has been suggested by others [43] that activation of signaling pathways such as Ras/MAPK and PI3K/AKT, which are controlled in part by ERBB2 signaling, might influence the alternative splicing balance of cells, by phosphorylation and activation of specific splicing factors.
The intrinsic molecular heterogeneity found between distinct human tumor samples as well as within a single breast tumor sample has been reported by many laboratories [7,44]. These differences appear to be strongly dependent upon microenvironmental factors [5,45]. Despite differences in molecular character- istics between cells in vivo and in vitro, our approach allowed us to identify 4 genes, the expression of which is likely mediated by ERBB2. We highlight LOX downregulation in C5.2 cells as well as in tumor samples overexpressing ERBB2. Furthermore, our data also revealed a higher level of LOX expression in C5.2 cells after exposure to rapamycin (4-fold change), data indicating that LOX is potentially regulated by the ERBB2/mTOR pathway. LOX encodes an extracellular copper-requiring enzyme that initiates collagen and elastin crosslinking and enhances tumor cell invasion and metastasis [46]. Conversely, the 18-kDa LOX propeptide was found to be an effective inhibitor of the more invasive phenotype of breast cancer cells driven by ERBB2 and has been suggested to improve treatment in this subtype of breast cancer [47].
Altogether, the results presented here demonstrate that our approach is suitable for whole transcriptome interrogation, with single or multiple samples in parallel sequencing by the 454-ROCHE platform, from which an accurate quantitative and qualitative portrait of complex transcriptomes can be generated.

Cell lines and tumor samples: treatment and RNA purification
Two human breast cell lines, HB4a [16] and C5.2 [17], were provided by Dr. Michael O'Hare -LICR, NY, and were grown as described [16]. The high expression of ERBB2 transcript in C5.2 compared to HB4a was confirmed by quantitative RT-PCR (data not shown). For rapamycin treatment, both cell lines were plated in 25 cm 2 flasks and maintained at 40-50% confluence. Cells were treated with vehicle (0.01% absolute ethanol) or 20 nM rapamycin for 24 hours.
Breast tumor samples containing signed informed consents were retrieved from the A. C. Camargo Hospital Biobank. This study is in accordance with the Ethics Committee in Human Being Research from Fundação Antonio Prudente -Hospital of Cancer -A.C. Camargo/S.P and has been approved under number 952/ 07. Fresh-frozen tumor blocks were cut, fixed, and stained with hematoxylin and eosin (H&E) and reviewed by a pathologist. Total RNA was extracted with Trizol and treated with DNaseI. Samples were classified as ERBB2-high or basal expression according to protein and mRNA levels. Protein levels were evaluated by immunohistochemistry (signals 2+ or 3+ indicating high ERBB2 expression, and signals 0 or 1+ indicating basal expression). ERBB2 transcript was evaluated by quantitative RT-PCR. Only samples with concordant results at both protein and transcript expression level were included in the study.
For gene fusion events, reads from partly aligning subset, defined as reads with coverage $20% and #80% and identity $99.9% (pslReps parameters: minCover = 0.70, minAli = 0.96, nearTop = 0.005) were used. Next each region of the read could have only one single match in the human genome sequence and the two regions of the read should align to different genes. Chimeric events including one gene and its corresponding pseudogene were also discarded. Additionally, only reads mapping to exonic regions were selected.
To determine the sequence depth coverage of the fusion gene candidates, we performed alignments between the read reporting the fusion event and all reads generated by 454-sequencing using less stringent criteria.
The KnownGene [50] annotation track coordinates from the University of California Santa Cruz genome database (UCSC) was used as a reference for mapping of the reads in relation to annotated transcripts and exons. Gaps spanning more than 50 bp and having conserved splice sites were considered true intron sequences, and were used to identify putative splicing events not annotated in the KnownGene annotation track. Reads not mapped to any gene-track were used for the identification of putative novel genes. These candidates were analyzed for the presence of conserved splice sites.
For single nucleotide polymorphisms (SNPs) Blast-like BLAT alignment outputs were parsed by the use of a Perl script. A set of parameters such as base coverage, proximity to exon-intron boundaries, proximity to alignment ends, number of different sequenced bases for a specific genome location, ratio between divergent base and reference base, and presence in both libraries, was used to select putative SNPs, that were aligned against dbSNP (build 129) [51].
For all comparative analyses of both RNA-seq from HB4 and C5.2 cells, DNA barcoding identification was valid when barcoding was flanked by a 59-adapter sequence and a 39restriction enzyme site. To analyze the differential gene expression profile between HB4a and C5.2 cell lines, we first aligned confident reads against the RefSeq database with the MegaBlast tool (E-value#1610215, identity $96% and coverage $90%). Reads with significant alignments to different genes were excluded. The overall read count per gene was scaled to reads per million (RPM), and differential expression was calculated as the ratio of C5.2RPM/HB4aRPM. We used the SAGEbetaBin statistical approach to calculate the Bayes Error Rate (SAGEbetaBin available at http://bioinfo.lbhc.hcancer.org.br/sage/betabin// en/index.php) [52].
Association between RNA-Seq read count and qRT_PCR Ctvalues was assessed with Spearman correlation coefficient and p,0.01 was considered as statistically significant.

Validation of SNPs
Potentially novel SNPs exhibiting more than one mismatch in a 50 bp window or mapped to homopolymers and repetitive sequences were discarded. Primers were designed by Primer3 and were used for PCR amplification with DNA from HB4 and C5.2 cell lines. The amplicons were evaluated on 3% agarose gels and were sequenced in an ABI 3130xL (Applied Biosystems, Foster City, California).

Validation of gene fusions
For confirmation of gene fusion, two strategies were used. First, primers aligning at the extremities of the reads reporting the event were designed for RT-PCR and PCR amplification using both cDNA and genomic DNA, respectively, from both cell lines (Table  S7). Amplification products were analyzed on 8% polyacrylamide gels and were sequenced on the ABI3130 instrument (Applied Biosystems, Foster City, California).
Further, for detection of fusion events in cDNA two probes (left and right) were designed for each putative gene fusion event. The left probe was complementary to one of the genes involved in the event, exactly at the limit of the fusion; the right probe, directly adjacent to the left probe, was complementary to the other gene involved in the putative event. In addition, the left probe contained at its 59 end a recognition sequence of the forward PCR primer (59GGGTAGGCTAAGGGTAGGA39). The right probe was phosphorylated at its 59 end and contained a recognition sequence of the reverse PCR primer (59TCTA-GATTGGATCTTGCTGGCAC39) at its 39 end (Table S7). The probes were hybridized to pre-heated double-stranded cDNA and genomic DNA from HB4a and C5.2 cells at 54uC for 12 hours. The two probes hybridized to their target sequence were subsequently ligated by Ligase-65 (MC Holland, Amsterdam, Netherlands) to form a single probe that was PCR-amplified. As a negative control, hybridization in the absence of any template was performed for all probes, and the reaction was submitted to PCR. Amplification products were analyzed on 8% polyacrylamide gels and were sequenced on the ABI3130 instrument (Applied Biosystems, Foster City, California).

Validation of splicing variants
Primers were designed at the respective novel exon and at one adjacent exon (Table S8). cDNAs converted from 40 ng of DNase Itreated total RNA from HB4a and C5.2 cells was used in each reaction. PCR was performed in a total volume of 20 ml, 1 X reaction buffer, 2.5 mM MgCl 2 , 0.2 mM dNTP, 10 pmoles of each primer, and 1 unit Platinum Taq DNA polymerase (Life Technologies, Foster City, California) in 40 cycles at 95uC for 30 sec, 60uC for 30 sec, and 72uC for 30 sec, followed by a final extension at 72uC for 7 min. Amplification products were visualized on 8% polyacrylamide gels and were sequenced on an ABI3130 instrument (Applied Biosystems Foster City, California). For quantitative analyses, PCR amplification with the same pair of primers was performed in 20 ml 1 X SYBR Green PCR MasterMix (Applied Biosystems Foster City, California), containing 2-8 pmoles of each primer and cDNA converted from 100 ng of total RNA.

Validation of novel genes
Primers for the validation of 6 putative novel genes were designed at two distinct exons with the 454-read as a reference sequence (Table S9). cDNA converted from 40 ng of DNase Itreated total RNA from HB4a and C5.2 cells was used in each reaction. PCR reactions were performed in 20 ml containing 1 X buffer, 2.5 mM MgCl 2 , 0.2 mM dNTP, 10 pmoles of each primer, and 1 unit Taq DNA polymerase incubated at 95uC for  30 sec, 60uC for 30 sec, and 72uC for 30 sec for 40 cycles, followed by a final extension at 72uC for 7 min. Amplification products were visualized on 8% polyacrylamide gels and were sequenced on an ABI3130 instrument (Applied Biosystems Foster City, California).
Validation of differential gene expression cDNA converted from 400 ng of total RNA was used as a template for the evaluation of 96 distinct transcripts (target genes and endogenous controls) in duplicate. Expression levels of selected genes were verified by quantitative RT-PCR with customized TaqMan low-density arrays (Applied Biosystems Foster City, California) in an ABI7900 instrument. A total of 91 target genes (75 and 16 up-regulated genes in C5.2 and HB4a, respectively) was randomly selected; GUSB was selected, from the 5 endogenous genes tested, as a reference gene. Differential expression levels that exhibited a fold-change.2, were considered significant as determined by the 2 2DDCt method. The list of selected genes is shown in Table S5.

Gene Ontology (GO) and KEGG pathways annotation
FunNet tools were used for computation of the enriched GO and KEGG categories [29]. Significant themes were calculated for up-and down-regulated genes, with the 11,366 represented genes as the reference set. A decorrelated annotation procedure was performed by application of the Fisher exact test using corrected pvalues (p-value,0,01) and false discovery rate 5%. Table S5 Validation of differential gene expression modulated by ERBB2. The mRNA seq data is given as a fold-change between C5.2 and HB4a cell lines. When no reads were identified in the RNA-seq from one of the cell lines, we calculated fold-change by replacing ''0'' by ''1''. Positive and negative values correspond to higher expression in C5.2 and HB4a, respectively. The qRT_PCR results are given as fold-change obtained by 2 2DDCT . In grey: genes validated in the qRT-PCR by the criterion for differentially expressed genes as Fold-change.2. (DOC)

Supporting Information
Table S6 Effects of rapamycin treatment on genes influenced by ERBB2-mediated expression. The results from quantitative RT-PCR on rapamycin-treated cell lines for the 46 validated genes are given as fold-change between C5.2 and HB4a cell lines. Positive and negative values correspond to higher expression in C5.2 and HB4a cell lines, respectively. The fourth columns show the qRT_PCR results obtained from the C5.2 and HB4a cell lines and from the cells lines after rapamycin treatment. Response to rapamycin was considered when a decrease or inversion of fold-change between C5.2 and HB4a cell lines compared to non-treated cell lines was observed. Yes and No represent response or no response to rapamycin, respectively. (nd) Cycle threshold not determined.