Investigating transcriptional characteristics of breast tumor patient samples and normal tissue

A custom expression microarray (GEO accession number GPL13648) was used to analyze the expression patterns of protein-coding and non-coding transcripts in total RNA from 26 well characterized breast tumor samples [44] and 5 normal breast tissue samples from breast reduction operations (Table S1). Tumor tissue samples were selected to distribute equally between the five well-established mRNA subtypes – Luminal A, Luminal B, ERBB2, Basal-like, and Normal-like - based on the PAM50 molecular classifier [2], [45]. We are aware that the samples used for our analysis include samples with heterogeneous tissue composition. Nonetheless, we observed a widespread downregulation of tumor suppressors in breast cancer tissue samples versus normal samples and an upregulation of oncogenic RNAs in tumor, both at the coding and non-coding level (Tables S2 and S3). Moreover, we did not detect any enrichment for adipocyte-specific pathways (KEGG pathway identifiers 00061, 00062, 00071, 00532, 00533, 00534, 01040, 04975) suggesting that differential expression of novel or functionally unannotated non-coding transcripts relates mainly to the transition to tumor. We used Fisher's exact test to assess whether the observed overlap of significantly regulated transcripts with genomic annotation would have been detected by using randomly chosen probes from the array. Odds ratios were computed between the relative overlap of significantly differentially expressed probes (DE-probes) and the annotations over the relative overlap of the annotations and all probes contained on the microarray. We report the observed odds ratios, their 95% confidence intervals and p-values, accordingly.

Differentially expressed transcripts that mapped to intergenic or intronic space were considered as bona fide non-coding (subsequently referred as non-coding), if they did not exhibit any evidence for encoding functional open reading frames, as predicted by RNAcode () [46], nor any amino-acid sequence similarity to known proteins as annotated in the RefSeq database from March 7, 2012 () [47]. Additionally, we obtained a separate set of non-coding transcripts antisense to exons of known protein-coding genes; however, without any bioinformatic evidence for a functional open reading frame on the reading strand of the probe.

Non-coding transcription changed drastically between normal and tumor samples, independent of copy number changes

Our analysis revealed 20,605 probes corresponding to 19,245 unique loci (hg19) significantly differentially expressed (DE) between normal and tumor samples (), reflecting major differences in the transcriptional landscape of breast cancer in contrast to normal tissue (Figure 1A, Figure S4). Of these regulated genomic loci mapped 7075 (36%) to exons of known protein-coding genes (Gencode v12), 3882 (20%) were located in intergenic space, and 6047 (31%) in introns of protein-coding genes.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 1. Differential expression analysis.

The expression patterns of mRNA-probes and non-coding probes of 26 breast tumors and 5 normal breast tissues were investigated using the custom microarray. (A.) Fraction of unique genomic loci significantly differentially expressed () between normal and tumor samples located completely in exons of protein-coding genes (Gencode v12), in exons of known lncRNAs (lincRNAs, Gencode v12 lncRNAs, lncRNAs as annotated in lncRNAdb [51], and lncRNAs contained in chromatin [27]), in exons of transcripts of uncertain coding potential (TUCPs [23]), in exons of short RNAs (UCSC sno/miRNA track), in genomic loci with conserved secondary structure motifs (Evofold [59], RNAz [58], [97] and SISSIz [53]), in antisense-direction to known exons (Gencode v12), or in novel genomic regions. (B.) Fraction of unique genomic loci significantly differentially expressed () between Basal-like and Luminal tumors and located in genomic annotations as described for panel A. Numbers beside bars denote absolute number of unique DE-loci.

https://doi.org/10.1371/journal.pone.0106076.g001

DE-probes downregulated in tumor were highest enriched, according to Fisher's exact test, in exons of protein-coding genes (Figure 2A, and Table S4), many of these are known to act as tumor suppressors [48] (odds ratio 5.2 fold, , data not shown). DE-probes upregulated in tumor appeared to represent primarily novel transcripts (Figure 1A), which was also reflected by a comparably lower enrichment in exons of protein-coding genes (Figure 2A, Table S4). Many tumor-upregulated DE-probes were found antisense to exons of protein-coding genes (Figure 1A, Figure 2A), suggesting that an abundant repertoire of antisense transcripts is active in breast cancer tissue but not in normal breast tissue.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 2. DE-probe overlap with genomic annotation (Normal versus Tumor).

A.–D.: Number of DE-probes significantly differentially expressed between normal and tumor samples () and mapping to different genomic annotations. Log2 transformed odds ratios and their 95% confidence interval for the respective annotation dataset are shown. Odds ratios of observed versus expected probe overlaps were calculated and tested by Fisher's exact test for significant enrichment or depletion, with *** indicating , ** , and * , respectively. Results are shown (A.) for DE-probes located in annotated protein coding genes versus intergenic space based on Gencode release v12, (B.–D.) for intergenic or intronic non-coding DE-probes either located in several classes of known and predicted ncRNAs (B.), in non-coding transcripts regulated during cell cycle (CC), upon TP53 or Stat3 induction [43] (C.), or in regulatory sites (D.). (E.) Fraction of unique non-coding DE-loci in exons of known short and long ncRNAs, in genomic sites with conserved secondary structures, in antisense-direction to known non-coding exons (Gencode v12), or in novel sites. Numbers denote absolute number of DE-loci located in novel sites. For detailed output of Fisher's exact tests see Table S4, and Table S7 for detailed description of annotation datasets.

https://doi.org/10.1371/journal.pone.0106076.g002

An impressive number of 8591 (82.9%) DE-probes that mapped to 8282 unique genomic loci in intergenic or intronic space were non-coding, i.e. without any bioinformatic evidence for functional open reading frames or sequence similarity to known proteins. Non-coding DE-probes located antisense to exons of known protein-coding genes mapped to 1365 unique genomic loci.

It ought to be expected that at least some of the observed expression changes of non-coding loci might be driven by alterations at the DNA level, since it has been shown that 12% of the total variation of mRNA expression in breast cancer may be explained by copy number variation [49], [50]. Consequently, we evaluated the influence of DNA copy number variation on in-cis lncRNA expression changes. Correlation analysis of the 26 tumor samples utilizing data from 109k Illumina SNP array did not detect any in-cis correlation between copy number and lncRNA expression (absolute Spearman's correlation coefficients ) (Figure S1B). Following Pollack et al. 2002, we applied a linear regression model to estimate that the expression variation of 6% of the non-coding probes was explained by copy number variation [50]. These results indicate that the contribution of copy number changes to the observed variation in the lncRNA expression is marginal and that the differential expression of lncRNAs in breast cancer has more complex reasons than a trivial consequence of genomic aberrations.

Non-coding transcription was altered within tumor samples

To investigate the biological variation of lncRNAs within tumor samples independent of known mRNA subtypes, we applied unsupervised hierarchical clustering. Uncertainty of derived clusters was assessed by random sampling with replacement (bootstrapping with 10,000 iterations). Hierarchical clustering based on probes mapping to protein-coding exons largely reproduced the known mRNA subtypes and reflected TP53 status, while hierarchical clustering based on non-coding probes revealed a different pattern (Figure S3). However, due to the small number of patient samples and the associated limited discriminative power of unsupervised clustering methods, this discrepancy should be interpreted with caution.

An F-test was employed to assess, if the mean expression of a probe is equal for all five mRNA-based subtypes in our sample set. We identified 3175 probes that were significantly differentially expressed between any subtype (). We chose a less stringent false discovery rate to account for smaller sample sizes in each group, 382 of these probes mapped to non-coding regions of the human genome with a distinct expression pattern for the Basal-like tumors (Figure S2).

In breast cancer, the most extreme diverging mRNA subgroups are Luminal A and B versus the Basal-like subtype, different in hormon status, prognosis, and survival rate. Our analysis focused on the comparison of these outermost subgroups and we identified 3025 unique genomic regions significantly differentially expressed (), of which 682 were non-coding. The majority (60%) of unique loci that were upregulated in Basal-like tumors corresponded to exons of protein-coding genes (Figure 1B). Of the remaining, 324 (15%) loci were identified in antisense direction of known exons (275 to exons of known protein-coding genes), 134 (6%) corresponded to known lncRNAs or predicted lncRNAs with conserved secondary structures, and 156 (7%) loci were novel. In contrast, only 45% (416) of the unique genomic loci upregulated in the Luminal A and B tumors mapped to coding genes but approximately one third (249) mapped to novel loci. Differential expression of antisense transcripts was higher than expected from the composition of the custom microarray (Figure 3A).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 3. DE-probe overlap with genomic annotation (Basal-like versus Luminal A and B tumors).

A.–D.: Number of DE-probes significantly differentially expressed between Basal-like and Luminal A and B tumors () and mapping to different genomic annotations. Log2 transformed odds ratios and their 95% confidence interval for the respective annotation dataset are shown. Odds ratios of observed versus expected probe overlaps were calculated and tested by Fisher's exact test for significant enrichment or depletion, with *** indicating , ** , and * , respectively. Missing error bars denote no DE-probes overlapped with according annotation. Results are shown (A.) for DE-probes located in annotated protein coding genes versus intergenic space based on Gencode release v12, (B.–D.) for intergenic or intronic non-coding DE-probes either located in several classes of known and predicted ncRNAs (B.), in non-coding transcripts regulated during cell cycle (CC), upon TP53 or Stat3 induction [43] (C.), or in regulatory sites (D.). (E.) Fraction of unique non-coding DE-loci in exons of known short and long ncRNAs, in genomic sites with conserved secondary structures, in antisense-direction to known non-coding exons (Gencode v12), or in novel sites. Numbers denote absolute number of DE-loci located in novel sites. For detailed output of Fisher's exact tests see Table S4, and Table S7 for detailed description of annotation datasets.

https://doi.org/10.1371/journal.pone.0106076.g003

Differentially expressed non-coding transcripts were mostly novel

We observed that up to 80% of non-coding DE-loci, either with significant expression changes between normal and tumor or Basal-like and Luminal A and B samples, were novel. They were neither antisense to known exons, nor located in known or predicted short or long ncRNAs (Figures 2E and 3E). Some of these novel regions corresponded to non-coding transcripts, previously identified to be regulated by cancer-related pathways, like mitotic cell cycle and TP53 mediated apoptosis [43]. We observed an overrepresentation of those within the set of non-coding transcripts downregulated in breast cancer tissue (Figure 2C) or upregulated in Basal-like tumor samples (Figure 3C).

Apart from transcriptional changes in novel sites, we found also significant regulation of known lncRNAs. Considering all non-coding DE-probes downregulated in tumor samples, we detected an enrichment for lincRNAs [23], lncRNAs as annotated in Gencode v12, chromatin-associated lncRNAs (CARs [27]), manually curated lncRNAs (lncRNAdb [51]), transcripts originating from introns of protein-coding genes [52], transcripts of uncertain coding potential (TUCPs [23]), and small RNAs (Figure 2B, Table S4). Non-coding DE-probes upregulated in tumor displayed significant enrichments for lncRNAs as annotated in Gencode v12, for small RNAs, and for loci of conserved secondary structures (SISSIz [53]).

Within the tumor samples, DE-probes upregulated in Basal-like tumors showed significant enrichment for known lncRNAs (Gencode v12, lncRNAdb [51]), for non-coding RNAs transcribed from introns [52], and for small RNAs (Figure 3B and Table S4).

Chromatin-associated lncRNAs were mostly downregulated in tumor samples

One of the highest scored enrichments of non-coding DE-probes, which were significantly differentially expressed between normal and tumor tissue, was observed for lncRNAs previously detected to be contained in chromatin of human fibroblast cells [27]. A total of 88 chromatin-associated lncRNAs (CARs) were represented by at least one probe on the custom microarray, of which 64 showed significant changes in expression between normal and tumor samples (). 43 CARs displayed consistent downregulation in breast cancer, while only 17 were found to be upregulated in tumor, and 4 CARs contained probes with inconsistent expression changes (Table S5). A total of 9 CARs overlapped with annotated lncRNAs, of which NEAT1, MALAT1, and MEG3 are known to be regulated by cancer pathways [19], [42], [54], [55]. Further, 9 CARs were found in introns of protein-coding genes (e.g. in introns of ARID5B, CALD1, EXT1), and 39 spanned introns and exons of a protein-coding gene (e.g. FOS, HNRNPH1). Expression levels of CARs appeared to be constant among Basal-like versus Luminal A and B tumor samples (Figure 3B and Table S4). Hence, CARs might be regulated by pathways being responsible for the onset and progression of primary tumors; however, not by pathways controlling the genesis of a particular subtype. The expression changes detected by the custom microarray of three randomly selected CARs were successfully validated by RT-qPCR (Figure S5).

Differentially expressed non-coding transcripts showed low sequence conservation, but partly secondary structure conservation

Previous studies have indicated that the primary sequence conservation of lncRNAs is in general lower than for protein-coding genes [22], [23], [56]; however, a substantial fraction displays evidence for purifying selection on RNA structures in mammals [57]. Hence, we investigated, if non-coding DE-probes regulated in breast cancer are conserved on sequence and/or on secondary structure level.

Non-coding DE-probes showed less conserved sequence composition than DE-probes that mapped to protein-coding genes; however, they were more conserved than neutrally evolving sequences (Figure S6A). DE-probes significantly differentially expressed in the comparison of Basal-like versus Luminal A and B tumor samples displayed less sequence conservation than DE-probes significantly differentially expressed between normal and tumor tissue samples.

Assessing the structural conservation of non-coding DE-probes, we found these enriched for genomic loci with conserved structure motifs as predicted by SISSIz [53]; however, not for motifs predicted by RNAz [58] and Evofold [59] (Figures 2B, 3B, and Table S4). The three computational methods require different degrees of sequence conservation in the alignment, which may explain the observed imbalance. SISSIz predicts secondary structure motifs at loci with primarily low sequence conservation, while RNAz and Evofold reach optimal classification rates only for alignments with moderate or high sequence conservation, respectively [57], [60]. This result suggests an involvement of ncRNAs with structural motifs in breast carcinogenesis; however, only for those molecules with comparable low sequence conservation. Further, our studies revealed that DE-probes were in general less conserved in their sequence than probes with detectable secondary structure motifs (Figure S6B). Hence, many DE-probes corresponded to genomic loci for which reliable classification, whether the locus contains a secondary structure motif or not, is not possible due to low sequence conservation. Also, the detection of structured differentially expressed ncRNAs by microarrays is hampered by a lower affinity of those to array probes. This was reflected by lower signal intensities of probes, located in genomic loci with conserved secondary structure motifs compared to the remaining probes (Figure S6C).

Differentially expressed non-coding transcripts were enriched in regulatory sites

The genomic location and frequency of non-coding DE-probes was assessed investigating promoters, enhancer regions, transcription factor binding sites, or transcriptionally active or in-active regions (see Table S7 for detailed information about annotations). Non-coding DE-probes, regulated between normal and tumor tissue samples, were enriched in promoter sites as defined by CpG islands and/or H3K4 trimethylation sites, in chromatin marks that are characteristic for enhancers (H3K4 monomethylation and H3K27 acetylation), and in genomic regions actively transcribed (Pol II binding sites, H3K36 trimethylation sites) (Figure 2D, Table S4).

The discrepancy between high (low) odds ratios for DE-probes upregulated (downregulated) in tumor samples and transcriptionally active sites (H3K36 trimethylated) - as opposed to DE-probes downregulated (upregulated) in tumor samples and transcriptionally inactive sites (H3K27 trimethylated) - might be a consequence of the unequal composition of ENCODE ChIP-Seq data, which is mainly derived from normal cell lines (Figure 2D, Table S4).

A less pronounced, but significant, enrichment was observed for transcription factor binding sites [7]. DNase-I hypersensitivity sites, characterized by regions where the chromatin is open in such a way that transcription factor binding is in general possible, were significantly enriched (Figure 2D).

Within tumor samples, non-coding DE-probes, which were upregulated in Basal-like tumors compared to Luminal A and B tumor samples, were significantly enriched in promoters (H3K4 monomethylation), enhancers (H3K27 acetylation), and in regions known to be actively transcribed (Pol II binding sites, H3K36 trimethylation, Figure 3D). Non-coding DE-probes downregulated in Basal-like tumor samples were significantly enriched in promoters (CpG islands and H3K4 trimethylation), enhancers (H3K27 acetylation), transcription factor binding sites, Pol II binding sites, and in DNase-I hypersensitivity sites.

In summary, we observed enrichment of differentially expressed lncRNAs in regulatory and epigenetically modified sites, which might interfere with expression of adjacent protein coding RNAs.

Differentially expressed non-coding transcripts were located in close proximity to protein-coding genes

Cabili et al. 2011 has suggested that intergenic lncRNAs are preferentially located in proximal regions of protein-coding genes [23]. To validate this result for lncRNAs regulated in cancer, we assessed the distance of all non-coding intergenic probes to their nearest protein-coding gene, independent of reading strand. In a comparative analysis of normal versus tumor tissue samples we found more intergenic non-coding DE-probes in close genomic proximity to protein-coding genes than expected from randomly selected intergenic regions preserving probe length of 60 bp (Figure S7A, Kolmogorov-Smirnov test ).

For intergenic non-coding DE-probes significantly differentially expressed between Basal-like versus Luminal A and B tumor samples a similar distance distribution to neighboured transcription start sites of protein-coding genes was observed (Kolmogorov-Smirnov test ); however, with a larger shift to proximal regions of protein-coding genes for DE-probes upregulated in Basal-like tumor samples (Figure S7B).

Differentially expressed non-coding transcripts and nearest protein-coding genes displayed non-synonymous expression changes

Our results of significantly regulated non-coding loci (abundantly located in proximal regions of protein-coding genes, and enriched in regulatory sites) together with reports of previous studies about frequent cis-regulatory mechanisms of lncRNAs [14], [26], [27], [61], [62] encouraged us to analyze the expression variation for all non-coding DE-transcripts in comparison to their nearest protein-coding gene (Gencode v12). Exclusion criteria for potentially non-annotated distant exons of protein coding genes were: (1) If the non-coding DE-probe and the protein-coding gene were located on the same reading strand, only those lncRNA-mRNA pairs were accepted that exhibited significant expression changes in opposite directions. (2) If the non-coding DE-probe and the protein-coding gene were located on different reading strands, all lncRNA-mRNA pairs with significant expression changes were accepted.

We detected 416 protein-coding genes in nearest genomic proximity of 782 intergenic non-coding DE-probes, where both displayed significant differential expression variation between normal and breast tumor tissue (intergenic lncRNA-mRNA pairs, , Figure 4A). The majority (75%) of those protein-coding genes showed non-synonymous expression changes with at least one intergenic non-coding DE-probe: 279 (32) protein-coding genes were found upregulated (downregulated) in tumor in relation to their nearby downregulated (upregulated) intergenic non-coding DE-probes in contrast to 137 protein-coding genes with synonymous expression variations (Figure 4A).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 4. Proximal lncRNA – mRNA pairs.

For non-coding DE- probes significantly differentially expressed between normal and tumor samples (FDR) the protein-coding gene (Gencode release v12) with closest genome coordinates was identified, and the pair retained if the protein-coding gene was differentially expressed at the same FDR cutoff. Log2 fold change of the non-coding probe (x-axis) and the maximal log2 fold change of probes located in exons of the protein-coding gene (y-axis) is depicted as a bivariate histogram using hexagonal binning (R package hexbin). Pairs with converse fold changes are shown in the left upper and right lower quadrant. Pairs with consistent fold changes but opposite reading direction are shown in the left lower and right upper quadrant (see also panel describing direction of expression changes for each quadrant). Numbers in quadrant correspond to number of unique genes depicted. (A.) Proximal pairs, where the non-coding probe is intergenic. (B.) Pairs where the non-coding probes is in an intron of the protein-coding gene. (C.) Pairs where the non-coding probe and the protein-coding gene are on opposite strands and overlap at least partially.

https://doi.org/10.1371/journal.pone.0106076.g004

Further, a total of 1276 actively transcribed and putatively regulatory non-coding transcripts in introns of 655 regulated protein-coding genes were identified (intronic lncRNA-mRNA pairs, Figure 4B). Again, the majority of those protein-coding genes displayed non-synonymous expression changes with at least one intronic non-coding DE-probe. We found 441 (72) protein-coding genes upregulated (downregulated) in tumor while at least one intronic DE-probe was downregulated (upregulated).

For non-coding DE-probes antisense to protein-coding genes, we observed a total of 865 lncRNA-mRNA pairs comprising 565 unique protein-coding genes (Figure 4C). Here, we identified a balanced fraction of pairs with synonymous or non-synonymous expression changes with the exception of a small number of protein-coding genes (29) that were upregulated in tumor and with an antisense lncRNA downregulated. An example of a protein-coding gene downregulated in breast cancer with a non-coding antisense transcript significantly upregulated in breast cancer versus normal tissue is HDAC3 (histone deacetylase 3) on chr5 (Figure 5). HDAC3 belongs to the class I of histone deacetylase family of proteins, which regulates gene expression by histone deacetylation at promoter sites thus transforming the chromatin into a more “packed” state (heterochromatin) leading to transcriptional repression of genes.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 5. HDAC3 (histone deacetylase 3) mRNA and its putative regulatory antisense lncRNA.

(A.) Genomic locus of HDAC3 on chromosome 5 and the antisense transcript downstream of HDAC3 with genomic positions of strand-specific RT-qPCR primers/products. Annotation track DE-TAR corresponds to genomic loci significantly downregulated upon TP53 induction [43]. Both transcripts appear to be significantly differentially expressed on the custom microarray (), exhibiting a non-synonymous expression pattern (B.). The transcription start site of the annotated antisense RNA overlaps with the transcription start site of DIAPH1. Genome-wide predictions of functional open reading frames (RNAcode, ) correspond mainly to exons of HDAC3 mRNA, while some short putative open reading frames overlap the antisense transcript. (C.) Strand-specific RT-qPCR validations relative to normal sample “RP38” for both, the HDAC3 mRNA and the antisense transcript.

https://doi.org/10.1371/journal.pone.0106076.g005

A KEGG pathway enrichment analysis for all protein-coding genes, which were significantly differentially expressed between normal and tumor tissue and which were a member of the above described lncRNA-mRNA pairs revealed many different cancer-related pathways among the significant KEGG terms (, Table S6).

We observed a less pronounced number of lncRNA-mRNA pairs for the comparison of Basal-like versus the Luminal A and B tumor samples (). Only 15 protein-coding genes with nearby non-coding DE-probes in intergenic space were detected, 17 with non-coding DE-probes in their introns, and 117 with non-coding DE-probes in antisense direction.

In summary, these findings directly support the hypothesis that expression of protein-coding genes associated to cancer relevant pathways might be affected by lncRNAs located in close genomic proximity.

The landscape of non-coding transcription shows dramatic changes in breast cancer

Breast cancer is a highly heterogeneous disease with distinct molecular subtypes characterized by patterns of protein-coding gene expression [2], [3] and genomic DNA alterations [50], [63]. The genomic changes may contribute to alterations at the expression levels; tumors of the same mRNA expression subtype often share similar copy number variations, with the most prevalent genomic alterations observed for the Basal-like and ERBB2 subtypes [63]–[65]. Further, it has been demonstrated that 12% to 40% of the expression variation of the protein-coding genes may be explained by genomic alterations [50],[66]. We investigated whether the observed differential expression of lncRNAs is merely a consequence of genomic aberrations. We found that the observed correlation between copy number variations and lncRNA expression was very low, only 6% of the observed expression changes may be explained by copy number variation. Thus, and in opposite to previous findings for protein-coding genes [49], [50], we conclude that the majority of the observed differential expression of lncRNAs is influenced by other factors than underlaying DNA aberrations.

Overall, 9647 unique non-coding loci were significantly differentially expressed between 26 breast tumor and 5 normal tissue samples. These non-coding transcripts were located in the intergenic space (3194) and in introns of protein-coding genes (5088), further in antisense or bidirectional genomic location relative to protein-coding genes (1365). Several of the detected differentially expressed lncRNAs were well described members of cancer-related pathways, including tumor suppressors and oncogenes (e.g. MEG3 [55], Xist [67], [68], MALAT1 [19], [42], H19 [39], [69], GAS5 [40], and HOTAIR [25], [70]). However, the majority of differentially expressed lncRNAs corresponded to novel transcripts of unknown function. Differentially expressed mRNAs were enriched in cancer- but not in adipocyte-specific KEGG pathways, suggesting that the observed differences between normal and tumor tissue samples are not histological artefacts. We therefore conclude that the observed differential expression of lncRNAs is mainly reflecting the difference between normal and tumorous epithelial cells and is thus associated with breast cancer.

LncRNA differential expression between tumor subtypes was on a much lower scale than between normal and tumor tissue. This may be a consequence of smaller effect sizes and larger within group heterogeneity combined with comparably smaller sample sizes. However, it might also reflect a bias of the custom array. Apart from database derived lncRNAs it contained in house identified lncRNAs from cell cycle, oncogenic, and tumor suppressor pathways. Alterations in these processes are common to many tumors, but transcripts associated with breast cancer subtypes, like estrogen- or progesterone receptor controlled lncRNAs, are likely underrepresented on the array.

Investigating evolutionary conservation of tumor-related non-coding RNAs we observed that selective pressure on primary sequence is small compared to protein-coding exons, which is in line with reports by others [22], [23], [56]. In addition, we detected conservation of secondary structure motifs in mammals for about 10% of non-coding DE-regions regulated between normal and breast tumor tissue (9.8%) and between Basal-like versus Luminal A and B tumor samples (11.5%). This fraction was only marginally explained by the observed number of regulated short ncRNAs usually folding into distinct secondary structures. Given that the majority of DE-probes were less conserved in primary sequence than it is required for optimal signal-to-noise ratios in RNA secondary structure motif detection tasks, a notable number of conserved structure motifs may still be hidden among our set of novel non-coding tumor-related non-coding RNAs. Further, we noticed on average less sequence conservation for lncRNAs differentially expressed within tumor tissue (Basal-like versus Luminal A and B tumor samples) than for lncRNAs differentially expressed between normal and tumor tissue.

Non-coding RNAs may regulate transcription of protein-coding genes in cis

Several lncRNAs have been found to regulate gene expression in cis [14], [26], [27], [61] or in enhancer regions [71]. We hypothesize that cis acting lncRNAs regulate expression of tumor suppressors and (to a lesser extent) oncogenes in a direct or indirect way, and thus contribute to tumorigenesis. Overall, we found a significantly smaller distance of differentially expressed non-coding regions to the proximal protein coding gene than expected by randomly drawing from the transcripts represented on the custom array. We detected a massive deregulation of non-coding transcription from regulatory DNA sites upstream of protein coding sequences. These may partly correspond to short RNAs (50bp to 200 bp length) that are transcribed upstream of coding genes, which are known to be a target of polycomb proteins and can induce repression of coding genes in cis [72].

Further, we identified 311 protein-coding genes with proximal long non-coding RNAs in intergenic space and non-synonymous expression changes, suggesting that these lncRNAs may interfere with mRNA expression. Synonymous pairs were only assessed for RNAs on opposite strands to avoid counting of unknown mRNA exons and these synonymous pairs were significantly fewer than non-synonymous pairs. For the majority of non-synonymous pairs the protein-coding gene was found down- and the intergenic lncRNA was identified being upregulated in tumors samples. We noticed an enrichment of tumor suppressor protein-coding genes for mRNAs downregulated in tumor. Together with the fact that the majority of mRNAs were down- and the majority of non-coding transcripts were upregulated in tumors one may speculate whether the upregulation of lncRNAs contributes to the downregulation of tumor suppressor genes and thus to the progress of cancer.

In few cases, non-synonymous expression in opposite reading direction, exhibiting overexpression of breast cancer-associated mRNAs along with an antisense lncRNA downregulated in tumor samples, was detected. Cabili et al. 2011 found that the expression patterns of lincRNAs and their nearest protein-coding neighbour are not more correlated than pairs of two neighbouring protein-coding genes [23]. Thus, synonymous pairs may just reflect that transcription of a particular genomic locus is accompanied by transcription of neighboured loci because the surrounding chromatin is open [73]. However, lncRNAs can positively affect the transcription of adjacent genes by inducing chromatin looping between the ncRNA loci and the neighbouring genes [74]. Ørom et al. 2010 report intergenic non-coding RNAs acting as enhancers for their nearest protein-coding gene, which results in synonymous expression changes between non-coding RNAs and mRNAs [14]. In summary, our results suggest that lncRNA s may act in the regulatory control of adherent cancer-related genes and their malfunction may present an important factor in the development and progression of breast cancer.

An example for a strongly anticorrelated, non-synonymous pair with a tumor-downregulated mRNA is HDAC3 and its antisense lncRNA transcript. We found HDAC3 expression in normal samples and strongly downregulated but detectable in tumor tissue samples. Class I histone deacetylases, like HDAC3, are repressing the transcription machinery for various genes in cancer [75]. On the other hand, studies reported that HDAC3 is an inhibitor of migration of metastatic breast cancer cells [76].

Chromatin-associated lncRNAs may contribute to epigenetic changes during carcinogenesis

In the breast cancer samples used for our investigations, we observed a massive downregulation of lncRNAs previously described to be a component of chromatin [27]. Chromatin-associated lncRNAs (CARs) have been found to regulate transcription in cis by mediating chromatin modifications in close genomic proximity. We identified downregulated CARs in vicinity or overlapping the oncogenes IGF1R, MYLK [77], the breast cancer drug target OGT (O-GlcNAc transferase) [78], and the breast cancer-related cyclin CCNL1 [79]. CARs also overlap with lincRNAs that were found to bind to PRC2 (polycomb repressive complex 2) or RCOR1 (REST corepressor 1) and regulate transcription in trans through establishing repressive chromatin marks at distant sites [27]. We found 5 of these CARs/lincRNAs (MALAT1, MEG3, NEAT1, NBPF1, and AC058791.1) significantly downregulated in primary tumor samples compared to normal tissue. The observed downregulation of CARs in tumor may thus be responsible for a reprogrammed chromatin state in breast cancer leading to the activation of oncogenes or repression of tumor suppressors in cis or trans.

In summary, the observed massive deregulation of lncRNAs may be an important characteristic of breast cancer development and progression. The frequently observed anticorrelation between lncRNAs and adjacent breast cancer relevant onco- and tumor suppressor genes may give rise to novel drug targets at the non-coding RNA level. Finally, the perturbation of chromatin-associated lncRNAs encourages for a more detailed investigation of the role of lncRNAs in epigenetic reprogramming in breast cancer etiopathology.

Ethics Statement

All studies are approved by the Norwegian Regional Committee (REC) for Medical and Health Research Ethics (REC South East, reference numbers S97103 and 429-04148). All patients are informed and have declared written informed consent that their samples are used for research.

Tissue Samples

Fresh frozen tumor biopsies from early breast cancer cases were collected from 920 patients included in the Oslo Micrometastasis (MicMa) Study - Oslo I from various hospitals (a collaboration between Buskerud-, Bærum-, and different sections at the Oslo University Hospital, Norway) between 1995 and 1998 [28], [44], [80]. A total of 26 breast carcinomas have been selected for lncRNA expression study using the nONCOchip microarray. The samples have been classified into five clinically relevant tumor subclasses based on their mRNA expression levels [44], [81]. In addition, 5 breast tissue samples from breast reduction operations were provided from the Colosseum Clinic, Oslo in co-operation with Akershus University Hospital, Lørenskog and are herein defined as being normal breast tissue.

Custom expression microarray design - nONCOchip

We used the nONCOchip, a custom-designed Agilent microarray containing 203,527 probes covering protein-coding and non-coding genomic loci (GEO accession number GPL13648). The nONCOchip interrogates probes for lncRNAs regulated by three tumor-relevant pathways (mitotic cell cycle, anti-proliferative and pro-apoptotic p53, and pro-proliferative and anti-apoptotic Stat3), known lncRNAs, as well as mRNAs [43].

In detail, probes of 60bp length have been designed following Agilent's standard design protocol for expression exon microarrays as available from eArray (https://earray.chem.agilent.com/earray/). eArray design was performed according to the base composition methodology where probes are equally distributed across the target sequence, and uniqueness of probes have been checked against all known RefSeq (m)RNAs. Target sequences were grouped into three length categories defining the required number of probes. Target sequences of length are represented by exactly one probe, while target sequences of length are represented by up to five probes. Target sequences longer than 1000bp were split into intervals of 1000bp and number of probes selected according to length of subsequence. For target sequences of unknown reading strand, i.e. sequences originating from the transcriptome-wide study of cancer-related pathways [43] or from ncRNA predictions [57], [59], probes for the plus and minus strand have been designed. It should be noted that probes of 60bp length exclude the detection of the majority of smaller ncRNAs, like mature miRNAs.

Custom expression microarray processing

All samples assembled for expression analysis with the nONCOchip custom microarray were prepared for microarray performance using the Agilent Quick Amp Labeling Kit for single color following manufacturer's instructions. RNA quality was checked using Agilent's 2100 Bioanalyzer; only samples with a entered the processing. Instead of the Oligo-dT/T7-primer delivered with the Quick Amp Labeling Kit, we used 120 pmol of a random N6/T7 primer synthesized by Metabion. 1 µg RNA was used as input for the labeling procedure. cRNA quantity was checked using a NanoDrop ND-1000 UV-VIS Spectrophotometer, as enlisted in the manufacturers instructions. 1.5 µg of labeled cRNA was used for hybridization with the 244k custom microarray following manufacturers instructions. After hybridization the arrays were washed according to the manual and scanned using the Axon GenePix 4200 Scanner and GenePix Pro 6.1 Scan software with the following settings for scanning: 100% laser power; focus 0; 5 µm pixel size; 2 lines to average; wavelength at 532 nm with standard green filter. Result tables were extracted after grid placement using GenePix Pro 6.1 Software. Result tables were used for subsequent data analysis.

Custom expression microarray data analysis

Differentially expressed probes were identified by using R [82] and the Bioconductor library limma [83]. Quality control of arrays were performed by checking distribution of “bright corner”, “dark corner” probes, and relative spike-in concentration versus normalized signal. The controls confirmed high quality of the results and consequently all microarray data were included in the downstream analysis. To retrieve a set of probes mapping to unique genomic positions in hg19 we used BLAT [84] with the parameter -minIdentity = 93 allowing to detect probes spanning splice sites. All probes mapping to more than one distinct genomic region were discarded. Normalization between arrays was done by using quantile normalization [85]. In order to reduce the number of t-tests nonspecific filtering was applied as follows: The expression of a probe must be larger than background expression in four arrays. Background expression is defined by the mean intensity plus three times the standard deviation of negative control spots (Agilent's 3xSLv spots). In addition, a probe must exhibit a nonspecific change of expression of at least IQR . Finally, a linear model was fitted using the R package limma and reliable variance estimates were obtained by Empirical Bayes moderated t-statistics. False discovery rate was controlled by Benjamini-Hochberg adjustment [86].

Quality control of the custom expression microarray

For quality control of the expression data, we investigated whether the subtype classification obtained in previous studies for the same set of tumor samples was reproducible using the nONCOchip. Gene expression of PAM50 genes was compared to previous profiling of the same samples using Agilent's Whole Human Genome Oligo 44k Microarrays [81]. High concordance to the original subtype assignments for all tumor samples was revealed depicting sufficient Pearson's correlation coefficients (Figure S1A). The results indicate high molecular reproducibility for mRNA expression profiles applying the nONCOchip, thereby suggesting high qualitative performance for lncRNA expression profiling.

aCGH microarray data analysis

The genotypes aCGH dataset has been published previously [87]. For each sample, the copy number data were log2-transformed, segmented, and probe values were replaced by segment averages, using the Piecewise Constant Fitting (PCF) algorithm [63], [88]. In order to obtain matching copy number and expression data sets, the following procedure was applied to each sample and to each of the non-coding transcripts in the dataset to obtain a corresponding copy number value: A PCF value was found for each of the expression probe position through interpolation of the piecewise constant regression function. The average PCF value over all expression probes associated with a particular probe is calculated. This average then defines the copy number value for that probe in the given sample.

Defining a set of non-coding probes

A bona fide set of non-coding probes in intergenic and intronic regions was constructed from the significantly differentially expressed probes as follows: (1) All probes overlapping with at least one nucleotide with protein-coding exons (independent of reading strand) as annotated in Gencode release v12 [89], UCSC genes [90], RefSeq [47], or Ensembl genes [91] were discarded. (2) Probes overlapping with a significant RNAcode [46] segment () contain de-novo short open reading frames and were discarded from the set of non-coding probes. In detail, genome-wide Multiz alignments [92] for 46 vertebrate genomes have been downloaded from http://hgdownload.cse.ucsc.edu/goldenPath/hg19/multiz46way/ and split into shorter alignments of at most 400bp length to enable parallel processing. Each multiple alignment was assessed by RNAcode in all six possible reading frames, and the longest segment in the reference sequence of maximal score reported. All segments with an RNAcode define de-novo protein-coding regions. We refrained from adjusting p-values for multiple testing, as we are not interested in a set of highly reliable protein-coding segments (i.e. reducing the number of false positives), but in reducing the number of regions falsely interpreted as non-coding (i.e. reducing the number false negatives). An RNAcode p-value of resulted in sensitivity (according to known protein-coding exons annotated in Gencode v12) and specificity (according to 10,000 sampled intergenic intervals preserving length distribution and repeat content of protein-coding exons). (3) All probes which have not been classified by RNAcode (due to low sequence conservation) but overlap with at least one nucleotide genomic regions of translated sequences patterns similar to amino-acid sequences of known human proteins may be of coding origin and, hence, were discarded (tblastn [93] with parameters -word-size 3 -evalue 0.05 and RefSeq database from March 7, 2012). Applying these filters results in 53,577 bona fide non-coding probes in intergenic regions and 71,228 in introns of protein-coding genes. For simplicity, we termed probes passing these three filters as non-coding probes.

Probes antisense to a protein-coding exon, but not containing a significant RNAcode hit () on the same strand define a separate set of antisense non-coding probes comprising in total 14,272 probes.

Statistical analysis of annotation overlaps

Overlap with annotation sets was calculated using R [82] and the Bioconductor library genomeIntervals [94]. We further used the R library Snow to enable parallel processing [95]. For each contrast of interest the overlap with a particular annotation set was computed in terms of (1) absolute number of differentially expressed probes (DE-probes) overlapping with a particular annotation and (2) the odds ratio of observed relative overlap versus relative overlap of a background list consisting of all probes on the nONCOchip. Observed odds, randomized odds and odds ratios are defined as follows:(1)(2)(3)

The observed number of overlapping probes is given as , while corresponds to the number of overlapping probes in the background list. The overall number of DE-probes is denoted as , and the overall number in the background list as . Significance of odds ratios was assessed by using Fisher's exact test as implemented in R. We also report the 95% confidence interval of the odds ratio, which is larger than 1 in case of an enriched number of overlapping nucleotides and less than 1 in case of depletion. A probe was interpreted as an overlapping probe if it maps to at least 90% to an interval of the according annotation set.

Annotation categories

A detailed description and listing of annotations sets used is given in Methods S1 and in Table S7.

KEGG pathway enrichment analysis

KEGG pathway enrichment analysis was performed using the R library GOstats [96] with pathway annotations as stored in KEGG.db and org.Hs.eg.db. Significance of enrichment was assessed by a one-sided hypergeometric test where the universe contains all genes of the nONCOchip which passed nonspecific filtering.

Validation of lncRNA expression analysis by RT-qPCR

For quantitative real time PCR 250 ng of total RNA was reverse transcribed using random primers and the High-Capacity cDNA Reverse Transcription Kit following the manufacturer's recommendations (Applied Biosystems, now Life Technologies). Transcript expression analysis was performed using Fast SYBR Green Master Mix according to the manufacturer's instructions (Applied Biosystems, now Life Technologies). 5 µl of 1∶12.5 diluted cDNA was used per reaction. Experiments were performed in triplets and all transcript quantification data were normalized to β-Actin mRNA. All primers were designed using Primer3 (v0.4.0, default parameters, Table S8). The UCSC In-Silico PCR option (http://genome.ucsc.edu/cgi-bin/hgPcr) was used to check for unspecific primer matches.