Annotation and functional characterization of long noncoding RNAs deregulated in pancreatic adenocarcinoma

Abstract

Purpose

Transcriptome analysis of pancreatic ductal adenocarcinoma (PDAC) has been useful to identify gene expression changes that sustain malignant phenotypes. Yet, most studies examined only tumor tissues and focused on protein-coding genes, leaving long non-coding RNAs (lncRNAs) largely underexplored.

Methods

We generated total RNA-Seq data from patient-matched tumor and nonmalignant pancreatic tissues and implemented a computational pipeline to survey known and novel lncRNAs. siRNA-mediated knockdown in tumor cell lines was performed to assess the contribution of PDAC-associated lncRNAs to malignant phenotypes. Gene co-expression network and functional enrichment analyses were used to assign deregulated lncRNAs to biological processes and molecular pathways.

Results

We detected 9,032 GENCODE lncRNAs as well as 523 unannotated lncRNAs, including transcripts significantly associated with patient outcome. Aberrant expression of a subset of novel and known lncRNAs was confirmed in patient samples and cell lines. siRNA-mediated knockdown of a subset of these lncRNAs (LINC01559, LINC01133, CCAT1, LINC00920 and UCA1) reduced cell proliferation, migration and invasion. Gene co-expression network analysis associated PDAC-deregulated lncRNAs with diverse biological processes, such as cell adhesion, protein glycosylation and DNA repair. Furthermore, UCA1 knockdown was shown to specifically deregulate co-expressed genes involved in DNA repair and to negatively impact DNA repair following damage induced by ionizing radiation.

Conclusions

Our study expands the repertoire of lncRNAs deregulated in PDAC, thereby revealing novel candidate biomarkers for patient risk stratification. It also provides a roadmap for functional assays aimed to characterize novel mechanisms of action of lncRNAs in pancreatic cancer, which could be explored for therapeutic development.

Supplementary Information

The online version contains supplementary material available at 10.1007/s13402-022-00678-5.

Introduction

Pancreatic ductal adenocarcinoma (PDAC) is the most common type of pancreatic neoplasia accounting for more than 80% of cases [1]. PDAC is an extremely lethal disease with an overall 5-year survival rate of less than 5% and is one of the leading causes of cancer-related death worldwide, with an incidence rate close to the mortality rate [2] and projections indicating that it will become the second leading cause of cancer death in the USA within the next two decades [3]. PDAC is refractory to available radio/chemo/immunotherapies and a complete surgical removal of the tumor at early stages remains the best curative option. However, the disease is frequently asymptomatic and, as a result, 80–90% of cases are surgically unresectable [4]. Thus, it is crucial to identify novel biomarkers for early diagnosis and targets for therapeutic intervention.

Global transcriptome analyses have been performed to identify differentially expressed genes and to highlight key molecular pathways that are perturbed in PDAC. Most studies examined polyadenylated-enriched RNA fractions using DNA microarray platforms that interrogated predefined sets of protein-coding genes [5–8]. More recently, next generation sequencing has been employed to interrogate with greater resolution the transcriptional landscape of hundreds of PDAC samples, revealing clinically relevant molecular subtypes of the disease [9–12]. However, these latter studies did not evaluate the composition and changes in the transcriptome of PDAC relative to non-malignant pancreatic tissue and focused on the expression of curated gene catalogs, such as GENCODE [13]. Thus, there is still a lack of knowledge regarding transcriptional changes in PDAC that involve poorly annotated non-coding gene loci.

High throughput transcriptome analysis of the human genome has revealed thousands of long non‐coding RNAs (lncRNAs) that may act in cis or in trans to regulate gene expression through multiple mechanisms in physiological and pathological processes [14–17]. Consistently, lncRNAs are emerging as important players in promoting oncogenesis in different types of cancer, including pancreatic cancer [18]. Even though a number of deregulated lncRNAs functionally associated with malignant phenotypes have been identified in pancreatic cancer [18–21] and the expression levels of subsets of lncRNAs in PDAC samples have been shown to correlate with patient survival [22, 23], the exploration of lncRNAs that could act as biomarkers or therapeutic targets in pancreatic cancer is still in its infancy [24–26], particularly as lncRNA genes are still poorly annotated and validated. As lncRNAs tend to show expression patterns restricted to certain tissues [27], a comprehensive analysis of lncRNAs with aberrant expression in PDAC is an attractive approach to identify novel PDAC-specific biomarkers and therapeutic targets.

In this study, we performed total RNA-Seq and examined the transcriptional landscape of matched samples of tumor and nonmalignant adjacent pancreatic tissue samples from 14 PDAC cases. We catalogued known as well as novel lncRNAs expressed in the exocrine pancreas and identified 86 intergenic lncRNAs with aberrant expression in PDAC, 10 of which were found to be associated with patient outcome. Loss of function assays performed with 5 lncRNAs demonstrated their contribution to sustain malignant phenotypes in pancreatic tumor cells. By constructing protein coding/lncRNA co-expression networks in PDAC, we identified the likely biological processes and molecular pathways associated with the aberrantly expressed oncogenic lncRNAs identified. Among these networks, we found that lncRNA UCA1 was co-expressed with genes involved in DNA repair and we experimentally confirmed that UCA1 is required for efficient DNA damage repair induced by ionizing radiation in PDAC cells. Taken together, our results show that aberrant lncRNA expression in PDAC is pervasive, and that aberrantly expressed lncRNAs have functional importance in pancreatic cancer, suggesting that these transcripts can be used as prognostic biomarkers and/or be explored as therapeutic targets.

Materials and methods

Clinical samples and cell lines

Fragments of tumor and patient-matched peritumoral nonmalignant tissues from 14 PDAC were retrieved from the A.C. Camargo Cancer Center biorepository. Samples were collected with informed consent during tumor surgical resection and immediately snap frozen in liquid nitrogen. All tissue fragments were reviewed by a pathologist for histological confirmation of the malignant/nonmalignant status. Samples containing non-neoplastic tissues, fibrosis, adipose tissue or other contaminants were manually dissected. Whenever necessary, tumor fragments were macro-dissected to guarantee that 70% or more of the sections used for gene expression analysis were composed of neoplastic tissue. Clinicopathological information is provided as supplementary material (Table S1). A panel of 6 patient-derived tumor xenografts (PDXs) from PDAC cases established and maintained in athymic nude mice in our laboratory was used in the validation experiments (Table S2). Experiments involving animals were performed in compliance with protocols approved by the Institutional Animal Care and Use Committee.

Pancreatic cancer cell lines were purchased from the American Tissue Culture Collection (ATCC): Capan-1 (ATCC® HTB-79), AsPC-1 (ATCC® CRL-1682), MIA PaCa-2 (ATCC® CRL-1420), BxPC-3 (ATCC® CRL-1687) and PANC-1 (ATCC® CRL-1469). The PDX-8 cell line was derived from a PDAC PDX generated in our laboratory. MRC-5 fibroblast cells were kindly provided by Dr. C.F. Menck. The cells were propagated as follows: Capan-1, MIA PaCa-2, PANC-1, MRC-5 and PDX-08 were cultured in Dulbecco’s modified Eagle’s medium (DMEM) and AsPC-1 and BxPC-3 cells were cultured in RPMI-1640 medium. The media were supplemented with 10% fetal bovine serum (except Capan-1, for which 20% fetal bovine serum was supplemented, and MIA PaCa-2, for which 10% fetal bovine serum and 2.5% horse serum plus 1% penicillin–streptomycin were supplemented. Cells were cultured in a 5% CO₂ humidified incubator at 37 °C.

RNA isolation

Tissue samples with up to 30 mg of tissue were homogenized in 400 ul RLT lysis buffer with β-Mercaptoethanol (Qiagen, CA, USA) using Precellys® equipment (Bertin Technologies, Montigny-le-Bretonneux, France). Tissue RNA isolation was performed using a QIAsymphony RNA extraction kit (Qiagen, CA, USA) on a QIASymphony apparatus (Qiagen, CA, USA), according to the manufacturer’s protocol. The samples were DNase-treated during the purification process to remove possible DNA contaminants. RNA integrity was evaluated using a RNA nano 6000 analysis chip (Agilent, cat# 5067–4627) on a BioAnalyzer 2000 series instrument (Agilent Technologies). All samples showed good quality, with an RNA integrity value ≥ 7. Cell line RNA isolation was performed using Trizol reagent (Invitrogen) following the manufacturer’s protocol.

Total RNA library preparation and sequencing

RNA-Seq libraries were generated using an Illumina TruSeq Stranded Total RNA LT sample preparation kit with Ribo-Zero Gold (cat# RS-122–2001) according to the standard manufacturer’s protocol. Briefly, the ribosomal RNA (rRNA) depletion step was performed using 1 μg starting RNA and the fragmentation step aimed at producing libraries with an insert size in the range of 120–200 bp. Complementary DNA was then synthesized from the rRNA depleted fragmented RNA using SuperScript II Reverse Transcriptase (Invitrogen, (cat# 18,064) and random primers. The resulting cDNA was converted into double-stranded DNA in the presence of dUTP to prevent subsequent amplification of the second strand and thus to maintain the strandedness of the library. Following 3′ adenylation and adaptor ligation, libraries were subjected to 15 cycles of PCR to produce RNA-Seq libraries ready for sequencing. Library quality was verified by capillary electrophoresis (Agilent High Sensitivity DNA Analysis kit, Cat# 5067–4626) on a BioAnalyzer 2000 series instrument (Agilent Technologies). Library quantification was performed by qPCR using a KAPA Library Quantification Kit—Illumina/Universal (KAPA Biosystems, Cat# KK4824). Libraries were pooled and sequenced at low-coverage in an Illumina MiSeq platform to confirm their quality, and then sequenced at higher coverage in a HiSeq 2500 system with TruSeq SBS Kit v3—HS (200-cycles) reagents (Illumina, Cat# FC-401–3001), to generate 100 bp paired-end reads.

RNA-Seq data processing and assembly

Following demultiplexing of sequenced reads, quality control was performed using FastQC v0.10.1 [28]. Adapters and low quality sequences were filtered out with Trimmomatic [29] using the following parameters: sliding window 4:15, minlen: 35. Reads passing the previous filters were aligned against the human genome reference (version hg19) using TopHat2 software [30]. Tumor and non-tumor sample aligned reads were assembled separately using StringTie [31]. To limit the inclusion of possible sequencing artifacts, low abundance (FPKM < 0.1) and short RNAs (length < 200nt) were excluded, and the two sets of assembled transcripts (“initial sets”) were combined in a non-redundant “merged set” using the Cuffmerge tool [32]. Next, the pancreatic transcriptome assembly was annotated according to transcripts listed in the GENCODE human gene catalog (v.19) using the Cuffcompare tool [33] to identify known and unannotated coding and noncoding transcripts expressed in pancreatic tissues from PDAC patients. A “discovery set” was produced after filtering unprocessed RNAs and specific GENCODE biotypes as follows: (i) unannotated intergenic and antisense mono-exonic transcripts without evidence of processing by the splicing machinery, (ii) novel transcript isoforms with an aberrant large exon (e.g. longer than 2,420 nt, which encompasses 99% of GENCODE transcripts) not confirmed by the miTranscriptome catalog [34], (iii) GENCODE biotypes “Sense intronic”, “miscRNA”, “3 prime overlapping ncRNA”, “IG_C_gene”, “IG V gene”, “mitochodrial rRNA”, “non stop decay”, “retained intron”, “nonsense mediated decay”, “sense overlapping”, “snoRNA”, “snRNA”, “TR C gene”, “TR V gene”, “possible pre-mRNA fragments”, “intronic RNAs” and “polymerase run-on fragments”. Novel intergenic/antisense transcripts and splicing variants in the discovery dataset were further annotated using lncRNA catalogs from miTranscriptome v.2, a computational reconstruction of the human transcriptome based on over 7,000 RNA-Seq libraries from tumor/nontumor tissues and cell lines [34], and NONCODE v.4, a curated lncRNA database that collects data from various resources [35].

The abundance levels of gene loci and transcripts in our Discovery set (Transcript Per Million—TPM) were estimated using RSEM [36]. RSEM was also used to quantify the abundance (in TPM units) of known/unannotated transcripts from our discovery set in independent PDAC cases using RNA-Seq data generated by TCGA (n = 113) and ICGC (n = 76), retrieved and processed through the Cancer Genome Collaboratory cloud computing resource (https://cancercollaboratory.org). Only transcripts detected (TPM > 0) in at least 20% of the TCGA/ICGC samples were considered as expressed in these PDAC datasets.

Unsupervised clustering and differential gene expression analysis

Log-transformed normalized counts were used for principal component analysis (PCA) and hierarchical clustering analysis using DESeq2 [37] built-in functions. Visualizations of PCA and hierarchical clustering with the top 500 most variant (inter-sample variance) coding and noncoding transcripts in the “discovery set” were generated using the ‘ggplot2’ R package [38]. Raw read counts were used as input for differential expression analysis between patient-matched tumor and non-tumor samples performed using DESeq2. A threshold of fold change >|10| and adjusted p value ≤ 0.001 was used to select differentially expressed transcripts.

Survival analysis

Cancer specific survival data of PDAC cases from ICGC (PACA-AU, n = 76) and TCGA (PAAD-US, n = 113) were retrieved from the consortia data portals. The expression quantification of the transcripts from our “Discovery set” in the TCGA/ICGC RNA-Seq data was described above (see “RNA-Seq data processing”). For each detected transcript, patients were ranked according to their expression level and assigned to “high” or “low” expression groups based on the median expression level of all samples. Kaplan–Meier (KM) survival curves were generated with R ‘survival’ package [31] and differences between KM curves were ascertained using the log-rank test. Hazard ratios (HR) were calculated relative to the “high” expression group (i.e. HR > 1 conveys that the higher expression group is at higher risk) using Cox regression and a 95% confidence interval. P values ≤ 0.05 were considered statistically significant.

Association with cis-regulatory DNA elements

Genomic coordinates of candidate cis-regulatory elements (ccREs) in transcriptionally active genomic regions experimentally defined by the presence of DNase hypersensitive sites and further annotated using ChIP-Seq data of histone modifications (H3K4me3, H3K27ac) were obtained from the ENCODE project (http://screen.encodeproject.org). The ccREs are subclassified in ‘Promoter-like’ (enriched in H3K4me3 and DNase signals, within 200 nt of a transcription start site, TSS) or ‘Enhancer-like’ (enriched in H3K27ac and DNase signals). The BEDTools package [39] was used to annotate the ccRE closest to the TSS of novel splicing isoforms, intergenic and antisense transcripts, as well as GENCODE annotated protein-coding RNAs and lncRNAs. To infer the statistical significance of the association of specific ccREs with TSSs, we created 50 control datasets of randomly selected genomic sequences and used the Kolmogorov–Smirnov (KS) test statistics to compare the average distance distribution of ccREs to the TSS for each transcript type and those observed for each random control set. Distance distributions of regulatory motifs from intronic/intergenic lncRNAs/protein-coding mRNAs were considered significantly different from those obtained by chance if all KS p values calculated using each random set were smaller than 0.05.

Coding potential analysis

The protein-coding potential of unannotated intergenic and antisense RNAs and novel splicing variants of GENCODE transcripts was evaluated using two computational approaches: CPC [40] and PhyloCSF [41]. CPC uses Open Reading Frame (ORF) features and sequence alignment information against protein databases to generate a support vector machine (SVM) learning classifier [40]. PhyloCSF is a comparative genomics method that uses multispecies nucleotide sequence alignments to derive phylogenetic models of codon substitution frequencies and to evaluate the coding potential of transcript models [41]. Based on the coding-potential score estimated by each method, the transcripts were classified as ‘non-coding’ (CPC score < -1; or PhyloCSF score ≤ -5), ‘non-coding weak’ (-1 ≤ CPC score < 0; or -5 < PhyloCSF score < 0), ‘coding weak’ (0 ≤ CPC score < 1; or 0 ≤ PhyloCSF score < 5) or ‘coding’ (CPC score ≥ 1; or PhyloCSF score ≥ 5). ORF analysis was performed using TransDecoder [42], and only ORFs encoding > 100 residues were considered in similarity searches against the Swiss-Prot database using BLASTP [43] and considering an E-value < 1E-5 as a significance threshold.˃

Gene co-expression network and functional enrichment analysis

Weighted Gene Co-expression Network Analysis (WGCNA) R package [44] was used to investigate correlation patterns between protein-coding mRNAs and lncRNAs expressed in tumor/nontumor pancreatic samples. Only transcripts with an average TPM > 0.075 and standard deviation ˃ 0.75 across all samples were considered for network assembly. WGCNA calculates a transformed Pearson’s correlation matrix for transcripts, in which transcripts are network nodes connected with a variable connection strength (similarity) defined as the absolute value of the Pearson correlation raised to the exponent β. A β parameter of 8 was chosen as the value that best approximates a scale-free topology. Therefore, WGCNA assigns a connectivity measurement that describes how central (or hub) the transcript is in a given network. The WGCNA function ‘dynamic tree cut’ was used to obtain 22 modules, i.e., clusters of connected transcripts. The procedure was performed 100 times using 99% of the dataset each time, and network nodes that did not have a clustering consistency of at least 50% were discarded. Transcripts in each module ranked from most to least connected were used as input for functional enrichment analysis using the gProfiler tool [45], considering annotations from Gene Ontology [46], KEGG [47] and Reactome [48] and all transcripts in the network as the background distribution. A significance threshold of p < 0.05 was used to select enriched biological processes and molecular pathways. For each lncRNA, the average network similarity to all protein-coding transcripts in the pathway was compared to the average network similarities of all protein-coding transcripts assigned to the pathway, to define its relative connectivity (“hubness”). An average network similarity greater than the median similarity value measured within the pathway, means that the lncRNA displays more network similarity to that pathway than half of its annotated members, indicating a strong pathway interaction. The correlation of the first principal component of the expression matrix of each module (i.e., the module eigengene) with the sample histology trait ‘tumor’ or ‘nontumor’ was used as a measure of the association between the respective submodules and the malignant phenotype.

Single cell RNAseq data analysis

A processed count matrix with scRNA-Seq expression data from 57,530 cells from 24 PDAC and 11 nontumor pancreatic samples from [49] was retrieved from GSA (https://ngdc.cncb.ac.cn/gsa/; accession CRA001160). For each candidate lncRNA, the relative count frequency of positive cells (1 or more reads/cell) in the 10 cell types detected in the tumor microenvironment was compared between tumor ductal cells type 1 and 2. The chi-square independence test was used to estimate the significance of differences between observed and expected frequencies.

Validation of novel transcripts

The expression of a subset of unannotated transcripts detected in the transcriptome assembly in PDAC was confirmed by reverse transcription of RNA followed by endpoint PCR using primer sequences that flanked predicted exon junctions (Table S3). Five μg total RNA from PDAC derived cell lines was reverse transcribed using random primers and SuperScript III reverse transcriptase (Invitrogen, Cat#. 18080044) according to the manufacturer's protocol. One-tenth of the cDNA reaction was used as template for 35 rounds of PCR amplification with each primer pair. The PCR reactions were cleaned up using a QIAquick PCR cleanup kit (Qiagen, Cat#28,104) and quantified using a Nanodrop spectrophotometer. An equal mass (50 ng) of DNA from each reaction was then loaded in each lane of a 2.0% agarose gel, post-stained with ethidium bromide and documented in a UV transilluminator.

Validation of lncRNAs aberrantly expressed in PDAC by RT-qPCR

Total RNA (0.5-1 µg) from pancreatic tissue (tumor, nontumor) and PDX samples were isolated and reverse transcribed as indicated above. The differential expression in PDAC of a subset of unannotated and known lncRNAs was confirmed by RT-qPCR using an Applied Biosystems® 480 7500 Real-Time PCR System (Thermo Fisher). PCR assays were performed in triplicate using 1 µl cDNA template, 0.6 µM gene-specific forward and reverse primers designed using Primer Express 3 software (Thermo Fisher Scientific) and 6 µl 2 × SYBR Green Master Mix (Thermo Fisher) using the following cycling parameters: 95 °C for 1 min, 40 cycles of 95 °C for 10 s, 60 °C for 30 s and 72 °C for 10 s, followed by 72 °C for 5 min. For each amplicon, melting curves were examined to confirm the amplification specificity. HMBS was used as a reference for normalization across all samples. The sequences used for each primer pair are listed in Table S2. Relative levels of gene expression were calculated using the 2^−ΔΔCt Ct method [50] and expressed as fold change in tumor relative to non-tumor samples.

siRNA transfections

Transient knockdown was used to investigate the contribution of lncRNAs upregulated in PDAC to malignant phenotypes. AsPC-1 cells were seeded in 6-well plates at a density of 1 × 10⁵ cells/well 12 h before transfection. siRNA transfections were performed with 50 nM of either a non-targeting siRNA control or siRNA SMART pools (siGENOME siRNA, Dharmacon) targeting pre-selected lncRNAs (LINC001559, LINC001133, CCAT1, LINC00920 and UCA1) using Lipofectamine 3000 (Invitrogen) according to manufacturer's instructions. Cells were collected for RNA analysis, and cell-based functional assays were performed 48 h after transfection. The efficiency of lncRNA depletion relative to control was evaluated by RT-qPCR as described above.

Proliferation and cell motility assays

siRNA-transfected AsPC-1 cells were seeded at low density (2 × 10⁴) in 24-well plates in triplicate (in duplicate for control siRNA) after which cell growth was monitored in real time for 8 days in a zenCELL Owl live cell imaging microscope (innoME GmbH, Germany). Images were collected at 20 min intervals and the equipment software calculated the percentage of confluence of adherent cells at each data point [51]. The data were used to adjust linear regression fits, and differences in cell growth rates between test and control cells were determined by testing for statistically significant differences (p < 0.001) in regression line slopes (GraphPad Prism 8, GraphPad Software Inc.). The motility in test and control cells was estimated by measuring the migration distance of 15 individual cells per condition at every 20 min for 6 h (18 points in total) using ImageJ software [52].

Transwell migration/invasion assay

AsPC-1 cells were transfected with siRNAs as described above and 48 h post-transfection, 1 × 10⁵ cells were re-suspended in 300 μl serum-free medium and layered into the upper chambers of 24-well transwell inserts with 8 μm pore size membrane filters with or without Matrigel® coating (Corning, NY, USA) for migration and invasion assays, respectively. The inserts were mounted on top of a lower chamber containing 500 μl RPMI-1640 medium supplemented with 20% FBS and incubated for 24 h at 37 °C in 5% CO₂. Non-migrating/invading cells were removed from the upper surface of the membrane using a cotton swab, after which migrating/invading cells on the bottom surface were fixed with methanol and stained with 0.1% crystal violet. Images were obtained under an IX51 Inverted Microscope (Olympus). The area covered by cells (stained area) from five random fields from three independent experiments were measured using ImageJ software.

DNA damage and repair assay

siRNA-treated AsPC-1 cells (6 × 10⁴) were seeded in 12-well plates in duplicate, incubated for 48 h and subjected to ionizing gamma irradiation at a dose of 10 Gy. Irradiated cells were kept in culture for recovery for 15 min, 30 min, 1 h or 2 h and subsequently processed for alkaline comet assays. Non-irradiated control cells were processed in parallel. For the comet assays, cells were trypsinized, resuspended in 0.5% (w/v) low-melting agarose, layered on top of glass slides previously prepared with a thin layer of 1.5% (w/v) agarose, covered with coverslips to ensure homogeneous spread of cells over the slide surface and incubated overnight at 4 °C in lysis solution (2.5 M NaCl, 100 mM EDTA, 10 mM Tris pH10) for membrane disruption. Next, samples were subjected to DNA denaturation (30 min, 4 °C) and subsequent electrophoresis in alkaline pH buffer (300 mM NaOH, 1 mM EDTA, pH > 13) at 25 V and 300 mA for 30 min at 4 °C. After electrophoresis, the slides were immersed in neutralization solution (0.4 M Tris, pH 7.5) for 15 min followed by fixation in absolute ethanol for 5 min and staining with EtBr (20 µg/ml) for DNA visualization. Slide images were captured in a fluorescence microscope Olympus BX51 and analyzed using Andor Komet 6.0 software (Oxford Instruments Inc.), which measures the amount of genomic DNA damage expressed as “Olive tail moment units”, i.e., the product of the tail length and the fraction of total DNA in the tail. One hundred cells were examined in each condition/time point (50 cells/replicate).

Statistical analysis

All values are presented either as average ± S.D. or as representative images of at least three independent experiments. The statistical tests used in the bioinformatics analyses performed are described above. For all RT-qPCR and functional assays, the independent replicate measurements were tested to determine if the data followed a normal distribution, and statistical inference of differences between conditions was performed using unpaired parametric Student’s t-test or nonparametric Mann–Whitney test, as appropriate. One-way ANOVA with Tukey post-hoc test assuming a confidence interval of 99% was used to infer differences between test and control samples in the comet assay.

Results

Reconstruction of the PDAC transcriptome reveals thousands of unannotated transcripts

We generated rRNA-subtracted total RNA-Seq data from surgical fragments of matched tumor and adjacent histologically nonmalignant pancreatic tissues from 14 PDAC cases. The associated clinicopathological information is presented in Table S1. On average, 14.7 million genome aligned 100-nt read pairs were obtained from each sample, representing a 3.4 Gb mean coverage of the pancreatic transcriptome. As expected, this dataset was enriched in non-polyadenylated RNAs in comparison to RNA-Seq data with greater coverage generated from poly-A + selected libraries by TCGA, exemplified by detection at higher levels of RNAs encoded by RPPH1, TERC and histone genes (Supplementary Fig. 1). A hierarchical approach was used to reconstruct the transcriptome of PDAC and nontumor pancreatic tissues (Table 1). Reads from tumor or non-tumor sample libraries were assembled separately and the resulting transcripts were filtered to remove poorly detected (FPKM < 0.1) and short (length < 200nt) RNAs to generate two ‘initial sets’ of reconstructed transcripts. Next, tumor and non-tumor assembled transcripts were combined to generate a ‘merged set’ that was annotated using the GENCODE catalog (v.19). Unannotated intergenic and antisense mono-exonic transcripts (n = 18,453 and 1,495, respectively) were excluded from the analysis, as they are less likely to represent bonafide processed transcripts. The merged set was further filtered to remove unprocessed RNAs with aberrant large exons and other possible artifacts (see Materials and methods for details). The resulting dataset, referred to as the ‘discovery set’ (Table 1) comprised 48,950 transcripts (88% of total) originating from 18,221 GENCODE loci as well as 6,711 unannotated transcripts (12% of total) from 4,426 loci. The GENCODE annotated transcripts were predominantly protein-coding (78%), but also comprised thousands of long non coding transcripts (Table 1); most annotated lncRNAs are transcript variants without an observable ORF originated from protein-coding loci (Ensembl biotype “processed transcript”, n = 5,993), followed by transcripts from expressed pseudogene loci (n = 1,566), antisense RNAs (n = 1,526) and long intergenic RNAs (n = 1,473). Multi-exonic unannotated transcripts predominantly represented novel splicing variants from known genes (n = 6,188), whereas a minor fraction mapped to intergenic regions (n = 339) or were antisense (n = 184) to Ensembl gene models (Table 1).

Table 1.

Reconstruction of the pancreatic transcriptome

	loci		transcripts
	tumor	non-tumor	tumor	non-tumor
‘Initial set ‘	30,625	28,081	73,613	58,762
‘Merged set’	39,435		90,522
‘Discovery set’	19,164		55,661
Annotated (Ensembl biotypes)	18,221		48,950
protein-coding	14,069		38,392
processed transcript	3,968		5,993
pseudogene	1,527		1,566
lincRNA	1,051		1,473
antisense RNA	1,162		1,526
Unannotated	4,426		6,711
splicing variant	4,002		6,188
intergenic RNA	298		339
antisense RNA	145		184

Number of genes and transcripts detected in the pancreatic transcriptome comprising Ensembl annotated and unannotated transcripts (see Methods for details)

We sought additional evidence to support the existence of the novel transcripts detected in our PDAC transcriptome reconstruction by comparing their structural and regulatory features with those from GENCODE transcripts with the same biotype (Fig. 1A-E). Similar to annotated long intergenic (lincRNAs) and antisense ncRNAs, the novel intergenic and antisense RNAs were found to have fewer exons (Fig. 1A) and to be 1 order of magnitude less abundant than protein-coding mRNAs (Fig. 1B). On the other hand, unannotated splice variants were found to be similar to protein-coding genes in exon number distribution and transcript abundance (Figs. 1A and B). The length distribution of the novel intergenic/antisense RNAs detected in our reconstruction is comparable to the length distribution of those annotated in GENCODE (mean ~ 500 nt; median ~ 1500 nt), but to include a subpopulation of longer transcripts with several thousand nucleotides (Fig. 1C). Similar results were obtained for the unannotated splice variants, which tend to have a length distribution similar to protein coding genes, but to include a larger fraction of longer transcripts (Fig. 1C). We also evaluated the distribution of cis-regulatory DNA elements associated with transcriptionally active chromatin relative to the transcription start site (TSS) of known and unannotated transcripts. Similar to GENCODE transcripts, all unannotated transcript classes showed an excess of ‘promoter-like’ (Fig. 1D) or ‘enhancer-like’ (Fig. 1E) ccREs closer to their TSSs compared to control sets of randomly selected sequences (KS test p < 0.05, see Materials and methods for details). Of note, ~ 40% of the novel intergenic and antisense RNAs exhibited enhancer–like signals within 1 kb from their TSSs (Fig. 1E, red and dark blue lines, respectively), which is approximately twice those observed for GENCODE annotated transcripts in these classes. We also found that 56% of the novel intergenic and 42% of the novel antisense transcripts displayed exactly the same exon/intron structure of lncRNA models reported in the miTranscriptome [34] or NONCODE [35] catalogs (Fig. 1F). Approximately half of the novel splicing variants overlapped with miTranscriptome models, but only a small fraction of these comprised lncRNAs (Fig. 1F). Together these results indicate that our transcriptome reconstruction comprises thousands of yet unannotated bonafide RNAs present in cancerous and nonmalignant adjacent pancreatic tissues.

Fig. 1.

Novel transcripts display structural regulatory genomic features similar to known protein-coding and intergenic/antisense RNAs. Distributions of number of exons (A), abundance (B) and length (C) of the reconstructed transcripts. Distance distribution of GENCODE annotated promoter (D) and enhancer-like (E) regions relative to the TSS of known novel transcripts. All distributions were significantly different (Kolmogorov–Smirnov test, p < 0.001) to those observed for a randomly selected set of sequences. Transcript types are color-labeled as follows: lilac, potentially novel isoform; red, novel intergenic RNA; blue: novel antisense RNA; black, GENCODE protein coding mRNA; orange: GENCODE lincRNA; light blue, GENCODE antisense RNA; gray: random sequences. The overlap between each class of unannotated transcripts detected in our study lncRNAs from miTranscriptome NONCODE catalogs is shown (F)

Annotation of novel lncRNAs expressed in pancreatic tissues

We hypothesized that a fraction of the unannotated transcripts detected in our transcriptome assembly might represent novel lncRNAs. To address this, we employed two computational methods, CPC and PhyCSF, to infer the protein-coding potential of these transcript models (see Materials and methods for details). Using CPC or PhyloCSF, most unannotated intergenic and antisense RNAs were classified as noncoding or with low coding potential (59% and 68%, respectively) or (89% and 77%, respectively) (Fig. 2A). Conversely, we found that the majority of the novel splicing variants encoded polypeptides according to CPC (97%) and PhyloCSF (54%). Interestingly, PhyloCSF classified almost half (46%) of unannotated splice variants as noncoding, indicating that this method is more stringent than CPC in assigning coding potential, which might be explained by the distinct assumptions considered by the algorithms implemented in CPC and PhyloCSF [40, 41]. Consistently, the overlap observed between transcripts classified as “noncoding/coding weak” by each method was partial (Fig. 2B). To further document the potential of these novel transcripts to be translated into cellular polypeptides, we searched for ORFs encoding > 100 residues and compared the ORFs with the sequence of known proteins. Overall, we confirmed that PhyloCSF coding calls were more stringent, with a greater fraction of transcripts with ORFs similar to known proteins being classified as noncoding, which was particularly evident in the set of novel splicing isoforms of protein coding genes (Fig. 2B). Approximately half of the unannotated intergenic and antisense transcripts encoded small ORFs, but ~ 85% of these showed no similarity to the sequence of known proteins (Fig. 2B), being consistent with the hypothesis that these sequences are for the most part novel lncRNAs expressed in pancreatic tissues.

Fig. 2.

The protein-coding potential of unannotated intergenic/antisense RNAs and novel splicing variants of GENCODE transcripts was inferred using the CPC or PhyloCSF programs (A) based on scores calculated with each approach (see Materials methods for details). (B) Venn and pie diagrams showing the intersection between non-coding/non-coding weak transcripts (coding potential score < 0) detected by CPC or PhyloCSF Open Reading Frame (ORF) analysis performed using TransDecoder (https://transdecoder.github.io), respectively. Only ORFs with more than 100 residues were annotated. Similarity to the Swissprot database (based on a 1.0 E⁻⁵ significance threshold) was evaluated using BLASTP. Dark blue; transcripts without an ORF; medium blue: ORFs without similarity to known proteins; light blue: ORFs with significant similarity to Swissprot

Molecular signatures of lncRNAs in PDAC

To assess whether the abundance levels of known and novel lncRNAs in our discovery dataset could indicate molecular differences between tumor and patient-matched adjacent nontumor tissues we performed two unsupervised analyses, namely PCA and hierarchical clustering. In both analyses we found that sample clusterization considering solely intergenic/antisense lncRNAs was able to distinguish tumor from nontumor pancreatic tissues (Fig. S2, panels C and D), with a performance comparable to the one obtained using protein-coding mRNAs (Fig. S2, panel A). Similar results were obtained for novel splicing isoforms (Fig. S2, panel B). Next, we evaluated differential expression between tumor and nontumor samples of the 55,661 transcripts in the ‘discovery set’ (Table 1). Applying a stringent significance cutoff (absolute fold change > 10; adjusted p value ≤ 0.001, see Materials and methods for details), we identified a total of 975 differentially expressed transcripts, including 86 intergenic lncRNAs, 49 antisense lncRNAs, as well as 171 novel splicing isoforms and 669 protein-coding mRNAs (Fig. 3, panels A − D). Interestingly, most differentially expressed intergenic lncRNAs (52 out of 86, 60%) and antisense lncRNAs (35 out of 49, 71%) are as yet unannotated in the GENCODE catalog. To obtain independent validation of the differential expression of intergenic/antisense lncRNAs and splicing variants detected in our analysis, we examined their abundance in a larger number of PDAC cases using RNA-Seq data generated by the TCGA and ICGC consortia. As observed for known protein-coding transcripts and lncRNAs, nearly all novel lncRNAs and splicing isoforms differentially expressed in patient matched tumor/nontumor adjacent tissues were detected in both the TCGA and ICGC datasets (Table 2). Subsequently, we used the available clinical information to ask whether the relative expression of these transcripts could have prognostic value for PDAC patients. In addition to 187 known and 22 novel splicing isoforms of protein-coding genes, we identified 10 intergenic lncRNAs (LINC00675, LINC00941, LINC01559, LINC02577, MIR210HG, UCA1 and four novel lincRNAs: TCONS_00009076TCONS_00017253, TCONS_00024799, TCONS_00036574) and 4 antisense lncRNA (HLA-AS1, MUC12-AS1 and two yet unannotated antisense lncRNAs in the MT-ND1 and TRBV7-6 loci) whose expression levels were significantly associated (log rank test < 0.05) to the overall survival of PDAC patients in TCGA or ICGC datasets (Table 2 and 3).

Fig. 3.

GENCODE annotated novel transcripts differentially expressed in PDAC. Only transcripts with fold change >|10| p-adj ≤ 0.001 were considered significant (see Materials methods for details). Heatmaps with the expression of (A) protein coding (n = 669), (B) novel splicing isoforms (n = 171), (C) intergenic RNAs (n = 86; 34 known, 52 novel) (D) antisense RNAs (n = 49; 14 known, 35 novel) are shown. White bars: non-tumor pancreatic tissue patient samples; black bars: tumor samples. Z-score normalized expression values are shown

Table 2.

The expression of known and novel transcripts differentially expressed in PDAC was evaluated using RNA-Seq from PDAC cases from TCGA and ICGC

Differentially expressed transcripts		ICGC		TCGA		overlap
		# detected	# progn	# detected	# progn	# detected	# progn
protein coding (n = 669)		658	86	650	117	647	16
novel splicing (n = 171)		160	13	162	10	156	1
intergenic RNA (n = 86)	known (n = 34)	33	2	33	4	32	0
	novel (n = 52)	50	2	49	2	48	0
antisense RNA (n = 49)	known (n = 14)	14	1	14	2	14	1
	novel (n = 35)	33	1	34	1	33	1

The numbers of transcripts detected in the ICGC (PACA-AU, 76 samples) and TCGA (PAAD-US; 113 samples) datasets and the number of transcripts significantly associated with patient overall survival are shown (see Methods for details). The overlap between transcripts detected and with prognostic value in both datasets is indicated

Table 3.

Known and novel lincRNAs and antisense lncRNAs significantly associated with OS of PDAC cases

Type	Gene Symbol	Pvalue KM	HR high	CI_HR high	Pvalue HR	Dataset
LincRNA	LINC00675	0.008	0.46	0.25 – 0.82	0.009	ICGC
LincRNA	LINC00941	0.007	2.25	1.23 – 4.11	0.008	TCGA
LincRNA	LINC01559	0.020	2.00	1.10 – 3.67	0.023	TCGA
LincRNA	LINC02577	0.025	1.96	1.08 – 3.57	0.027	TCGA
LincRNA	MIR210HG	0.018	1.99	1.11 – 3.56	0.020	ICGC
LincRNA	UCA1	0.032	1.92	1.05 – 3.52	0.035	TCGA
LincRNA	TCONS_00009076	0.036	1.88	1.03 – 3.43	0.039	TCGA
LincRNA	TCONS_00017253	0.047	1.82	0.99 – 3.31	0.050	ICGC
LincRNA	TCONS_00024799	0.028	0.51	0.28 – 0.94	0.031	TCGA
LincRNA	TCONS_00036574	0.044	0.55	0.31 – 0.99	0.047	ICGC
Antisense	HLA-AS1	0.036	0.53	0.29 – 0.97	0.039	TCGA
Antisense	HLA-AS1	0.001	0.36	0.20 – 0.66	0.001	ICGC
Antisense	MT-ND1	0.039	0.53	0.29 – 0.98	0.041	TCGA
Antisense	MUC12-AS1 (RP11-395B7.4)	0.009	0.46	0.25 – 0.83	0.039	TCGA
Antisense	TRBV7-6	0.001	0.36	0.19 – 0.69	0.001	TCGA
Antisense	TRBV7-6	0.022	0.51	0.28 – 0.92	0.025	ICGC

Hazard ratios (HR) refer to the risk of death in the “high” expression group of patients. See Methods for details

Validation and prognostic potential of lncRNAs aberrantly expressed in PDAC

Our discovery set comprised 339 unannotated intergenic RNAs with little or no coding potential, ranging from 204 to 74,228 nt (average 4.6 kb, median 1.6 kb). Of these, 86 were found to be aberrantly expressed in PDAC compared to nontumor adjacent tissues (Fig. 3C and Table 2), indicating that we have identified novel yet uncharacterized PDAC-associated lincRNAs. A subset of 6 novel lincRNAs overexpressed in PDAC samples (TCONS_00085964, TCONS_00009076, TCONS_00087289, TCONS_00006750, TCONS_59572, TCONS_00036574) that range between 4.6 kb to 21.6 kb in length was selected for validation (Fig. 4A). Of note, two of these candidates (TCONS_00087289 and TCONS_00085964) were reconstructed with exact the same primary structure found in the miTranscriptome catalog. Following reverse transcription of RNA from distinct PDAC cell lines and end-point PCR assays with primers flanking adjacent exons, amplicons with the predicted length were obtained for the 6 lincRNA candidates, further supporting their primary structure and expression in pancreatic tumors (Fig. 4B). Next, we sought to confirm by RT-qPCR the differential expression of these novel lincRNAs in PDAC primary tumors and patient-derived xenografts (PDXs). We found that all 6 candidate lincRNAs were upregulated (3 to 12-fold on average) in primary tumors or PDXs compared to non-tumor pancreatic tissues (Fig. 4C). All six lincRNA candidates were also detected in PDAC RNA-Seq data from the TCGA/ICGC consortia (Table 2). Subsequent comparison of Kaplan–Meier overall survival curves revealed that lower levels of the lincRNA TCONS_00036574 were significantly associated with a shorter patient survival (log rank p = 0.044; hazard ratio, HR = 0.55), whereas higher expression of lincRNA TCONS_00085964 was marginally associated with a worse prognosis (log rank p = 0.085; HR = 1. 66) (Fig. 4D) in the ICGC dataset.

Fig. 4.

Validation and prognostic potential of novel intergenic transcripts detected in PDAC. (A) genome structure of novel intergenic lncRNAs selected for RT-PCR validation. The arrows represent the exon border regions flanked by the PCR primers. (B) expression of selected candidates tested by end-point RT-PCR in PDAC-derived cell lines (AsPC-1, MiaPaCa-2, BXPC-3, PANC-1, PDX-08) fibroblasts (MRC-5). Control reactions with genomic DNA (gDNA), RNA without reverse transcription (no RT) or with no template (NTC) were run in parallel. A 2% agarose gel is shown. Molecular size markers are shown in lane 1. PCR amplicons with the expected size were observed in different cell lines: TCONS00085964 (149 bp), TCONS00009076 (145 bp), TCONS00087289 (188 bp), TCONS00006750 (114 bp), TCONS00059572 (159 bp), TCONS00036574 (153 bp). (C) aberrant expression in PDAC of novel intergenic lncRNAs confirmed by RT-qPCR in clinical samples of PDAC, nontumor adjacent tissue (NT) patient-derived tumor xenografts (PDXs). Results are shown as log2 fold-change values relative to the expression in the NT samples (which was set to 1) following normalization to a reference gene (see Materials methods for details). Bar graphs represent average ± 1S.D. of three independent experiments. Statistical significance was determined by Mann–Whitney test (* p ≤ 0.05) (** p ≤ 0.01) (*** p ≤ 0.001). Groups being compared are indicated by horizontal bars. (D) Kaplan-Meyer survival curves according to TCONS00036574 (upper) or TCONS00085964 (lower) expression using patient samples with clinical expression data generated by the public ICGC consortia (PACA-AU ICGC dataset, n = 76). The log-rank test (p < 0.05) was used to ascertain the statistical difference between survival curves of cases with high (above median) or low (below median) expression of the indicated lncRNAs. Log-rank p values hazard ratios (with 95% confidence intervals) are shown

The discovery set detected in pancreatic tissues comprised thousands of GENCODE annotated non protein coding transcripts originating from intragenic (processed transcripts, antisense RNAs), intergenic (lincRNAs) or pseudogene loci (Table 1). All these transcript classes include subsets that are differentially expressed in PDAC samples (Fig. 3C-D, Table S4). Given the emergence of lincRNAs as regulators of malignant phenotypes in various types of tumors including pancreatic cancer, we sought to confirm the aberrant expression of a subset of lincRNAs in PDAC tissues by RT-qPCR (Fig. 5A). Consistent with our RNA-Seq results, we confirmed upregulation in primary tumors and PDXs of LINC01559 (~ 4-fold), LINC0133 (~ 8-fold), LINC01614 (~ 2.5 fold), LINC00920 (~ 5-fold), LINC02577 (~ 2-fold), CCAT1 (5 to 20-fold), UCA1 (4 to 10-fold) and HOTAIR (~ 3-fold) and downregulation (> 10-fold) of lncRNAs PCAT29 and DRAIC (Fig. 5A). Interestingly, LINC01614 was not detected in PDAC-derived PDX samples and tumor-derived cell lines, suggesting that the expression of this lncRNA is restricted to the tumor stroma and is lost upon xenografting (Fig. 5A). In addition, we evaluated the expression of these lncRNAs across cell types reported in the single cell RNA-Seq analysis of 57,530 cells from 24 PDAC and 11 nontumor pancreatic samples [49]. We found that LINC01133, LINC01559, LINC00920, CCAT and HOTAIR are preferentially expressed in malignant ductal cells type 2 relative to ductal cells type 1 (11, 14, 2, 13 and 7-fold, respectively) and other cell types present in the pancreatic tumor microenvironment (Fig. S3). Finally, we evaluated the expression of these lincRNAs in five different PDAC cell lines by RT-qPCR and noted a variable expression pattern (Fig. 5B left panel). Nonetheless, the RT-qPCR results were highly correlated (r_s > 0.6) to independent measurements obtained by RNA-Seq from the same cell lines [53] (Fig. 5B, middle and right panels).

Fig. 5.

Expression and prognostic potential of lncRNAs differentially expressed in PDAC. (A) expression of GENCODE annotated lncRNAs in clinical samples of PDAC PDXs relative to adjacent non-tumor tissues (NT) confirmed by RT-qPCR. For each transcript, the relative expression refers to the average expression in NT samples (which was set to 1) following normalization to a reference gene (see Materials methods for details). Box plot graphs represent normalized gene expression levels. The box represents the 25th-75th percentile distribution the line inside the box represents the median (50th percentile).The whiskers represent the 9th-91st percentile individual samples are represented as circles. Statistical significance was determined by Mann–Whitney test (* p ≤ 0.05) (** p ≤ 0.01) (*** p ≤ 0.001). Groups being compared are indicated by horizontal bars. (B) relative expression of candidate lncRNAs in different PDAC cell lines measured by RT-qPCR (left panel; expressed as fold-change relative to the abundance of LINC02577 in AsPC-1) or RNA-Seq (middle panel; FPKM values, from ArrayExpress ID: E-MTAB-2706). For each lncRNA, the Spearman correlation between the expression measured by RT-qPCR or RNA-Seq is shown (right panel). (C) Kaplan-Meyer survival curves according to lincRNA expression as indicated using patient samples with clinical expression data generated by the public TCGA consortia (PAAD-US dataset, n = 113). The log-rank test (p < 0.05) was used to ascertain the statistical difference between survival curves of cases with high (above median) or low (below median) expression of the indicated lincRNAs. Log-rank p values hazard ratios (with 95% confidence intervals) are shown

Since these lincRNAs were also detected in TCGA/ICGC RNA-Seq data from PDAC cases (Table 2), we tested whether their expression was associated with patient outcome. Kaplan–Meier survival analysis showed that higher expression levels of LINC01559 (HR = 2.01; p = 0.021), LINC02577 (HR = 1.96; p = 0.025) and UCA1 (HR = 1.92; p = 0.032) were significantly associated with a shorter patient survival (Fig. 5C).

LncRNA depletion reduces proliferation, migration and invasion of PDAC cells

To ask whether the candidate oncogenic lncRNAs detected as upregulated in PDAC (LINC01559, LINC01133, LINC00920, CCAT1 and UCA1) could contribute to sustain malignant phenotypes in vitro, we performed loss of function experiments using RNA interference. We found that compared to AsPC-1 cells transfected with non-targeting control siRNA, AsPC-1 cells transfected with LINC-specific siRNAs displayed robust knockdown of expression (65 to 90%) of each lincRNA analyzed (Fig. 6A). Next, proliferation of siRNA-transfected cells was monitored for 8 days by real-time imaging and confluence values collected at every 2.5 h were used to fit a linear function (Fig. 6B). By comparing the growth kinetics of cultured cells, we observed a significant (p < 0.001) reduction in the confluence of cells with siRNA-mediated inhibition of LINC01559 (20% reduction), LINC01133 (50% reduction), CCAT1 (60% reduction) and UCA1 (50% reduction). No effect on cell confluence was observed upon LINC00920 knockdown (Fig. 6B). The real-time imaging data were also used to evaluate the impact of lncRNA depletion on cell motility by recording the migration distance of individual cells. For each condition, the trajectory of randomly selected cells (n = 15) was measured every 20 min for 6 h and plotted on a cartesian plane (Fig. 6C, left panel) after which the average distance travelled was calculated (Fig. 6C, right panel). We observed, not only a shorter trajectory of cells with siRNA-mediated inhibition of LINC01559, LINC01133, CCAT1 and UCA1 (Fig. 6C, left panel), but also a significant (p ≤ 0.01 to 10^–4) reduction (35 to 50%) in the average distance travelled (Fig. 6C, right panel). No significant change was observed in motility of cells after siRNA-mediated LINC00920 inhibition. The effect of lncRNA depletion on cell motility was independently tested by transwell migration measurements, which corroborated the results obtained by individual cell tracking analysis for LINC01559, LINC01133, CCAT1 and UCA1 (Fig. S4). However, contrary to the individual cell tracking results, in the transwell assays, LINC00920 silencing also reduced cell migration (Fig. S4B). Finally, transwell invasion assays with siRNA-transfected cells revealed a significant (p ≤ 0.01 to 0.05) reduction (40 to 80%) in the number of tumor cells invading through matrigel compared to that of controls for all tested lincRNAs (Fig. 6D and Fig. S5).

Fig. 6.

siRNA-mediated targeting of PDAC-associated lncRNAs reduces proliferation, migration invasion of AsPC-1 cells in vitro. AsPC-1 cells were transfected with siRNAs targeting LINC01559, LINC01133, CCAT1, LINC00920, UCA1 or control siRNA (siCTRL) as described in Materials methods. (A) relative expression (% of control) of target lncRNAs analyzed 48 h post-transfection by RT-qPCR as indicated. Bar graphs represent average ± 1S.D. of three independent experiments. (B) cell growth monitored by real time imaging for 8 days after transfection with test (3 replicates, purple points) or control siRNAs (2 replicates, green points). Dotted vertical lines indicate the change of culture medium at day 2. Values collected at every 2.5 h were used to fit a linear function statistically significant differences in the slope of the linear fits (p < 0.001) are indicated. (C) cell motility was measured by recording the migration distance of individual cells (see Materials methods for details). For each condition, the trajectory of randomly selected cells (n = 15 per condition) was measured at every 20 min for 6 h (left panel) the average migration distance of 15 cells in test (3 replicates) control (2 replicates) cells was calculated (right panel). The left panel includes representative images of individual cells’ trajectories on a bidimensional Cartesian plane normalized do the (0,0) coordinate. The trajectory of each individual cell is represented by a different color. The right panel includes a box plot graphical representation of the distance travelled. The box represents the 25th-75th percentile distribution the line inside the box represents the median (50th percentile).The whiskers represent the 9th-91st percentile individual samples are represented as circles. (D) Transwell invasion assays were performed 48 h post-transfection as described in Materials methods the total area covered by invading cells was determined using ImageJ software. Bar graphs represent average ± 1 s.d of three independent experiments. Statistically significant differences were inferred using Mann–Whitney test (proliferation invasion) or unpaired Student’s t-test (expression migration) the following cut-offs: * p ≤ 0.05; ** p ≤ 0.01; **** p ≤ 10^–4. Groups being compared are indicated by horizontal bars

UCA1 modulates the expression of DNA repair genes and siRNA-mediated UCA1 knockdown reduces DNA repair in vitro

Based on a ‘guilt-by-association approach’ [54], we performed gene co-expression network analysis to infer possible mechanisms of action of lncRNAs that contribute to malignancy in PDAC. First, the ‘discovery set’ of transcripts detected in tumor and nontumor adjacent pancreatic tissues (Table 1) was filtered to remove low abundant transcripts and transcripts with invariant expression across tumor/nontumor adjacent tissues. Next, the remaining transcripts were used to assemble a protein-coding/lncRNA co-expression network using the WGCNA tool [44] (see Materials and methods for details), resulting in twenty-two network modules (Fig. S6A). The assembled network comprised a total of 23,951 transcripts, including over 98% of those deemed as differentially expressed in PDAC samples (i.e., absolute FC > 10 and adj. p < 0.001). One module (gray module) comprised transcripts with low connectivity (similarity) and was not considered further. Next, functional enrichment analysis was performed with transcripts in the 21 remaining modules using the gProfiler tool [45]. Over one hundred significantly enriched biological pathways (q-value < 0.05) were identified, and selected cancer-related pathways are depicted in Fig. S6B. Notably, genes associated with the glandular function of the pancreas were predominantly downregulated in tumor samples, which can be explained by the dedifferentiation and loss of exocrine function of the malignant tissues, whereas categories relevant for tumor maintenance and progression such as “O-linked glycosylation of mucines” (‘purple module’), “cell cycle”, “WNT signaling”, “RHO GTPase effectors” (‘green’ module), “ubiquitin mediated proteolysis” (‘brown’ module), “immune system process” (‘brass’ module) and “Cell adhesion” (‘red’ module) were upregulated in the tumors (Fig. S7B). Each network module comprised a variable fraction of lncRNAs differentially expressed in PDAC compared to nontumor adjacent tissues (light shaded bar areas, Fig. S6B). Several subnetwork modules enriched in specific molecular pathways considered of interest in the context of PDAC and containing differentially expressed lncRNAs with average network similarity greater than the median similarity measured in the pathway (see Materials and methods for details) were identified (Fig. 7A, Fig. S7). Among these, a subnetwork in the ‘green’ module enriched in genes associated with ‘DNA repair’ contained 10 lncRNAs differentially expressed in PDAC, including seven GENCODE annotated lncRNAs (UCA1, LINC02577, LINC00941, TERC, HOXA-AS2, RP11-55,418.2 and RP1-249H1.4) and 3 novel intergenic lncRNAs (TCONS_00088362, TCONS_00009727 and TCONS_00070350) (Fig. 7A). We hypothesized that highly connected lncRNAs may participate in regulatory networks that sustain the expression of co-expressed protein coding genes. To test this hypothesis, we evaluated the effect of UCA1 siRNA-mediated knockdown on the expression levels of SPIDR and PPP4C, two positively correlated genes that are involved in double-strand break DNA repair [55, 56]. We found that UCA1 knockdown resulted in a significant reduction (~ 90%, p < 0.05) in the abundance of SPIDR and PPP4C (Fig. 7B). Conversely, UCA1 depletion caused no effect on the abundance of MUC13, which is upregulated in the PDAC dataset, but does not belong to the DNA repair subnetwork module, indicating that the modulation observed for DNA repair genes is likely biologically relevant. This result prompted us to investigate the functional effect of UCA1 in DNA damage repair. Using an alkaline comet assay, we found that siRNA-mediated inhibition of UCA1 reduced the ability of AsPC-1 cells to repair DNA breaks induced by ionizing radiation (Fig. 7C). 15 min after radiation exposure, the level of DNA damage was comparable in siUCA1-treated and control cells, but the ability to repair this damage, which was measured at time intervals up to 2 h after radiation exposure, was significantly reduced in the UCA1 silenced cells. These results indicate that lincRNA UCA1 is required for the efficient repair of single strand DNA breaks, possibly by modulating the expression of genes involved in the DNA repair machinery.

Fig. 7.

lncRNA UCA1 modulates the expression of DNA repair genes and is required for efficient DNA repair in vitro. (A) a pancreatic gene expression network module enriched in DNA repair protein coding genes co-expressed lncRNAs is shown. Tumor upregulated (red bars) downregulated (blue bars) transcripts are indicated (statistically significant differences are shown in dark colors). Pairwise connectivity values are depicted in gray scale. The upper green bar indicates the relative connectivity (“hubness”) of the lncRNAs in the module. (B) AsPC-1 cells were transfected with a siRNA SMARTpool targeting UCA1 (siUCA1) or control siRNA (siCTRL) as described in Materials methods. Relative expression of genes belonging to the DNA repair co-expression module (SPIDR PPP4C) or from a distinct network module (MUC13) were evaluated 48 h post-transfection. Bar graphs represent average ± 1S.D. of three independent experiments. Significant changes in Mann–Whitney tests are indicated (* p ≤ 0.05). Groups being compared are indicated by horizontal bars. (C) AsPC-1 cells were transfected with a siRNA SMARTpool targeting UCA1 (siUCA1) or control siRNA (siCTRL) as described in Materials methods. At 48 h post-transfection, DNA fragmentation was analyzed by alkaline comet assay before or at different timepoints (15, 30, 60, 120 min) after gamma-irradiation (10 Gy). Representative DNA fragmentation images from individual nuclei are shown in the lower panel the dot plot graphs (upper panel) represent relative levels of DNA damage expressed as OTM units (“olive-tail moment”) in each sample time point as indicated. Horizontal bars represent average relative OTM units in each sample group (siCtrl vs siUCA1). Data were obtained from 50 nuclei per condition measured in three independent replicate experiments. Statistical significance was evaluated by one-way ANOVA with post-hoc Tukey test (* p ≤ 0.05) (**** p ≤ 10^–4) groups being compared are indicated by horizontal bars

Discussion

Comprehensive analyses of mammalian transcriptomes performed in the last two decades, especially of human and model organisms, have revealed the existence of pervasive transcription of lncRNAs that are still poorly catalogued and characterized. Albeit there is still debate regarding the fraction of lncRNAs that are merely transcriptional noise resulting from RNA polymerase misfiring [57], the existence of functional lncRNAs that have regulatory and structural roles in cells and organisms [58], and whose deregulation may elicit oncogenic or tumor suppressive processes in cancer, including PDAC, is undisputable [14, 59]. Therefore, understanding the mechanisms of action of lncRNAs is important, as they represent potential novel targets for cancer therapy. In addition, lncRNAs usually display more distinct tissue-specific expression patterns than protein-coding genes and this observation has motivated the search, among this class of transcripts, for biomarkers with potential for PDAC diagnosis and/or prognosis [23–25].

Previous RNA-Seq studies aimed at interrogating the pancreatic adenocarcinoma transcriptome at high resolution used poly-A-selected [12] or total RNA [11, 60], and focused on lncRNAs already annotated in reference datasets such as GENCODE, which admittedly do not represent the entire complement of the noncoding transcriptome in every human cell type or tissue. Also, these studies sampled hundreds of tumors but did not sequence patient-matched nonmalignant pancreatic tissues, precluding the direct identification of transcriptional changes specifically associated with the malignant phenotype of PDAC tumors. Thus, knowledge of the full complement of lncRNAs expressed in the exocrine pancreas that are deregulated in pancreatic cancer is still incomplete. To fill this gap, we sequenced rRNA-depleted total RNA-Seq libraries from 14 matched tumor and adjacent pancreatic tissues without histological evidence of malignant transformation, and reconstructed transcripts expressed in pancreatic tissues using a hierarchical approach and strict filtering criteria (see Materials and methods). The discovery dataset comprised ~ 55 thousand transcripts originated from ~ 19 thousand distinct loci, most of which (88% and 95%, respectively) were annotated in GENCODE. The annotated PDAC transcriptome mainly comprised protein-coding transcripts (78%), followed by different classes of lncRNAs, including processed transcripts (12%), transcribed pseudogenes (3%), antisense RNAs (3%) and lincRNAs (3%). The majority of the lncRNAs detected in pancreatic tissues (5,993 transcripts) were annotated using the Ensembl biotype “processed transcript”, i.e., transcripts originated from protein-coding loci that do not retain introns and do not contain an ORF, followed by pseudogenes (1,566 lncRNAs). Even though members of these classes of lncRNAs have been shown to be involved in PDAC [61, 62], only a few of these lncRNAs has been characterized in detail. Nonetheless, we identified 70 noncoding processed transcripts and 21 pseudogene lncRNAs that are differentially expressed in PDAC. Further studies will be required to ascertain whether these transcripts represent promising biomarkers and/or therapeutic targets.

Our discovery set also comprised 1,473 lincRNAs and 1,526 antisense lncRNAs, approximately 10% of the total number of annotated lncRNAs in each class. For comparison, 30% of the annotated protein coding transcripts were detected, which is in agreement with the reported greater tissue specificity of noncoding RNAs. The number of lncRNAs with Ensembl biotytpes “lincRNA” and “processed transcript” detected in our analysis (7,466) is comparable to those identified in a RNA-Seq-based transcriptome analysis of 76 high purity PDAC samples performed by the TCGA [12]. In addition to known lncRNAs, our reconstruction of the exocrine pancreas transcriptome revealed hundreds of novel intergenic and antisense RNAs that are as yet unannotated in curated gene catalogs. All intergenic RNAs are located more than 1 kb apart from UTRs of known gene loci, indicating that these are indeed bonafide transcripts rather than uncharacterized untranslated regions of incomplete mRNAs. The number of transcripts expressed in pancreatic tissues is likely to be even greater, since we noted that few highly abundant polyA⁻ RNAs (e.g. 7SL and 7SK small RNAs, mitochondrial RNAs) and mRNAs encoding pancreatic enzymes, dominated the libraries and represented nearly 40% of all mapped reads, thus limiting the ability to detect transcripts expressed at lower levels at the sequencing depth used.

Next, we focused on multi-exonic spliced transcripts to avoid possible artifacts originating from the presence of unprocessed RNAs in our rRNA-depleted total RNA-Seq libraries. Interestingly, we found that the length distribution of the novel intergenic and antisense transcripts identified in our analysis is skewed to longer sizes compared to lincRNAs and antisense lncRNAs annotated in GENCODE, including several RNAs with a length of 10,000 to 70,000 nt. Approximately half of the novel antisense/intergenic lncRNAs was also detected in the miTranscriptome and NONCODE catalogues with exactly the same exon structures. Since the miTranscriptome transcript models were assembled from experimentally generated RNA-Seq data [34], this observation provides an independent validation for their existence and expression in exocrine pancreatic tissues. We also detected over 6 thousand novel splicing variants from protein-coding and lncRNAs, with at least half of them being confirmed in the miTranscriptome dataset. Finally, experimental expression validation was obtained in patient samples, PDXs and tumor cell lines for 6 novel candidate lincRNAs with lengths ranging from ~ 5,000 to ~ 22,000 nt. These included 4 lincRNAs not detected by miTranscriptome, further indicating that our RNA-Seq protocol was effective in identifying novel unannotated transcripts expressed in pancreatic tissues.

In addition, the low protein coding potential of these transcripts, coupled to the presence of regulatory elements close to their putative TSSs, suggest that, for the most part, these transcripts are lncRNAs expressed in a regulated manner. Approximately half of the novel intergenic and antisense lncRNAs harbor candidate regulatory regions at their TSSs that are enriched in H3K27ac and devoid of H3K4me3 marks, consistent with them being enhancer-associated lncRNAs (elncRNAs) [63, 64]. elncRNAs can modulate the recruitment of looping factors that alter chromatin topology and facilitate the recruitment of chromatin-activating complexes to promoter regions of target genes [63]. Also, elncRNAs may play pro-oncogenic roles as exemplified by SWINGN, a elncRNA that positively regulates the activation of GAS6 and other additional distant loci by facilitating binding and activation of the SWI/SNF chromatin remodeling complexes at their promoter regions, thereby inducing malignant phenotypes [65]. Enhancer-associated ncRNAs can also promote tumor suppression, as exemplified by TP53-induced enhancer ncRNAs found to be involved in p53-dependent cell cycle arrest in cancer cell lines [66]. In fact, it has been shown that more than 80% of the genes in canonical cancer signaling pathways exhibit expression that is highly correlated with specific enhancer ncRNAs in at least one tumor type, suggesting important regulatory roles for this class of lncRNA in cancer [67]. Of all the putative elncRNAs identified in our study, 45 were differentially expressed in PDAC samples, suggesting that they may represent potential novel biomarkers or therapeutic targets.

It has been reported that lncRNAs with small ORFs (sORFs) may encode peptides that contribute to malignant phenotypes or that have prognostic value in colon, liver, breast, melanoma and other types of cancer [68, 69]. In this regard, we should note that one third of the 523 novel intergenic/antisense lncRNAs identified in this work harbors sORFs. Although most of them (> 85%) display no similarity to known protein domains, it is plausible that at least a fraction of the 56 that are deregulated in PDAC encode micro-peptides that are functionally relevant in the tumor context.

A number of studies based on focused gene panels or individual genes have reported lncRNAs with prognostic potential in pancreatic cancer. Microarray-based expression profiling of tumor and nontumor pancreatic tissues has shown the existence of molecular signatures of lncRNAs that are associated with malignancy and metastasis [70, 71]. Using both unsupervised and supervised analyses, we identified molecular signatures comprising known and novel lncRNAs of various classes that are deregulated in PDAC. The aberrant expression in PDAC of a subset of known and novel lncRNAs was confirmed by RT-qPCR in tumor and nontumor adjacent samples, PDX models and immortalized cell lines, indicating the robustness of the results obtained in the global differential expression analysis. Nearly all deregulated transcripts in our genome-wide analysis were also detected in tumor samples from the TCGA and ICGC datasets, which allowed an evaluation of the prognostic potential of the PDAC-associated lncRNAs identified in this study in a larger PDAC patient cohort. We found that six annotated lincRNAs (LINC00675, LINC00941, LINC01559, LINC02577, MIR210HG, UCA1) and four novel lincRNAs (TCONS_00009076, TCONS_00017253, TCONS_00024799, TCONS_00036574) with aberrant expression in PDAC displayed a significant association with the overall survival of PDAC cases. Further corroborating the validity of our findings, 5 of the 6 annotated lincRNAs with prognostic value identified in our study have been shown to be associated with clinicopathological features of PDAC patients [72–76]. Of note, lncRNA TCONS_00036574 was upregulated in patient and PDX samples, but counterintuitively, higher expression levels of the lncRNA were found to be associated with a better patient prognosis. There are reports of other cancer-associated lncRNAs that display opposing associations with patient prognosis, depending on the disease type [77]. A plausible explanation may be that lncRNA TCONS_00036574 does not have an oncogenic role or is required to sustain malignant PDAC phenotypes. Rather, in the context of more aggressive tumors, lncRNA TCONS00036574 may reduce tumor progression and, hence, the observed association between a higher expression and a better prognosis. One possibility may be that lncRNA TCONS_00036574 expression is part of a tumor suppressive response in PDAC. In this regard, the ability to induce the expression of this lncRNA may favorably impact disease outcome, as has been observed for tumor suppressor lncRNA GAS5 in glioma [78]. These hypotheses should be explored in loss/gain of function experiments using PDAC models.

Even though a higher expression of LINC00675 has been shown to be an independent predictor of a poor prognosis in a cohort of 90 PDAC cases [76], we found that a higher expression of LINC00675 was significantly associated with a favorable survival in the ICGC PDAC dataset (log rank p = 0.008; HR = 0.46) and that no association of LINC00675 expression and overall survival was observed in the TCGA dataset. Of note, LINC02577 has previously been found to be upregulated in metastatic nasopharyngeal carcinomas [79] and, more recently in colorectal cancer, being associated with a poor prognosis [80]. This is the first report associating higher LINC02577 expression in PDAC with a poor patient survival.

Next, we performed loss-of-function analyses in order to ascertain the functional relevance of the PDAC-associated lncRNAs identified. We found that siRNA-mediated knockdown of lncRNAs CCAT1, LINC00920, LINC01133, LINC01559 and UCA1 impaired the migration and invasion of PDAC cells. With the exception of LINC00920, lncRNA silencing also reduced cell proliferation in vitro. These results corroborate and expand previous functional studies of these lncRNAs, some of which have been associated with PDAC for the first time.

We confirmed an oncogenic role of CCAT1 in promoting the proliferation and migration of PDAC cells [81] and demonstrated its requirement for PDAC cell invasion in vitro which, together with the observation that CCAT1 expression correlates with the expression of EMT markers [81], suggests a role for CCAT1 in modulating metastasis in pancreatic cancer.

LncRNA LINC00920 has recently been reported to act as an oncogenic lncRNA in prostate [82] and colon cancer [83, 84], where it associates with hnRNP-L to induce the expression of the receptor tyrosine kinase AXL, thus enhancing the metastatic potential of tumor cells [84]. Here we show that LINC00920 is upregulated in PDAC patient samples and that its knockdown impairs the migration and invasion of PDAC cells in vitro. Interestingly, the LINC00920 target Axl has been implicated in cellular transformation and PDAC tumor progression [85] and we found that its expression is also significantly upregulated in PDAC tumors (2.5-fold; data not shown). Thus, it is plausible that, like in colon cancer, LINC00920 promotes PDAC oncogenesis by upregulating the expression of AXL.

LINC01133 has been proposed to act as an oncogene in PDAC by promoting epigenetic silencing through promoter methylation of DKK1, a negative regulator of the Wnt/β-catenin signaling pathway in tumor cells [86] and recently it has been found to be expressed in tumor-derived exosomes and to be required for promoting proliferation, migration, invasion and EMT in SW1990 and CFPAC-1 PDAC cells in vitro, as well as tumor growth, EMT and metastasis in vivo [87]. We have independently identified LINC01133 as a top candidate biomarker of PDAC and corroborated its oncogenic role in an additional cell line model.

UCA1 is a well-known oncogenic lncRNA found to be over-expressed in various cancer types and to promote tumorigenesis through different mechanisms in a context-dependent manner (reviewed in [88]. In pancreatic cancer, UCA1 has been found to be differentially expressed between basal-like and classical molecular subtypes of PDAC [12] and functional experiments in cell lines and animal models have shown that UCA1 is required for promoting colony formation, proliferation, migration and invasion and to inhibit apoptosis [72, 89, 90]. In this study, we have confirmed the contribution of UCA1 to the proliferation, migration and invasion of PDAC cells in vitro.

Contrary to bulk tissue expression analyses, which report average gene expression patterns of various cell types present in the tumor microenvironment (TME), including both transformed and non-transformed cells of different origins, single cell analyses allow the identification of cell type-specific expression changes. To gain further insight in the cell type-specific expression of the PDAC-associated lncRNAs identified here, we evaluated their abundance in a single cell RNA-Seq dataset collected from tumor and nonmalignant pancreatic tissues from PDAC patients [49]. By this analysis, we could determine that the lncRNAs upregulated in bulk tumor tissue samples and validated in this study, LINC00675, LINC00920, LINC01133, LINC01559, CCAT1 and HOTAIR, were predominantly detected in type 2 ductal cells, which are expanded and exclusively present in malignant tissues, containing highly proliferative cell subpopulations [49]. Although this cell type comprises 27% of the total number of cells profiled, it accounts for 64% up to 94% of all cells in which the expression of these lncRNAs was detected. This observation further supports the idea that the PDAC-associated lncRNAs identified are expressed preferentially in malignant tumor cells.

Even though the loss-of-function and single-cell expression analyses performed in this work and/or reported in previous studies show that several lncRNAs deregulated in PDACs affect their malignant behavior and are upregulated specifically in the tumor cells, they unveil little information regarding their mode of action. Lack of knowledge regarding the rules that govern structure–function relationships of regulatory lncRNAs remains an obstacle for predicting their mechanisms of action and for the rational design of experiments able to test their involvement in specific biological pathways in cells and animal models [15]. As an alternative, systems biology approaches based on protein coding/noncoding gene co-expression network analysis have been successfully used to establish gene regulatory hubs and to infer putative functions to lncRNAs expressed in human and mouse genomes [91–93]. In the context of pancreatic cancer, WGCNA has been employed to prioritize lncRNAs with biomarker potential in PDAC, leading to the identification of 11 hub lncRNAs present in co-expression modules that are deregulated in two independent PDAC networks [94], 4 of which appeared to be upregulated in PDAC samples in our analysis (ITGB2-AS, LINC00675, LINC01133 and MIR155HG). Considering that the RNA-Seq-based lncRNA signature identified in our study discriminates between patient-matched samples of tumor and adjacent non-malignant tissues with comparable accuracy to the one observed with signatures of protein coding mRNAs, we reasoned that a lncRNA/mRNA co-expression network analysis followed by functional annotation of co-expressed gene modules would be informative to putatively assign lncRNAs with aberrant expression in PDAC to specific biological processes that are modulated to facilitate tumor progression and metastasis. Different from the study from Giuliette et al. [94], we opted to construct a single co-expression network with PDAC and nontumor matched samples and to search for submodules significantly enriched in transcripts encoding proteins with molecular functions/biological processes that were also correlated to the malignant phenotype and contained highly connected differentially expressed lncRNAs.

We identified 21 modules comprising the co-expression network, with 12 being positively correlated to the tumor trait. Those tumor-associated modules contained submodules significantly enriched in genes that participate in processes and pathways that are important to sustain malignant behavior in PDAC, including protein glycosylation [95], O-linked glycosylation of mucines [96], cell cycle regulation [97], DNA repair [98] and WNT [99], as well as Rho GTPase signaling [100]. Interestingly, several of these pathways have been reported as being enriched in WGCNA network submodules identified using proteomics data from 20 PDAC cases [101]. A number of known and novel PDAC-associated lncRNAs are highly connected to protein coding genes in these submodules (Supplementary Fig. S6 and S7) and, based on the co-expression pattern, we postulate that some of these lncRNAs may exert regulatory roles in these specific pathways. This seems to be the case for LINC01559 and UCA1, which were each found to belong to a network submodule enriched in O-linked glycosylation of mucines and in DNA repair genes, respectively. LINC01559 silencing significantly reduced the expression of the co-expressed B3GNT3 and GALNT3 genes in the O-linked glycosylation of mucines module (data not shown). Likewise, UCA1 silencing, not only specifically decreased the abundance of co-expressed genes SPIDR and PPP4C in the DNA repair pathway module, but also negatively impacted the ability of PDAC cells to repair DNA damage induced by ionizing radiation. Interestingly, both SPIDR and PPP4C are known to exert functions in repairing double-strand breaks in DNA [55, 56] and are upregulated in PDAC in our analysis. SPIDR associates with RAD51 and BLM helicase to scaffold a multi-protein complex required for homologous recombination-associated DNA repair and is critical for maintaining genome stability [55]. PPP4C encodes the catalytic domain of the protein phosphatase 4 (PPP4) complex, a ubiquitous serine/threonine phosphatase that modulates the NF-κB and mTOR signaling pathways and, thus, regulates a variety of cellular functions, including the stabilization of stalled replication forks during the DNA damage response [102]. Since the alkaline comet assay used mainly detects single-strand breaks, it is conceivable that UCA1 regulates the expression/function of additional genes in PDAC cells required for efficient repair of single-strand DNA breaks. Moreover, we cannot rule out that UCA1 may affect DNA repair efficiency by directly interacting with other proteins involved in the DNA damage response. Nonetheless, this is the first report linking UCA1 to DNA repair, confirming that the WGCNA approach may uncover novel mechanisms of action, even for well-studied lncRNAs.

Conclusions

Our unbiased transcriptome-wide analysis of expression changes in PDAC has highlighted a number of known lncRNAs with robust prognostic value in pancreatic cancer, as well as identified many unannotated transcripts and revealed pervasive alternative splicing in cancer-related genes with biomarker potential yet to be explored. Through co-expression network analysis we were able to attribute specific molecular functions to PDAC-associated lncRNAs with phenotypic and clinical relevance and validated this approach experimentally by showing that lncRNA UCA1 is required for efficient DNA repair. Taken together, we not only identified novel lncRNAs that have functional importance in pancreatic cancer, but also provided an updated catalog of lncRNAs aberrantly expressed in PDAC, which will be a valuable resource for future studies on novel regulatory mechanisms mediated by lncRNAs and the identification of prognostic biomarkers and therapeutic targets in pancreatic cancer.

Supplementary Information

Below is the link to the electronic supplementary material.

Authors' contributions

Conceived and designed the experiments: VFP, OJS, DVSP, DSB, EMR. Performed the experiments: VFP, OJS, DVSP. BD, TBC, ERB, DOP, VPO, RACZ, LCR. Analyzed the data: VFP, OJS, DVSP, BD, TBC, ERB, DOP, VPO, RACZ, LCRV, FLF, HCF, DSB, EMR. Contributed reagents/materials/analysis tools: LBCAM, JJ, MCCM, JEF, JCS, MDB. Wrote the manuscript: VFP, OJS, EMR. All authors read and approved the final manuscript.

Funding

The study was funded by a research grant from Fundação de Amparo a Pesquisa do Estado de São Paulo (FAPESP) to EMR (grant nº 2013/13844–2) and DSB (grant nº 2016/19757–2). OJS, TB, LBCAM received fellowships from FAPESP. VFP, DVSP, ERB received fellowships from Coordenação de Aperfeiçoamento de Pessoal de Nível Superior (CAPES). DOP received a fellowship from Conselho Nacional de Desenvolvimento Científico e Tecnológico (CNPq). EMR and DSB are recipients of established researcher fellowships from CNPq.

Data availability

All data analyzed during this study are included in this article and its supplementary information files. Raw and processed RNA-Seq data are available at the NCBI-GEO repository under accession GSE130688. The complete annotations of the lncRNAs detected in the PDAC transcriptome and discussed in the article are provided as supplementary material (Table S4).

Declarations

Ethics approval and consent to participate

Clinical samples used in this study were retrieved from the A.C. Camargo Cancer Center biorepository with informed consent. The study was approved by the Ethics Committee of the Institution (nº: 1839/13) and the study is registered at the research regulatory platform of the Brazilian Ministry of Health (CAAE: 15059213.0.0000.5432). The collection of human PDAC specimens for PDX generation was carried out with patient informed consent and approval of the Institutional Review Board of Ethics in Research from collaborating institutions (Hospital Sírio-Libanês, Hospital Alemão Oswaldo Cruz and Hospital 9 de Julho, Sao Paulo, Brazil) and the study is registered at the research regulatory platform of the Brazilian Ministry of Health (CAAE: 36005314.1.0000.5461).

Consent for participate

Not applicable.

Consent for publication

Not applicable.

Competing interests

The authors declare that they have no competing interests.