Towards resolution of the intron retention paradox in breast cancer

We retrieved data from nine tumour types and healthy adjacent tissue, including 615 BrCa patient samples generated by the TCGA. We used TCGA metadata to assign the samples to molecular subtypes (i.e. Luminal A, Luminal B, human epidermal growth factor receptor 2 (HER2)-enriched, and Basal-like) based on the PAM50 classification system. On one occasion (Additional file 1: Fig. S2C), we grouped samples based on the immunohistochemical (IHC) score for HER2.

Only samples for which sequencing had been performed at > 40 M read depth were selected for analysis. Moreover, only tumour types with at least 20 matched tumour/normal tissue samples were considered. RNA-seq data were downloaded as BAM files using the R/Bioconductor package TCGAbiolinks [20] and the command-line tool gdc-client v1.4.0 (github.com/NCI-GDC/gdc-client) under an approved data access application. All files were checked for integrity. Harmonized gene expression data in the form of HTseq counts (RRID:SCR_005514) [21] were downloaded using TCGAbiolinks (RRID:SCR_017683).

Total RNA was isolated from MCF7 and MCF10A cells using TRIzol (Invitrogen). The RNA quality was assessed using RNA 6000 Nano Chips on an Agilent Bioanalyzer (Agilent Technologies) to confirm an RNA integrity score of > 7.0. mRNA-seq was performed by Macrogen (Korea; RRID:SCR_014454) using the Illumina Hi-Seq 2000 platform. RNA-seq libraries were prepared from > 1 μg of total RNA using TruSeq RNA sample prep kit (Illumina) according to the manufacturers’ instructions.

IR was quantified using IRFinder v1.2.0 [17], using the Ensembl human genome (hg38, release 86; RRID:SCR_002344) as reference. The IRFinder algorithm measures 20 parameters for IR detection in each sample, including the median number of reads mapping to each nucleotide across the intron length (intron depth, ID), the ratio of nucleotides within an intron with mapped reads (coverage), the number of reads that map to the 5′ flanking exon and to another exon within the same gene (splice left, SL), the number of reads that map to the 3′ flanking exon and to another exon within the same gene (splice right, SR), the number of reads spanning the exon-exon junction (splice exact, SE) as well as the IR ratio:

Some selection criteria for IR events were chosen to minimize the chance of false positive IR calling while at the same time maintaining sufficient sensitivity to avoid too many false negative events. The following criteria were used for quantifying the number of IR events in a sample:

The number of IR events in a sample was determined based on introns with an IR ratio > 0.1 and meeting the filtering criteria described above. Introns not meeting these criteria were not considered as being retained. Beta regression was used to identify differentially retained introns (dIR) between cancer and adjacent normal tissues using the betareg R package [22]. Since IR ratios are proportional data with values between 0 and 1, we reasoned that beta regression was best suited to model IR and identify dIRs between normal and cancer tissues. An absolute difference in the IR ratio (ΔIR = IR_Cancer − IR_Normal) of more than 0.1 with FDR-adjusted p < 0.05 was considered significant.

Dimensionality reduction, i.e. principal component analysis (PCA), of IR profiles was performed using the package factoextra (github.com/kassambara/factoextra).

Differential gene expression between normal breast tissue and BrCa was performed using the DESeq2 package (RRID:SCR_000154) [23]. Genes with an average read count > 10 in all samples were selected for differential gene expression analysis (n = 23,072). Genes with an absolute log2 fold change > 1 and FDR-adjusted p < 0.05 were considered significant. To identify genes that were specifically differentially expressed in BrCa, we removed genes that were differentially expressed in any of the other 8 cancers and determined specificity by computing the z-score on log fold change using the log fold change observed in BrCa as reference.

Gene Ontology analysis was performed using the clusterProfiler package (RRID:SCR_016884) [24]. The false discovery rate (FDR) approach was used for multiple testing correction. The list of 1542 RBPs was taken from Gerstberger et al. [25].

Patient survival data were provided by the TCGA consortium. Survival analysis was performed using packages Surv and survminer (github.com/kassambara/survminer).

RNA-binding protein (RBP) motifs in position weight matrix format (PWM) were retrieved from the ATtRACT database (version 0.99β) [26], which contains 1196 motifs corresponding to 160 human RBPs. Sequences of 100 nt were extracted from the regions flanking retained and non-retained introns and scanned for the presence of motifs using the fimo tool provided by the meme suite [27].

To compare IR profiles across human cancers, we retrieved transcriptomics data for nine different solid tumours and matched adjacent normal tissues from The Cancer Genome Atlas (TCGA; Fig. 1A) and quantified IR using the IRFinder algorithm, which we have previously validated [17]. Overall, we identified a total of 11,943 unique IR events, of which 917 were shared among all nine cancers analysed.

Our analyses confirmed a previous report that BrCa is the only cancer in which the number of IR events is reduced compared to normal adjacent tissue (Fig. 1B) [28]. All other cancers exhibit increased IR compared to their matched adjacent normal tissue (Fig. 1B). However, we also noticed that the number of IR events in breast cancer itself was comparable with other cancers, while normal breast tissue presented with unusually high numbers of IR events (Fig. 1B). In fact, normal breast tissue had the highest IR frequencies compared to all other normal tissues (Fig. 1C). Therefore, reduced IR in tumours, which is unique to BrCa, can be largely attributed to normal breast tissue having the highest occurrence of IR events.

We applied beta regression models to identify differentially retained introns (dIRs) and found 3024 dIRs between normal and breast cancer (Additional file 1: Fig. S1A). Of these 210 were downregulated (in 160 genes) and 69 were upregulated (in 52 genes) with a ≥ 10% difference in the IR ratio (ΔIR ≥ 0.1). Downregulated IR events in BrCa are associated with processes related to cell cycle, nuclear division, and DNA replication among others (Additional file 1: Fig. S1B).

Next, we explored whether the specific pattern of IR in BrCa is associated with clinical features. As shown in Additional file 1: Fig. S2A and Fig. 1D, IR patterns were distinct between BrCa vs normal tissues as well as oestrogen receptors positive (ER⁺) vs negative (ER⁻) samples, respectively. The human epidermal growth factor receptor 2 (HER2) amplified molecular subtype had the lowest average number of IR events (n = 1731) compared to the other three main subtypes Luminal A (n = 2089), Luminal B (n = 2113), and Basal (n = 2018) (Fig. 1E). Strikingly, HER2-amplified tumours are associated with a 2.9-fold increased hazard ratio (p = 0.001, Additional file 1: Fig. S2B). Moreover, advanced stage tumours (Stage III) had the lowest average number of IR events (n = 1913) compared to Stage II (n = 2064) and Stage I (n = 2210) tumours (Fig. 1F). Likewise, those tumours with the highest immunohistochemical (IHC) staining score for HER2 (score: 3) exhibited the lowest average number of IR events (n = 1825) compared to score 1 (n = 2095) or score 2 (n = 2137) tumours (Additional file 1: Fig. S2C). We also found that a high number of IR events is associated with better survival in patients with the Luminal B subtype (Fig. 1G; Additional file 1: Fig. S2D). Though, Luminal B is the only subtype where high IR is associated with a survival advantage (Additional file 1: Fig. S2E). Interestingly, the trend is reversed (although not significant) in Her2 positive breast tumours that do not belong to the luminal B subtype.

To confirm that differences between tumour and normal breast tissue can be observed in vitro, we sequenced the transcriptomes of cultured ER⁺ MCF7 cells and non-tumorigenic MCF10A cells. Indeed, we observed a similar trend as in the TCGA cohort (Fig. 2A) and found that higher IR in the breast epithelial cell line (MCF10A) was associated with reduced gene expression (Additional file 1: Fig. S3). We also analysed sequencing data (Sequence Read Archive; SRA) of other cell lines representing each of the molecular subtypes (Additional file 1: Table S1). Comparing the number of IR events, we observed a similar trend as with the TCGA tumour samples, except for HER + HCC1419 cells, which have on average more IR events (though not significant) than cell lines of other subtypes (Additional file 1: Fig. S4).

To identify potential regulators of IR, we correlated IR frequencies in the TCGA-BRCA cohort with expression values (normalized RNA-seq counts) of ~ 23,000 genes (Additional file 2: Table S2). We performed Gene Ontology (GO) enrichment analysis on the top 5% genes with the highest (r > 0.41) and lowest (r < − 0.27) correlation coefficients to identify potential positive and negative regulators of IR, respectively. We found eight GO terms that were significantly enriched in positively correlated genes (p-adj. ≤ 0.05; Fig. 2B). Intriguingly, six of these GO terms were related to RNA splicing, with DDX39B, RBM25, PRPF39, and SRSF11 being among the genes that most highly correlate with the number of IR events in each sample (Fig. 2C).

The four genes that most strongly anti-correlate with the number of IR events include the mutase PGAM1, the membrane protein encoding SURF4, the mitochondrial transmembrane transporter SLC25A5, and the mitochondrial ATP Synthase F1 Subunit Beta (ATP5F1B) (Fig. 2D). Strikingly, the top 10 most significant GO terms (out of 399) associated with genes that negatively correlate with the number of IR events correspond to mitochondrial processes and cellular energetics (Additional file 1: Fig. S5).

Next, we investigated whether changes in IR frequencies are associated with changes to cellular states. Since the energy demand of a cell is tightly coupled with proliferation, we sought to examine a potential link between doubling times of BrCa cells and the number of IR events in 36 BrCa-related cell lines of the Cancer Cell Line Encyclopedia (CCLE). Indeed, we found that cells with a slower doubling time exhibit a higher number of IR events (Fig. 3A). While this correlation is fairly robust, it should be noted that CCLE doubling times are an error-prone surrogate for cancer cell proliferation.

To corroborate this result, we also investigated whether a common proliferation marker would inversely correlate with IR frequencies. Since immunohistochemistry (IHC) staining of Ki-67 is unavailable for the TCGA cohort, we tested whether the proliferation rate in tissues might be estimated based on MKI67 sequencing read counts. The MKI67 gene encodes the proliferation marker protein Ki-67. Using Human Protein Atlas data, we confirmed that MKI67 mRNA expression correlates with its Ki-67 protein staining intensities detected by IHC (Additional file 1: Fig. S6A). As expected, normalized MKI67 read counts were higher in all nine cancers when compared to the respective adjacent normal tissues (Additional file 1: Fig. S6B). This suggests that MKI67 read counts can be used as a proxy for IHC staining to estimate cellular proliferation rates. MKI67 expression was also found to be inversely correlated to the doubling time of 36 CCLE cell lines (Additional file 1: Fig. S6C). We observed that the number of IR events in samples of the TCGA-BRCA cohort negatively correlated with the proliferation rate (Fig. 3B). However, no correlation was observed when analysing normal and tumour samples separately (Additional file 1: Fig. S7).

Next, we investigated genes that were specifically deregulated in BrCa. We identified a set of 150 genes that were only differentially expressed between BrCa and normal breast tissues, of which seven were RBPs (Fig. 4A). We calculated the z-score of each gene’s log fold change in BrCa versus the log fold change in other cancers in order to estimate the level of specificity of a gene being differentially expressed in BrCa only (Fig. 4B). Among the genes that are highly specifically over-expressed in BrCa are two known RBPs: ZFP36L2 and TUT4. ZFP36L2 promotes poly(A) tail removal of mRNA transcripts [30], while terminal uridylyl transferase 4 (TUT4) adds uridines to deadenylated transcripts [31]. Thus, both RBPs are mediators of mRNA decay, which could explain the observed reduction of IR transcripts in BrCa.

We also determined the frequency by which BrCa-specific genes occur in RNA-related gene sets (n = 138) in the Molecular Signatures Database (MSigDB; total ~ 5000 curated gene sets) [32]. While known RBPs such as ZFP36L2 and SNRNP25 (part of the minor U12-type spliceosome) are annotated in multiple RNA-related gene sets, other genes, that are specifically differentially expressed in BrCa, did not show any potential RNA-binding capabilities (Additional file 1: Fig. S8A).

In addition, we analysed differentially retained introns for occurrences of RBP binding motifs. We found that differentially retained introns were enriched in NELFE and SRSF9 binding sites in upstream exons (5'-up) and the 3' terminal region, respectively (Fig. 4C). Moreover, retained introns have fewer HNRNPH1 and HNRNPK binding sites in their 3' terminal region compared to non-retained introns (Fig. 4C; Additional file 1: Fig. S8B).

We conclude that RPBs are among the factors that facilitate reduced IR in BrCa by enabling efficient splicing of introns from pre-mRNA transcripts. However, RBPs specifically differentially expressed in BrCa are not among those with enriched binding motifs within and around differentially retained introns. This suggests that more complex, multifaceted regulatory mechanisms are causing the consistent reduction of IR in BrCa.

Since the reduction in IR events in BrCa contrasts with all other cancer types analysed, we examined a possible contribution from the changing cell composition in the tumour microenvironment compared to healthy breast tissue. Gene signature-based and machine learning-based algorithms have been developed to deconvolute the cell-type composition in bulk RNA-sequencing data [33]. To compare cell environmental profiles of TCGA breast tumour samples versus healthy adjacent control samples, we used the cell-type deconvolution algorithm xCell [34], which was trained on 1,822 pure human cell-type-specific transcriptomes extracted from single-cell transcriptome profiling data. xCell analysis revealed that the breast tumour cell composition is distinct from other cancers (Additional file 1: Fig. S9). Among the most enriched cell types in BrCa are T helper cells, mesenchymal stem cells, and basophils. These predictions are supported by recent single-cell BrCa profiling studies [35‐37]. Normal breast tissue is enriched in endothelial cells, adipocytes, and dendritic cells (Fig. 5A).

Indeed, adipocyte and myeloid cell (M1 macrophages, basophils) enrichment are specific to normal breast tissue (Fig. 5B/C), which could explain the IR paradox in BrCa.

To determine whether cell types enriched in normal breast tissue have particularly high IR event frequencies, we retrieved RNA-sequencing data of 66 cell/tissue types from the ENCODE repository (Additional file 3: Table S3). Our analysis suggests that breast epithelial cells have the highest prevalence of IR followed by adipocytes (Fig. 5C), which could explain the drop in IR events in breast tumours.