Principles Governing A-to-I RNA Editing in the Breast Cancer Transcriptome

Abstract

Little is known about how RNA editing operates in cancer. Transcriptome analysis of 68 normal and cancerous breast tissues revealed that the editing enzyme ADAR acts uniformly, on the same loci, across tissues. In controlled ADAR expression experiments, the editing frequency increased at all loci with ADAR expression levels according to the logistic model. Loci-specific "editabilities," i.e., propensities to be edited by ADAR, were quantifiable by fitting the logistic function to dose-response data. The editing frequency was increased in tumor cells in comparison to normal controls. Type I interferon response and ADAR DNA copy number together explained 53% of ADAR expression variance in breast cancers. ADAR silencing using small hairpin RNA lentivirus transduction in breast cancer cell lines led to less cell proliferation and more apoptosis. A-to-I editing is a pervasive, yet reproducible, source of variation that is globally controlled by 1q amplification and inflammation, both of which are highly prevalent among human cancers.

Graphical abstract

Figure 1

Detection of A-to-I Editing (A) Substitution frequencies of RDDs. (B) Percentage of RDDs confirmed in the validation data set, n = 15 BCs (in red), and the DARNED database (in blue). The negative control set is composed of 1,000 sites selected at random positions in randomly selected Alu regions. Sites in immunoglobulin (Ig) hyper-variable regions were excluded; see the Supplemental Experimental Procedures. (C) Each dot represents a sample for which the frequency of edited AZIN1 transcripts has been measured with Illumina full transcriptome sequencing (x axis) and Roche FLX amplicon sequencing (y axis). ρ denotes the Spearman’s correlation. (D) Distribution of the 560 edited sites into functional categories. (E and F) Number of detected Alu A-to-I sites as a function of transcriptome and exome coverages, respectively. Green dots represent tumor-matched normal samples.

Figure 2

A-to-I Editing and ADAR Expression in Normal and Tumor Breast Tissue (A) Each dot represents a patient with the mean editing frequency in her normal (x axis) and her matched tumor breast tissue (y axis). (B) Same as (A), except that the AZIN1 editing frequency measured by Roche FLX amplicon sequencing is depicted. (C) The mean editing frequency of eight breast organoid cultures is compared to that of 15 breast tumors. (D) Representative ADAR staining of a luminal A tumor. (E) Zooming in (D) reveals that tumor staining (black arrows) is higher than in normal epithelium (green arrows) and lymphocytes (red arrows). (F) Zooming further in (E) reveals a higher staining of nucleoli (black arrows).

Figure 3

Model of A-to-I Editing (A) Each dot represents a sample with its RNA-seq-estimated ADAR expression on the x axis (in log₂ of fragments per kilobase per million mapped reads), and its mean editing frequency across all 560 Alu sites on the y axis. Green dots represent tumor-matched normal samples. The RNA-seq expression of ADAR is highly correlated with microarrays and qRT-PCR expression (Figure S2). (B) Each dot represents an Alu A-to-I editing site with the maximal edit frequency across all samples on the x axis and the number of samples in which it was detectably edited on the y axis. (C) Heatmap of editing frequencies across all Alu A-to-I edit sites in all samples. Both are ordered by increasing (down-to-up, left-to-right) mean editing frequencies. Smoothed contour lines labels give the percentage of edited transcripts. The bottom panel shows corresponding ADAR expression. Green dots represent tumor-matched normal samples. Negative controls are presented in Figure S2. (D) Model of A-to-I editing. Turning the ADAR “expression knob” clockwise increases ADAR expression. As a result, more transcripts are edited (red dots), and the editing frequency of all editable sites increases accordingly (compare green versus red bars). Moreover, the detection limit at some sites for which editing was previously undetectable is passed. The detection limit depends on sequencing coverage, which is lower on the right-most exon. Importantly, the ranking of editing frequencies of the different sites is unaltered by ADAR expression.

Figure 4

Validation of the A-to-I Editing Model and Quantitative Estimation of Site-Specific Editability (A) Effect of increasing ADAR expression in the cell line MCF7 on editing in four representative Alu regions and AZIN1. The full-length of sequenced regions are shown in Figure S3 for MCF7 and three more cell lines. Complete ADAR western blots quantifications underlying the color scale are provided in Figure 5E (see baseline t = 0 and IFN-α, t ∈ {1, 2, 5} days tracks) and in Figure S7. Increasing ADAR expression increases the editing frequency at all editable positions, as predicted by the model of Figure 3D. Similar results were obtained for IFN-β and IFN-γ (global, position-less, view Figure 5F). Arrows point at editing sites detectable only at higher ADAR expression in our assay. (B) Increasing sequencing coverage (x axis) or ADAR expression (color scale) increases the number of detectable editing sites (y axis). Coverage variation was implemented by down-sampling the total pool of sequencing reads, starting from 2,000×, down to 100×, and re-running the variant detection pipeline for each down-sampled alignment. Each data point is the mean of 30 down-sampling experiments. Error bars, SD. (C) Editing of individual mRNA molecules. Each black dot depicts an edited base in a given mRNA molecule. The y axis goes from 0 to 60 and corresponds to the adenosines in the ∼250-bp span that are edited in at least one of the 2,842 reads represented along on the x axis. Reads and adenosines were ordered by decreasing editing frequencies. 185 non-edited reads were omitted from the figure. (D) Dose-response curves for experiment in cell line BT474. ADAR was increased through IFN-α stimulation (as in A). We focused on 81 sites (color lines) with a baseline editing frequency >2.5% in order to avoid trivial nonlinear effects caused by lack of detection at low ADAR expression. (E) Example of a fit of the logistic model (line) to experimental data points (dots). The unit of ω is commensurate to the dimensionless ADAR relative expression and ε is the fraction of edited transcripts at saturation. (F and G) Distributions of ε and ω across the 81 sites. (H) The 81 edited sites are depicted as dots with the corresponding ε_i estimates derived from the BT474 cell lines on the x axis and their in vivo editing frequency on the y axis. (D)–(H) are part of a more comprehensive analysis presented in Figure S4. (I) DNA-based statistical model of editability. The model included three parameters: (1) the best Smith-Waterman global alignment score of the 51-bp sequence surrounding the editing site (green dot) within the 2,501-bp sequence surrounding the editing site on the reverse strand; (2) the distance separating the editing site from this best alignment; (3) the 20 nucleotides surrounding the editing site. These 1 + 1 + 20 = 22 variables were fitted with a linear model against the editing frequencies of half of 51,621 Alu editing sites with coverage ≥20× previously identified (Ramaswami et al., 2012). (J) Observed editing frequencies versus editabilities predicted from DNA for validation sites.

Figure 5

ADAR Amplification and the Interferon Response Are Independent Predictors of ADAR Expression in Cancer (A) The top panel shows the frequencies of amplifications/deletions along chromosome 1 in our series. The middle panel shows the genes whose expression is highly associated with that of ADAR. Nineteen genes not located on chr1 are omitted. The bottom panel shows the Spearman’s correlation coefficient and associated p values of non-segmented copy-number array probes with the sample-wise mean editing frequencies. (B) Dots represent tumor samples, with STAT1 expression on the x axis and ADAR expression on the y axis. (C) Same as (B) with ADAR expression adjusted for ADAR copy number. (D) Association p values of ADAR copy number and STAT1 expression with ADAR expression increase in a multivariate analysis, demonstrating that ADAR expression is independently associated with these two variables. (E) Seven breast cancer cell lines were exposed to interferon α, β, and γ for 1, 2, and 5 days. Western blots quantifications are depicted for each cell line, interferon, and time. Because expression dynamic ranges vary among cell lines, each line has its own color scale extending from low expression in green to high expression in red. The underlying gels are presented in Figure S7 and blot quantification in Table S6. Corresponding mRNA RT-PCR expression data are shown Figure S7 and detailed Table S7. (F) Editing frequencies in the absence of treatment (x axis) versus interferon treatment (y axis). Points depict the editing sites in AZIN1 and the four Alu regions of Figure S3. Points are above the identity line x = y (black diagonals); i.e., interferons increase editing frequencies at all sites. Library preparation failed for MCF7/IFN-γ at 5 days. Limited sequencing coverage precluded detection of some editing events for MDA-MB-231, t = 0 and t = 1 days.

Figure 6

ADAR Involvement in Cell Proliferation and Apoptosis (A) Western blot analysis of ADAR silencing after shRNA lentiviral transduction in MDA-MB-231, MCF7, and BT474 breast cancer cell lines. (B) ADAR silencing statistically decreases cell proliferation. Cell growth curves for ADAR-knockdown cells (shRNA ADAR) and control cells (shRNA control) in MDA-MB-231, MCF7, and BT474 BC cell lines. (C) ADAR silencing statistically increases cell apoptosis. Illustration of the percentage of apoptotic cells in ADAR-knockdown cells (shRNA ADAR) and control cells (shRNA control) in MDA-MB-231, MCF7, and BT474 BC cell lines. Error bars depict SDs of three independent experiments.

Figure 7

ADAR Amplification and the Interferon Response Predict ADAR Expression in Human Cancers We included all TCGA data sets and tumors (see “N” column) for which both copy-number and RNA-seq expression data (pipeline v.3) were available. Data sets are ordered by decreasing median ADAR expression (top to bottom). The three leftmost plots depict the distributions of ADAR expression, ADAR DNA copy number, and STAT1 expression across each data set. The two rightmost bar plots extend to TCGA data the calculation presented for our data in Figures 5B and 5C. In most cancers, adjusting ADAR expression for ADAR copy number increases the Spearman correlation, ρ, with STAT1 (cf. the dark blue bars to the light blue bars).