Widespread intronic polyadenylation inactivates tumour suppressor genes in leukaemia

Abstract

DNA mutations are known cancer drivers. Here we investigated whether mRNA events that are upregulated in cancer can functionally mimic the outcome of genetic alterations. RNA sequencing or 3'-end sequencing techniques were applied to normal and malignant B cells from 59 patients with chronic lymphocytic leukaemia (CLL)^1-3. We discovered widespread upregulation of truncated mRNAs and proteins in primary CLL cells that were not generated by genetic alterations but instead occurred by intronic polyadenylation. Truncated mRNAs caused by intronic polyadenylation were recurrent (n = 330) and predominantly affected genes with tumour-suppressive functions. The truncated proteins generated by intronic polyadenylation often lack the tumour-suppressive functions of the corresponding full-length proteins (such as DICER and FOXN3), and several even acted in an oncogenic manner (such as CARD11, MGA and CHST11). In CLL, the inactivation of tumour-suppressor genes by aberrant mRNA processing is substantially more prevalent than the functional loss of such genes through genetic events. We further identified new candidate tumour-suppressor genes that are inactivated by intronic polyadenylation in leukaemia and by truncating DNA mutations in solid tumours^4,5. These genes are understudied in cancer, as their overall mutation rates are lower than those of well-known tumour-suppressor genes. Our findings show the need to go beyond genomic analyses in cancer diagnostics, as mRNA events that are silent at the DNA level are widespread contributors to cancer pathogenesis through the inactivation of tumour-suppressor genes.

Extended Data Figure 1.. Validation of IPA isoforms by independent methods and identification of CLL-IPAs used for further analysis.

(a) RNA-seq data were used to validate the presence of IPA isoforms using a GLM. Within two 100 nt windows (green bars) separated by 51 nt and located up- and downstream of the IPA peak the RNA-seq reads were counted. The IPA peak was considered validated if Padj < 0.1 (see methods). Out of N = 5,587 tested IPA isoforms, N = 1,662 were validated by this method. Shown is MGA as a representative example. (b) As only a fraction of IPA isoforms were validated by the method from (a), additional methods were used to obtain independent evidence for the presence of the IPA isoforms. Independent evidence was obtained using untemplated adenosines from RNA-seq data or through the presence of the IPA isoform in other 3′-seq protocols (10). As the majority of immune cell types used in this study have not been investigated using other 3′ end sequencing protocols and IPA isoform expression is cell type-specific (2), highly expressed IPA isoforms (>10 TPM) were not excluded from further analysis even if no read evidence was found by other protocols. (c) Hierarchical clustering based on IPA site usage separates the 3′-seq dataset into four groups. It separates CD5+B from CLL samples and clusters CLL samples into three different groups. Shown is the usage difference of the 20% most variable IPA isoforms across the data set (N = 342). Four of 13 CLL samples cluster away from the rest of the samples and are characterized by a high number of IPA isoforms (CLL high). (d) The GLM (FDR-adjusted P value < 0.1, IPA usage difference ≥ 0.05, IPA isoform expressed in CD5+B < 8 TPM) identified 477 recurrent (significantly upregulated in at least 2/13 CLL samples by 3′-seq) and 454 non-recurrent (significantly upregulated in 1/13 CLL samples by 3′-seq). IPAs were validated in an independent RNA-seq data set containing 46 new CLL samples. Among the recurrent IPAs, 71% of testable IPAs were verified using another GLM (see a). Among the non-recurrent IPAs, 64% of testable IPAs were verified. (e) Plotting the number of CLL-IPAs per sample separates the CLL samples investigated by 3′-seq into two groups: 4/13 samples generate a high number of CLL-IPAs (CLL high, median of CLL-IPAs/sample, N = 100, range, 42 – 274), whereas the rest of the samples generate lower numbers (CLL low, median, N = 9, range, 5 – 28). Center bar shows median and the error bars show the interquartile range. Two-sided Mann-Whitney test, **, P = 0.003.

Extended Data Figure 2.. The normal B cell counterpart of CLL cells are CD5+B cells derived from lymphoid tissue.

(a) Hierarchical clustering of normal human B cells (naive B (NB), memory B (MemB) and CD5+B) derived from lymphoid tissues or peripheral blood based on mRNA expression obtained from RNA-seq. The heatmap shows the 20% most variable genes across the data set (N = 1,887). The gene expression profiles of B cell subsets derived from peripheral blood or lymphoid tissue differ substantially, although the same markers were used for purification. (b) As in (a), but RNA-seq data from CLL samples were added to the analysis. The heatmap shows the 20% most variable genes across the data set (N = 2,078). CLL samples cluster with tissue-derived and not with blood-derived normal immune cells. (c) Number of all differentially expressed genes from the analysis shown in (b).

Extended Data Figure 3.. 3′-seq and RNA-seq tracks of functionally validated CLL-IPAs.

Five CLL-IPAs were functionally validated. Their 3′-seq and RNA-seq tracks are shown here and in Fig. 2a. Data are shown as in Fig. 1b. The corresponding RT-PCRs are shown in Extended Data Fig. 5a.

Extended Data Figure 4.. CLL-IPAs generate truncated mRNAs and proteins.

Gene models and western blots of 10 candidates depicted as in Figures 1b and 2a show that CLL B cells generate full-length and IPA-generated truncated proteins. BLCL were used as control B cells and were included in the 3′-seq tracks. ACTIN was used as loading control on the same blots. For gel source data see Supplementary Fig. 1.

Extended Data Figure 5.. Validation of the IPA-generated truncated mRNAs and validation of their stable expression over time.

(a) Detection of full-length and IPA-generated truncated mRNAs by RT-PCR in normal B cells (CD5+B, BLCL) and CLL cells used in the western blot validations shown in Fig. 2a and Extended Data Fig. 4. All experiments were performed twice with similar results. Primers to amplify the mRNA isoforms are located in the first and last exons shown in the gene models and are listed in Supplementary Table S3. HPRT was used as loading control. (b) Induction of truncated mRNAs and proteins through shRNA-mediated knock-down of splicing factors. All experiments were performed twice with similar results. U2AF1 was knocked-down in HeLa cells, U2AF2 was knocked-down in HEK293 cells and hnRNPC was knocked down in A549 cells. Shown as in (a), except for NUP96 which is shown as in Extended Data Fig. 4. NUP96 is derived from NUP98 precursor. Induction of DICER1 IPA by transfection of increasing amounts of anti-sense morpholinos (MO) directed against the 5′ splice site of intron 23 of DICER1 in HeLa cells. Shown are RT-PCRs. (c) RT-PCRs, performed once, on expression of full-length (FL) and IPA isoforms for eight CLL-IPAs in samples from two CLL patients and control B cells (CD5+B, BLCL). The samples were collected over a time interval of over 6 years. CLL11: T1, 17 months (mo) after diagnosis, T2, 24 mo, T3, 44 mo; CLL6: T1, 16 mo, T2, 49 mo, T3, 91 mo (42 mo after treatment). Samples from all time points (except CLL6, T3) were obtained from untreated patients. The primers for amplifications of the products were located in the first and last exons shown in the gene models and are listed in Supplementary Table S3. Expression of HPRT serves as loading control. The same gel picture of HPRT is shown in Fig. 3b for CLL samples and Extended Data Fig. 5a, far right panel, for BLCL and CD5+ control samples. All tested CLL-IPA isoforms were detectable at several time points during the course of the disease. Compared with CD5+B cells, expression of FCHSD2 IPA was not significantly upregulated in CLL. (d) Western blots of full-length and IPA-generated truncated proteins from CARD11, DICER, and SCAF4. All experiments were performed twice with similar results. ACTIN was used as loading control on the same blot. Shown are samples from normal B cells (BLCL) and two CLL patients, both at two different time points 0.5 – 10 months apart. For gel source data see Supplementary Fig. 1.

Extended Data Figure 6.. IPA-generated truncated proteins resemble the protein products of truncating DNA mutations and have cancer-promoting properties.

(a) CARD11 IPA results in translation of intronic nucleotides (grey) until an in-frame stop codon is encountered. This results in the generation of 16 new amino acids (grey) downstream of exon 10. In the case of MGA IPA three new amino acids downstream of exon 9 are generated. (b) Western blot showing that TMD8 cells express similar amounts of CARD11 IPA as CLL samples. The western blot is depicted as in Fig. 2a and was performed twice. ACTIN was used as loading control on the same blot. (c) Western blot (as in b) showing full-length CARD11 as well as CARD11 IPA in TMD8 cells expressing a control shRNA (Co), an shRNA that exclusively knocks-down the full-length protein (FL) and two different shRNAs that exclusively knock-down the CARD11 IPA isoform (IPA). The experiment was performed twice with similar results. GAPDH was used as loading control on the same blot. (d) Endogenous phospho-NF-κB-p65 levels were measured by FACS in TMD8 cells expressing the indicated shRNAs from (c). Mean fluorescent intensity (MFI) values are shown in parentheses in FACS plots of a representative experiment out of three. (e) Immunoprecipitation of V5-DICER or V5-DICER IPA from HEK293T cells using an anti-V5 antibody. The experiment was performed twice with similar results. 2.5% of input was loaded. (f) The extent of miRNA processing depends on the expression levels of full-length (FL) DICER, but not IPA. Shown are wild-type (WT) and DICER knock-out (KO) HCT116 cells. Re-expression of different amounts of FL DICER1 protein in the KO cells (measured by western blot of DICER1 in the top panel) results in different levels of endogenous let-7 expression (measured by northern blot in the bottom panel; compare lanes 3 and 4). Expression of DICER IPA has no influence on miRNA processing (compare lanes 4 and 5). ACTIN and U6 were used as loading controls on the same blots, respectively. The experiment was performed twice with similar results. (g) Western blot of MGA. MGA and MGA IPA were cloned and expressed in HEK293T cells to confirm the predicted protein size. The experiment was performed twice with similar results. Shown is also the endogenous MGA expression in Raji cells. ACTIN was used as loading control on the same blot. *, denotes an unspecific band. (h) Protein models of full-length and FOXN3 IPA are shown as in Fig. 2b. The IPA-generated protein truncates the fork-head domain and is predicted to lose the repressive activity. (i) As in (a), but for FOXN3. FOXN3 IPA generates 32 new amino acids downstream of exon 2. (j) FOXN3 IPA significantly de-represses expression of the oncogenic targets MYC and PIM2. Fold-change in mRNA level of endogenous genes in MEC1 B cells after transfection of GFP-FOXN3 IPA compared with transfection of full-length GFP-FOXN3. HPRT-normalized values are shown as boxplots (as in Fig. 1e) from N = 5 biologically independent experiments, each performed in technical triplicates. Two-sided t-test for independent samples was applied, **, P = 0.002. For gel source data see Supplementary Fig. 1.

Extended Data Figure 7.. Inactivation of TSGs by CLL-IPAs independently of DNA mutations.

(a) The distribution of full-length protein size of genes that generate CLL-IPAs (N = 306) and B-IPAs (N = 2,690) is shown in amino acids (aa). Boxplots as in Fig. 1e. Two-sided Mann-Whitney test, P = 0.87. (b) TR rate (ratio of TR mutations over all mutations) is shown for known TSGs obtained from (5). Boxplots as in Fig. 1e. Two-sided Mann-Whitney test, P = E-155. (c) Known TSGs, obtained from (5) that are targeted by CLL-IPAs (N = 21) are shown. Dark green bars indicate the fraction of retained CDR for each IPA-generated protein. Black dots indicate the hot spot positions of TR mutations obtained from MSK cbio portal. CLL-IPAs mostly occur upstream or within 10% (of overall aa length) of the mutations (two-sided Wilcoxon rank test, P = 0.04). (d) Contingency table for enrichment of TSGs among genes that generate CLL-IPAs. P value was obtained from two-sided Fisher’s exact test. TSGs were obtained from (5). (e) TSGs and genes that generate CLL-IPA isoforms have longer CDRs than genes that do not generate IPA isoforms. Boxplots as in Fig. 1e. Two-sided Kruskal-Wallis test, P = E-80. (f) Five control gene lists (N = 306, each) with a similar size distribution as CLL-IPAs and expressed in CLL were tested for enrichment of TSGs. Shown is the number of TSGs found. Chi-square-test did not show a significant enrichment of TSGs among the control genes. (g) Contingency table for enrichment of TR mutation genes in CLL among genes that generate CLL-IPAs. P value was obtained from two-sided Fisher’s exact test. (h) ZMYM5 is truncated by a TR mutation and an IPA isoform in the same patient, but the aberrations are predicted to result in different truncated proteins. A 10 bp deletion in exon 3 results in a frame-shift leading to the generation of a truncated ZMYM5 protein, whereas ZMYM5 IPA (not yet annotated) produces a truncated protein containing 352 more amino acids in the same patient. The genes shown in Extended Fig. 7h and 7i are the only genes with simultaneous presence of a TR mutation and CLL-IPA out of N = 268 tested. The position of the TR mutation is indicated in green. CLL7 and CLL11 3′-seq and RNA-seq tracks are shown for comparison reasons. (i) MGA is truncated by a TR mutation and an IPA isoform in the same patient. The TR mutation affects the 5′ splice site of intron 7, thus generating two additional amino acids downstream of exon 7, whereas the IPA isoform encodes a truncated MGA protein containing three more amino acids downstream of exon 9. Mutation and 3′-seq analysis were performed once. CLL7 and CLL11 are shown for comparison reasons. (j) Shown are additional recurrent (N > 1) DNA mutations found by exome sequencing of CLL patient samples stratified by a high or low number of CLL-IPAs per patient. Only the top and bottom 16 samples with high or low CLL-IPAs are shown to normalize the number of samples analyzed. This analysis is only descriptive and no test was performed. (k) Significant enrichment of SF3B1 mutations in the group of CLL samples with abundant CLL-IPA isoforms. Two-sided Mann-Whitney test was performed. (l) Abundance of CLL-IPAs is not associated with IGVH mutational status. Shown is the number of CLL-IPAs per sample for patients with mutated (MUT, N = 30) or unmutated (UN, N = 21) IGVH genes. Boxplots as in Fig. 1e. Two-sided Mann-Whitney test, P = 0.4.

Extended Data Figure 8.. Novel TSG candidates and validation of CHST11 IPA as cancer-promoting isoform.

(a) As in Fig. 3c, but shown are known (red gene names) and novel TSG candidates (black gene names) among the abundant CLL-IPAs. CLL-IPAs seem to inactivate these genes as they mostly occur upstream or within 10% (of overall aa length) of the mutations. Two-sided Wilcoxon rank sum test performed on all 136 TSGs, P = E-8; two-sided Wilcoxon rank sum test performed on the novel TSGs, N = 119, P = E-8. Position of TR mutation was determined using the data obtained from the MSK cbio portal and indicates the hot spot mutation. Right panel, the fraction of CLL samples affected represents the fraction of CLL samples (out of 59) with significant expression of the IPA isoform. Genes were included if they were affected in at least 20% of samples investigated either by 3′-seq or RNA-seq. (b) Contingency table for enrichment of novel TSGs among highly recurrent CLL-IPAs. P value was obtained from two-sided Fisher’s exact test. (c) TSGs have larger protein sizes. Boxplots as in Fig. 1e. Two-sided Mann-Whitney test, **, P = 0.005. The increased overall mutation rate of known TSGs correlates with larger protein size. Spearman′s correlation coefficient, r = 0.74, P = E-6. (d) CHST11 IPA generates 18 new amino acids (grey) downstream of exon 1. (e) Experimental set-up to measure paracrine WNT activity produced by MEC1 B cells either expressing GFP, GFP-CHST11 or GFP-CHST11 IPA and using a WNT reporter expressed in HEK293T cells. Primary CLL cells and the CLL cell line MEC1 express several WNTs, including WNT5B. In the presence of CHST11 WNT (red dots) binds to sulfated proteins on the surface of WNT producing cells, whereas WNT is secreted into the media in the presence of CHST11 IPA. WNT-conditioned media activates a WNT reporter in HEK293T cells. This set-up refers to Fig. 4f and 4g. (f) Western blot, performed once, for WNT5 shown as in Fig. 4f, but including HeLa cells as positive control for WNT5 expression. ACTIN was used as loading control on the same blot.

Extended Data Figure 9.. Cancer-upregulated IPA isoforms are also detected in breast cancer and T-ALL.

(a) MAGI3 is a TSG that is preferentially targeted by IPA in breast cancer (27). Shown is the mutation profile obtained from MSK cbio portal. (b) Expression of IPA isoforms in T-ALL detected by RNA-seq. Shown are 3′-seq and RNA-seq tracks of a representative mRNA (out of N = 101) from CLL samples, T-ALL samples and normal thymus. The T-ALL RNA-seq data were obtained from (32). We detected N = 381 IPA isoforms in at least one T-ALL sample, N = 133 in at least one thymus sample, N = 104 in at least one T-ALL and one thymus sample, and N = 101 in at least two T-ALL samples, but not in any of the thymus samples.

Figure 1.. Hundreds of genes generate recurrent CLL-IPAs.

(a) Schematic showing full-length mRNA and protein expression in normal cells and generation of a truncated mRNA and protein through cancer-specific IPA, despite no difference in DNA sequence. Polyadenylation sites (pA) are shown in light green. Loss of essential protein domains (dark green boxes) through cancer-gained IPA may inactivate TSGs, thus contributing to cancer pathogenesis. (b) Representative CLL-IPAs (from N = 330) are shown. mRNA 3′ ends detected by 3′-seq are depicted as peaks whose height corresponds to transcript abundance shown in transcripts per million (TPM). The bottom panel shows RNA-seq reads and numbers correspond to read counts. MemB, memory B cells, NB, naïve B cells. Full-length and IPA-generated truncated proteins are depicted in grey, known domains are shown in green and the domains lost through IPA are named. CC, coil-coil. For CLL-IPA, the number of retained and novel amino acids (aa) and aa of full-length proteins are given. (c) Representative RNA-seq tracks from two independent CLL data sets are shown as in (b), one is indicated by ′L′. Zoom-in shows the exonized part of intron 23 of DICER1 (green). (d) Difference in relative abundance (usage) of IPA isoforms between CLL and normal CD5+B cells. A GLM was used to identify significant events. CLL-IPAs with significantly higher usage are shown in red (FDR-adjusted P value < 0.1, usage difference ≥ 0.05, TPM in CD5+B < 8) and CD5+B-IPAs are shown in blue. Grey, IPAs present in CLL and CD5+B cells without significantly different usage. (e) Number of CLL-IPAs per sample is shown as box plots, horizontal line, median; box, 25th through 75th percentile; error bars, range. CLL high, N = 21/59, median of CLL-IPAs/sample = 98 vs CLL low, N = 38/59, median = 29. Two-sided Mann-Whitney test, ***, P = 6E-10.

Figure 2.. IPA-generated truncated proteins resemble the protein products of truncating DNA mutations and have cancer-promoting properties.

(a) 3′-seq and RNA-seq data of functionally validated CLL-IPAs (N = 5) as in Fig. 1b. The remaining tracks are shown in Extended Data Fig. 3. Endogenous full-length proteins are detected by western blot analysis in CLL and normal B cells (BLCL), whereas IPA-generated truncated proteins (red arrows) are only present in primary CLL cells. ACTIN was used as loading control on the same blot. The experiment was replicated with similar results (CARD11, N = 4, DICER, N = 3, MGA, N = 2). For gel source data see Supplementary Fig. 1. *, indicates an unspecific band. (b) Protein models are shown as in Fig. 1b. The aa positions of recurrent TR mutations are shown in blue. (c) Endogenous phospho-NF-κB-p65 levels are shown as normalized mean fluorescent intensity (MFI) values as mean ± SD from N = 5 (shRNA Control (Co), shRNA1 full-length CARD11 (FL) or N = 6 (shRNA2/3 CARD11 IPA, N = 3, each) biologically independent experiments. Two-sided Kruskal-Wallis test, **, P = 0.002; P value of two-sided Mann-Whitney test was adjusted for multiple testing, *, Padj = 0.036. (d) miRNA cleavage assay, performed twice with similar results, showing processing of pre-let-7i into mature let-7i by V5-DICER. Mock, no protein was added. V5-DICER IPA shows a complete loss of function, but no dominant-negative activity. (e) qRT-PCR of endogenous MYC target genes after expression of full-length or MGA IPA in Raji cells. Shown are GAPDH-normalized values as mean ± SD from three biological replicates, each performed in technical triplicates. Two-sided t-test for independent samples, *, P < 0.05, **, P < E-3, NS, not significant. Exact P values are shown in Supplementary Fig. 1. MGA represses all MYC target genes. Binding sites, BS.

Figure 3.. TSGs are enriched among CLL-IPAs. CLL-IPAs and TR mutations in CLL target the same genes but in different patients.

(a) The fraction of retained coding region (CDR) is shown for genes that generate CLL-IPAs (N = 306, median fraction of retained CDR = 0.21; 112 aa) and B-IPAs (N = 2,690, median fraction of retained CDR = 0.45; 221 aa). ***, Two-sided Mann-Whitney test, P = E-16. Box plots as in Fig. 1e. (b) RT-PCRs on expression of full-length (FL) and IPA isoforms for two TSGs (DICER1, NUP98) in samples from two CLL patients that were collected over a time interval of several years. CLL11: T1, 17 months (mo) after diagnosis, T2, 24 mo, T3, 44 mo; CLL6: T1, 16 mo, T2, 49 mo, T3, 91 mo (42 mo after treatment). Shown are the exons that contain primers for amplifications of the products. BLCL serve as control cells. Expression of HPRT serves as loading control. (c) Genes that are targeted by TR mutations in CLL and CLL-IPAs are shown (N = 36). Dark green bars indicate the fraction of retained CDR for each IPA-generated protein. Black dots indicate the positions of TR mutations in CLL. CLL-IPAs occur mostly in the vicinity of TR mutations or upstream of them (two-sided Wilcoxon rank sum test, P = 0.004). Right panel, the fraction of CLL samples affected is shown for each gene and represents the fraction of CLL samples (out of 59) with significantly upregulated expression of the IPA isoform (CLL-IPA, grey; TR mutations, red).

Figure 4.. Novel TSG candidates are inactivated in CLL at the mRNA level and in solid tumors at the DNA level.

(a) Color-coded IPA usage for a subset of CLL-IPAs (97/199 of samples with significant expression of IPA in ≥ 20% of CLL samples). Gene names and number of affected CLL samples per CLL-IPA is indicated (blue bars, 3′-seq, green bars, RNA-seq). (b) Truncating mutation rates (number of TR mutations/all mutations) in solid tumors, obtained from the MSK cbio portal for genes that generate abundant CLL-IPAs, partially shown in (a). The bimodal distribution was separated at the local minimum (TR mutation/all mutations = 0.12, red line) into two gene groups: those rarely targeted by TR mutations and those with high TR mutation rates in solid cancers, defined as novel TSG candidates. (c) TR mutation rates of known and novel TSG candidates. Two-sided Mann-Whitney test, **, P = 0.0002. Box plots as in Fig. 1e. (d) As in (c), but for overall mutation rates. Two-sided Mann-Whitney test, ***, P = E-10. (e) CHST11 protein models as in Fig. 2b. Loops depict membrane domains. A chromosomal translocation in CLL results in fusion of the immunoglobulin heavy chain locus (IGH) with a truncated CHST11 (23). (f) Western blot of WNT5B, performed once, shown as in Fig. 2a, from cell lysates or conditioned media (CM) of B cells stably expressing GFP, GFP-CHST11 or GFP-CHST11 IPA. CM from cells expressing CHST11 IPA contains unglycosylated WNT5B (25). (g) CM from samples described in (f) was added to HEK293T cells expressing a WNT reporter. Shown is normalized luciferase activity as mean ± SD from N = 7 biologically independent experiments. Two-sided Kruskal-Wallis test: **, P = 0.002; P value of two-sided Mann-Whitney test was adjusted for multiple testing, **, Padj = 0.002.