3′ UTR shortening represses tumor-suppressor genes in trans by disrupting ceRNA crosstalk

Abstract

Widespread mRNA 3′ UTR shortening through alternative polyadenylation¹ promotes tumor growth in vivo². A prevailing hypothesis is that it induces proto-oncogene expression in cis through escaping microRNA-mediated repression. Here we report a surprising enrichment of 3′UTR shortening among transcripts that are predicted to act as competing-endogenous RNAs (ceRNAs) for tumor-suppressor genes. Our model-based analysis of the trans effect of 3′ UTR shortening (MAT3UTR) reveals a significant role in altering ceRNA expression. MAT3UTR predicts many trans-targets of 3′ UTR shortening, including PTEN, a crucial tumor-suppressor gene³ involved in ceRNA crosstalk⁴ with nine 3′UTR-shortening genes, including EPS15 and NFIA. Knockdown of NUDT21, a master 3′ UTR-shortening regulator², represses tumor-suppressor genes such as PHF6 and LARP1 in trans in a miRNA-dependent manner. Together, the results of our analysis suggest a major role of 3′ UTR shortening in repressing tumor-suppressor genes in trans by disrupting ceRNA cross-talk, rather than inducing proto-oncogenes in cis.

Widespread 3′ UTR shortening (3′US) through alternative polyadenylation (APA) occurs during enhanced cellular proliferation and transformation^1,5–8. Recently, we reported that NUDT21-mediated 3′US promotes glioblastoma growth, further underscoring its significance to tumorigenesis². A prevailing hypothesis is that a shortened 3′ UTR results in activation of proto-oncogenes in cis through escaping microRNA (miRNA)-mediated repression. Indeed, several well-characterized oncogenes, such as CCND1, have been shown to use 3′US to increase their protein levels, but mostly in cell lines⁵. However, in recent PolyA sequencing⁷ and our TCGA RNA sequencing (RNA-Seq) APA analysis¹ (5 and 358 tumor/normal pairs, respectively), most oncogenes with 3′US previously identified in vitro⁵ displayed almost no changes in their 3′ UTR lengths in tumors (Fig. 1a). For example, we identified 1,346 recurrent (occurrence rate >20%) 3′US genes in 358 tumor/normal pairs¹. However, CCND1 is not on that list as its 3′ US occurred in only a very small portion (8 out of 358; 2.2%) of tumors (Fig. 1b). Furthermore, similar to random genes, 3′US genes from all 5 previous APA studies have little overlap with the top 500 (P < 0.01) high-confidence oncogenes as defined on the basis of distinct somatic mutational patterns of >8,200 tumor/normal pairs⁹ (Fig. 1c). These results challenge the previous hypothesis and suggest a different function of 3′ US for tumorigenesis.

Fig. 1 |. 3′ US genes are not strongly associated with oncogenes.

a, TCGA RNA-Seq data for CCND1 demonstrates no change in 3′ UTR usage between tumors (yellow) and matched normal samples (blue). A similar pattern was also observed in PolyA-seq⁷ of CCND1. b, ΔPDUI values for 3′ US genes (red) and all genes (gray) in 358 TCGA tumor/normal pairs¹ (upper panel). A negative ΔPDUI represents 3′ UTR shortening. The lower panel shows ΔPDUI values for CCND1 across 358 tumor/normal pairs¹. Significant CCND1 3′ UTR shortening occurred only in a very small portion (8 out of 358; 2.2%) of tumors. c, Overlap P values and the ratios between previously identified 3′ US genes and oncogenes. ‘Random (n = 100)’ represents the averaged P value from 100 random sampling of 100 RefSeq genes. The error bar represents standard variation values of P values from 100 random trials.

Aside from regulating its cognate transcript in cis, the 3′ UTR has also been implicated in competing-endogenous RNA (ceRNA) regulation in trans¹⁰. Although the scope is not fully understood, ceRNA is generally thought to form global regulatory networks (ceRNETs) controlling important biological processes¹¹. For example, the tumor suppressor PTEN’s ceRNAs, CNOT6L and VAPA, have been shown to regulate PTEN and phenocopy its tumor-suppressive properties¹². As the ceRNA’s regulatory axis is mostly based on miRNA-binding sites on 3′ UTRs, we hypothesize that when genes with shortened 3′ UTRs no longer sequester miRNAs, the released miRNAs would then be directed to repress their ceRNA partners, such as tumor-suppressor genes, in trans, thereby contributing to tumorigenesis.

To test this hypothesis, we first used well-established strategies to reconstruct two ceRNETs from 97 TCGA breast tumors and their matched normal tissues, respectively, based on miRNA-binding-site overlap and co-expression^13,14 between genes of active ceRNA regulation (Methods). In general, transcripts are less correlated between each other in tumors than in normal tissues, partially due to tumor heterogeneity¹⁵ and global reduction of miRNA expression in tumors¹⁶ (Fig. 2a). As expected, the loss of co-expression results in a much smaller (tenfold reduced) ceRNET for tumors than for normal tissues (Fig. 2b).

Fig. 2 |. 3′UTR shortening contributes to ceRNET disruption.

a, Pearson’s correlation coefficients of 100,000 randomly selected transcript pairs with significant miRNA-binding-site overlap in breast tumors and matched normal tissues. b, The number of ceRNA pairs in breast tumor and the matched normal ceRNETs. The numbers in parentheses are normalized to the number of edges shared between tumor and normal. c, Gene expression of EPS15 (3′ US gene) and PTEN (ceRNA partner) on 68 estrogen-receptor-positive (ERP) breast tumors and matched normal samples. The horizontal lines represent the mean expression values of PTEN, which is decreased in tumors (FDR = 2.1 × 10⁻¹⁰). d, The upper heatmap exhibits significant APA events (rows) across 68 ERP tumor/normal pairs (columns), ranked by the number of 3′ US genes. The Venn diagrams show the number of ceRNA pairs in the normal and tumor ceRNET. The numbers in parentheses are normalized to the number of edges shared between tumor and normal tissues. The P value was calculated from a one-tailed Pearson’s chi-squared test.

To investigate the role of 3′ US in ceRNET disruption, we focused on estrogen-receptor-positive (ER⁺) breast tumors, which comprise the majority (68/97) of TCGA breast tumor samples. We built normal and tumor ceRNETs using the same procedure as above. Using the DaPars algorithm¹, we identified 427 3′ US genes recurring in >20% of tumors. Close inspection indicates that 3′ US is associated with ceRNET disruption. For example, we identified PTEN and EPS15 as a ceRNA pair in normal ceRNET (4 miRNA-binding-site overlap and ρ = 0.63 co-expression). However, since EPS15 underwent 3′ US in 23 (33.8%) out of 68 tumors, thereby losing its capability to compete with PTEN for miRNAs, it lost (ρ = 0.32) the co-expression (and ceRNA) relationship with PTEN in tumors (Fig. 2c). Globally, the top 100 ceRNAs with the most significant 3′US genes all lost their interactions in tumors, while 12 out of 100 ceRNAs lacking 3′ US retained (P = 0.0002) their interactions. Furthermore, in separate ceRNETs from 30 tumor/normal pairs with the least and most amount of 3′ US (upper panel in Fig. 2d), more 3′ US is clearly associated with more ceRNET loss (38.6 versus 16.4 in fold decrease, P < 1 × 10⁻¹⁶, lower panel in Fig. 2d). From these findings, we conclude that 3′ US is strongly associated with ceRNA network disruption in tumors.

To understand the function of 3′ US-mediated ceRNET disruption, we selected 381 3′ US genes and 2,131 of their ceRNA partner genes (3′ US ceRNAs), including 591 3′ US ceRNA hub and 1,540 3′ US ceRNA non-hub genes, in the normal ceRNET (Supplementary Table 1, Methods). We hypothesized that 3′ US genes released their miRNAs to repress their ceRNA partners in trans. Consistent with our hypothesis, expression changes of 2,131 3′ US ceRNA genes in tumors are anti-correlated (r=−0.21; P = 5 × 10⁻²⁴) with the degree of 3′ US of the associated 3′ US genes (Supplementary Fig. 1a). As a result, among 976 genes in normal ceRNET downregulated in tumors, 816 (83.6%) are ceRNAs of 3′ US genes. Surprisingly, 3′ US ceRNA hub genes are enriched in tumor-suppressor genes (P ~ 1 × 10⁻²⁰) but not in oncogenes (Fig. 3a), suggesting that the 3′ US represses tumor suppressors in trans. For example, 3′ US of EPS15 would contribute to downregulating its ceRNA partner PTEN in tumors (Fig. 2c). Globally, 160 expressed tumor-suppressor genes from 3′ US ceRNAs are more likely downregulated than 226 control tumor-suppressor genes not in ceRNET (P = 8 × 10⁻³, Fig. 3b), indicating a significant association between 3′ US and tumor-suppressor gene repression.

Fig. 3 |. 3′ UTR shortening represses tumor-suppressor genes in TCGA breast cancer.

a, Functional enrichment of 3′ US ceRNA hub genes (red), 3′ US ceRNA non-hub genes (blue), 3′ US genes (purple) and random RefSeq genes (gray). We randomly sampled each gene category to the same number (381) 100 times; averaged P values with standard deviation are plotted. b, Relative expression (tumor/normal) of tumor-suppressor genes that are 3′US ceRNAs (n = 160, left box) is lower than for those that are not in ceRNET (n = 226, right box) (P = 8 × 10⁻³). c, 3′ US genes (left) might repress their ceRNA partner PTEN in trans through miRNAs (middle) commonly released through 3′ UTR shortening. d, A heatmap in the top panel showing APA events for the nine 3′ US genes (rows). The boxplots in the bottom panel show PTEN expression levels in 10 tumors with the most (left) or least (right) 3′ UTR shortening. e, Western blot analysis of lysates from MCF7 cells treated with control (Con.) or EPS15-targeting siRNAs. The image is representative of three independent experiments. f, Quantification of luciferase activity from cell lysates derived from MCF7 cells transfected with a luciferase reporter containing the PTEN 3′ UTR and EPS15-targeting siRNAS. Data are the average luciferase activity ± standard deviation from three independent experiments (P = 0.011 and P < 0.001, two-sided t-test). g, PTEN 3′ UTR luciferase reporter activity in MCF7 cells transfected with EPS15- and DICER-targeting siRNAS. Data are the average luciferase activity ± standard deviation from three independent experiments (P = 0.045, P = 0.003 and P = 0.645, two-sided t-test). h, Indirect immunofluorescence of MCF7 cells transfected with either a heterologous reporter containing a vector-derived 3′ UTR (Con.) or the EPS15 3′ UTR together with a GFP construct. PTEN was detected by anti-PTEN antibody conjugated with Alexa Fluor-594. The arrows highlight PTEN⁺ transfected cells. A representative image is shown from three independent experiments. Scale bar, 20 μM. i, The number of PTEN-positive cells in the transfected cells with either the EPS15 3′ UTR (n = 335) or the control 3′ UTR (n = 357) from three images.

Additional analyses on sequence features partially explain why 3′ US genes, but not tumor suppressors in their ceRNA partners, are likely to have alternative proximal polyadenylation sites, leading to 3′ US (Supplementary Note). We have also analyzed TCGA 450K methylation array data and found that the 3′ US-mediated ceRNA repression is independent of promoter hypermethylation (Supplementary Note).

To better quantify the trans effects of 3′ US, we developed a mathematical model (MAT3UTR) based on its 3′ US gene(s) expression, 3′ US level, miRNA-binding site(s) and miRNA expression(s) (Methods). In 1,548 differentially expressed 3′ US ceRNAs, MAT3UTR can explain 47.6% of variation in gene expression (Supplementary Fig. 3c). In contrast, the MAT3UTR-control model, which considers miRNA expression but not 3′ US, explains only 27.2% of variation (Supplementary Fig. 3d), consistent with previous reports17 that miRNA alone has a weak role in regulating gene expression. The results suggest that the trans effects of 3′ US plays a major role in regulating ceRNA gene expression.

MAT3UTR predicts many trans-target genes of 3′ US, including PTEN, in ceRNA crosstalk^11–13 (top 1% MAT3UTR score, Supplementary Table 2). In normal ceRNET, PTEN is predicted to be a ceRNA of nine 3′ US genes (Fig. 3c). When we ranked 97 breast tumor/normal pairs by the amount of 3′ US across these nine genes (upper panel in Fig. 3d), tumors with more 3′ US showed more down-regulation of PTEN (P = 0.03, lower panel in Fig. 3d). Furthermore, MAT3UTR can explain 86.9% of the variation in PTEN’s expression across tumors (Supplementary Fig. 3g), suggesting that the trans effects of 3′ US play a major role in downregulating PTEN

To empirically test the hypothesis that 3′ US can downregulate PTEN in trans, we focused on EPS15 among the nine 3′ US genes(Methods). We observed that depletion of EPS15 by siRNA in MCF7 cells reduces PTEN expression (Fig. 3e). To ascertain whether this effect depends on miRNA-based targeting of the PTEN 3′ UTR, we used a luciferase reporter vector with the PTEN 3′ UTR (pLight-Switch-PTEN 3′ UTR) to test the effect of EPS15 knockdown on its expression. We observed that reduction of EPS15 reduces PTEN 3′ UTR luciferase activity (Fig. 3f). To further understand whether the crosstalk is miRNA-dependent, we depleted DICER1 to abolish miRNA biogenesis and found that loss of DICER1 can relieve the influence of EPS15 knockdown on PTEN 3′ UTR expression (Fig. 3g). Finally, overexpression of the EPS15 3′ UTR increased the number of PTEN-positive cells (Fig. 3h,i). Thus, EPS15 3′ US may impact PTEN expression.

To gain insights into the global cause-and-effect relationship between 3′ US and the repression of tumor-suppressor genes, we revisited our previous data from NUDT21-knockdown HeLa cells, since NUDT21 is one of the master regulators of 3′ US². We identified 1,168 3′ US ceRNAs in NUDT21-knockdown cells solely on the basis of significant miRNA-binding-site overlap with 1,450 3′ US genes, since co-expression cannot be effectively estimated from two replicates of our experiments. With 9,914 expressed RefSeq genes with no significant miRNA-binding-site overlap with 3′ US genes as random controls, the tumor-suppressor genes remain strongly enriched in 3′ US ceRNAs (P ~ 1 × 10⁻³⁸, Fig. 4a). Among 57 tumor-suppressor genes in 3′ US ceRNAs, 33 (57.9%) showed repression in NUDT21-knockdown samples; whereas a smaller portion (44.5%) of 339 control tumor-suppressor genes showed repression (P ~ 0.03, Fig. 4b), suggesting that NUDT21-mediated 3′ US represses tumor-suppressor genes in trans. In spite of potentially higher false positives due to lack of co-expression in ceRNA identification, these results are highly consistent with our observations in TCGA breast cancer. On the basis of these results, we posit that repression of tumor-suppressor ceRNAs would correlate with increased occupancy of AGO2 in the RISC complex. To formally test this hypothesis, we isolated cytoplasmic fractions from control or NUDT21-knockdown cells and conducted RNA immunoprecipitation (RIP) using anti-AGO2 antibodies. On average, we observed ~200-fold enrichment of ceRNAs in Ago2 RIP complexes relative to control IgG (Supplementary Fig. 4b). Reduced expression of NUDT21 does not impact AGO2/DICER1 expression and GAPDH messenger RNA binding to AGO2 (Fig. 4c,d and Supplementary Fig. 4b). Furthermore, we sequenced miRNAs from control and NUDT21-knockdown cells, and found that miRNAs are equally likely to be upregulated or downregulated (Supplementary Fig. 4d), ruling out a general effect on miRNA biogenesis. Importantly, we could detect increased association of multiple tumor-suppressor ceRNAs with AGO2 following NUDT21 depletion that ranged from 1.5-fold to nearly 7-fold (Fig. 4d). These results demonstrate that 3′ US can lead to reduction of tumor-suppressor genes through their increased association with repressive AGO2 complexes.

Fig. 4 |. NUDT21-mediated 3′ UTR shortening causes tumor-suppressor repression in trans.

a, Oncogene or tumor-suppressor gene enrichment of 3′ US ceRNAs (red), 3′ US genes (blue) and RefSeq genes (gray), in the NUDT21-knockdown (KD) experiment. We randomly sampled each gene category to the same number (n = 1,168) 100 times, and reported the averaged P values with standard deviation (error bar). b, Expression change of tumor-suppressor genes that are 3′US ceRNAs (n = 57, left box) or that are not connected to 3′US genes (n = 339, right box). 3′US ceRNA tumor-suppressor genes showed lower expression in NUDT21-knockdown samples (P = 0.03). c, Knockdown of NUDT21 in HeLa cells using CRISPR/Cas9 and reduced NUDT21 was detected by western blot analysis in three independent experiments. d, RIP was performed with anti-AGO2 antibody; normal mouse IgG served as a control. The RIP complexes were detected by western blot with a distinct AGO2 antibody from rat (inset). The indicated ceRNAs associated with AGO2 enrichment in NUDT21-knockdown cytoplasmic lysates versus the control are shown with average fold change ± standard deviation from three independent assays (P = 0.0002, P = 5.2 × 10⁻⁶, P = 0.0004, P = 0.0005, P = 0.0006, P = 0.01 and P = 5.47 × 10⁻⁷, two-sided t-test, **P < 0.001, *P < 0.01).

To further validate the miRNA-dependent, repressive trans effects of 3′ US, we monitored expression of the tumor-suppressor genes PHF6 and LARP1 and their ceRNA partners, YOD1 and LAMC1 (Supplementary Table 3). We consistently observed that PHF6 and LARP1 expression levels were decreased in NUDT21-knockdown cells while both YOD1 and LAMC1 expression levels were increased (Fig. 5a). To determine whether the 3′ UTR mediated this effect, we transfected luciferase reporters containing the 3′ UTR of either PHF6 or LARP1 into control or NUDT21-knockdown cells and measured luciferase activity. We found that both reporters were downregulated after NUDT21 knockdown (Fig. 5b). Both PHF6 and LARP1 have been shown as tumor-suppressor genes^9,18,19 and downregulation of PHF6 or LARP1 in HeLa cells increases cell growth, confirming their tumor suppressive activity (Supplementary Fig. 5).

Fig. 5 |. NUDT21-mediated 3′ UTR shortening represses the tumor-suppressor genes PHF6 and LARP1.

a, Western blot of 3′ US ceRNA tumor-suppressor genes (PHF6/LARP1) and 3′ US genes (LAMC1/YOD1) in NUDT21-knockdown cells. A representative image is shown from three independent experiments. b, Activity of the PHF6 3′ UTR and LARP1 3′ UTR luciferase reporter constructs in NUDT21-knockdown cells relative to control siRNA-transfected cells. The data are the average of luciferase activity ± standard deviation from three independent experiments (P = 0.037 and 0.05; P = 0.016 and 0.025, two-sided t-test). c, NUDT21 knockdown induces 3′ UTR shortening and upregulation of YOD1, allowing miR-3187–3p and miR-549 to repress PHF6. d, Western blot analysis using the indicated antibodies on lysates from HeLa cells transfected with siRNA for NUDT21 (si-NUDT21-4) and two antagomirs that block miR-549 and miR-3187–3p. The image is representative of two independent experiments. e, Activity of the PHF6 3′ UTR luciferase reporter construct in HeLa cells with the indicated siRNAs, miRNAs or antagomirs. The data are the average of luciferase activity ± standard deviation from three independent experiments (P = 0.004, P = 0.90, P = 0.018 and P = 0.015, two-sided t-test). f, Western blot analysis of cell lysates from cells transfected with either control siRNA or YOD1 siRNA. In the third lane, the cells were transfected with YOD1 siRNA and then transfected with YOD1 cDNA. The data are representative of three independent experiments. g, Activity of the PHF6 3′ UTR luciferase reporter in cells treated with the same experimental design as in f. The data are the average of luciferase activity ± standard deviation from three independent experiments (P = 0.016 and P = 0.01, two-sided t-test). h, Activity of the PHF6 luciferase reporter in cells transfected with the indicated siRNAs. The data are the average of luciferase activity ± standard deviation from three independent experiments (P = 0.025, P = 0.009 and P = 0.99, two-sided t-test).

To further investigate the mechanism of tumor-suppressor ceRNA downregulation, we chose PHF6 on the basis of MAT3UTR analysis and experimental results (Methods). We selected two miRNAs targeting PhF6 (Fig. 5c), which were released by 3′ US of YOD1 (miR-3187–3p as the highest and miR-549 as the sixth highest in terms of ^βmiR_jin equation (3); Methods and Supplementary Table 4). Neither of these miRNAs was found to change its expression following NUDT21 knockdown (Supplementary Fig. 4d). However, PHF6 expression was partially rescued by an antagomir blocking the activity of miR-549 and completely rescued by an antagomir targeting miR-3187–3p (Fig. 5d). Moreover, PHF6 3′ UTR-mediated luciferase activity was partially rescued by the miR-3187–3p antagomir or YOD1 siRNA (Fig. 5e). To understand whether reduced expression of PHF6 depends on YOD1 levels, we transfected YOD1 cDNA into cells depleted of YOD1 and found that re-expression of YOD1 could not restore either the expression of endogenous PHF6 (Fig. 5f) or the expression of the PHF6 3′ UTR-mediated luciferase (Fig. 5g), suggesting that the trans effect on PHF6 is due to the 3′ UTR of YOD1 Finally, to determine whether the crosstalk between PHF6 and YOD1 is miRNA-dependent, we also showed that depletion of DICER1 abolishes PHF6 and YOD1 crosstalk (Fig. 5h). Collectively, the data strongly suggest that NUDT21-mediated 3′ US causes tumor-suppressor gene repression in trans in a miRNA-dependent manner.

Although analyzing ceRNA crosstalk in light of 3′ US has been briefly suggested^20–22, our MAT3UTR analysis of 97 breast cancer RNA-Seq data followed by functional validation suggests a wide-spread causal role of 3′ US in repressing tumor-suppressor genes in trans. While the trans effect further emphasizes the importance of APA in tumor progression, it also provides an additional layer of gene regulation and underscores the need for further investigation into other potential mechanisms^23,24 that could per-turb ceRNA crosstalk, such as RNA editing and competition with RNA-binding proteins.

Methods

Methods, including statements of data availability and any associated accession codes and references, are available at https://doi.org/10.1038/s41588–018-0118–8.

Tumor-suppressor genes and oncogenes.

The tumor-suppressor genes and oncogenes used in this study were defined by the TUSON algorithm from genome sequencing of >8,200 tumor/normal pairs⁹, namely residue-specific activating mutations for oncogenes and discrete inactivating mutations for tumor-suppressor genes. TUSON is a computational method that analyzes patterns of mutation in tumors and predicts the likelihood that any individual gene functions as a tumor-suppressor gene or oncogene. We ranked genes by their TUSON prediction P values from the most to the least significant and used the top 500 genes (P < 0.01) as the reference tumor-suppressor genes or oncogenes. After removing 30 genes in common, 470 tumor-suppressor genes and oncogenes were used for the enrichment analysis. Note that there were very few breast tumor-specific tumor-suppressor genes and oncogenes (36 and 3 with breast q-value ≤ 0.5, respectively) and 90% of them were found in the top 500 pan-cancer predictions.

Previously identified 3′ US genes in cancers.

Xia et al. identified 1,187 3′ US genes across 7 TCGA cancer types¹. Mayr and Bartel selected 23 3′ US genes from 27 cancer cell lines⁵. Fu et al. identified 428 3′ US genes in human breast cancer cell lines⁶. Lin et al. reported 120 3′ US genes in major cancers and tumor cell lines⁷. Morris et al. found 286 3′ US genes in human colorectal tumor samples⁸. The 3′ US genes of Xia et al. were randomly sampled to 100 genes for a fair comparison.

Selection of miRNA-binding sites.

Predicted miRNA-binding sites were obtained from TargetScanHuman version 6.2²⁵. Only those with a preferentially conserved targeting score (Pct) more than 0 were used¹. Experimentally validated miRNA-binding sites were obtained from TarBase version 5.0²⁶, miRecords version 4²⁷ and miRTarBase version 4.5²⁸. The binding sites found in indirect studies such as microarray experiments and high-throughput proteomics measurements were filtered out²⁹. Another source is the microRNA target atlas composed of public AGO-CLIP data³⁰ with significant binding sites (q-value <0.05). The predicted and validated binding site information was then combined to use in this study.

TCGA breast tumor RNA-Seq and miRNA-Seq data.

Quantified gene expression files (RNASeqV1) for primary breast tumors (TCGA sample code 01) and their matching solid normal samples (TCGA sample code 11) were downloaded from the TCGA Data Portal³¹. We used 97 breast tumor samples that have matched normal tissues. A total of 10,868 expressed RefSeq genes (fragments per kilobase of transcript per million mapped reads (FPKM) ≥ 1 in >80% of all samples) were selected for downstream analyses. To better quantify gene expression in the presence of 3′ US, we used only coding regions (CDS) to quantify mRNA expression. Exon and CDS annotation for TCGA data and miRNA expressions (syn1445790) were downloaded from Sage Bionetworks’ Synapse database.

CeRNA identification in TCGA breast tumors.

CeRNAs were identified by miRNA-binding-site overlap and expression correlation^13,14. Only microRNAs with intermediate expression (between 0.01 and 100 in averaged fragments per million mapped fragments (FPM)) were used to capture dynamic interactions¹⁴. After removing genes with fewer than six such miRNA-binding sites, gene pairs with significant miRNA-binding-site overlap (<0.05 in Benjamini–Hochberg-corrected P value) were selected. Among them, pairs correlated (>0.6 in Pearson’s correlation coefficient) (P < 1 × 10⁻¹⁰) in gene expression were defined as ceRNAs. To account for mRNAs with variable 3′ UTRs, we used only CDS to quantify mRNA expression. Genes that are connected with >500 ceRNAs were defined as hub genes.

Model-based analysis of trans effect of 3′ US (MAT3UTR).

Suppose transcript x has a constitutive proximal 3′ UTR (pUTR) and a distal 3′ UTR that might be shortened in tumors (dUTR) (Supplementary Fig. 3a). We define MiRs(x, miR_j) as the amount of binding sites for miRNA miR_j in x.

$$\text{MiRs}\left(x,{\text{miR}}_j\right)=(\text{pUTR}\left(x,{\text{miR}}_j\right)+\text{dUTR}\left(x,{\text{miR}}_j\right)\times \text{PDUI}(x))\times \text{FPKM}(x)$$ (1)

where pUTR(x, miR_j) and dUTR(x, miR_j) are the numbers of miR_j binding sites in pUTR and dUTR of x, and FPKM (x) is expression of x. PDUI indicates the percentage of dUTR usage index¹. Note that equation (1) can also estimate for genes with no distal 3′ UTR by setting PDUI = 1.

To estimate the trans effect of 3′ US on gene y′, we define X to be a set of 3′ US genes that are ceRNA partners of y′ (Supplementary Fig. 3b) and Y to be a set of ceRNA partners to x ∈ X, including y′. Only moderately expressed miRNAs are considered, since they are likely to bind all possible binding sites. Thus, we can roughly use the amount of miRNA-binding sites to represent the miRNA function. The miR_j-binding effect on each copy of y′ can be defined as follows:

$$\text{TransE}\left(y^{\prime },{\text{miR}}_j\right)=\frac{\text{FPM}\left({\text{miR}}_j\right)}{{\sum }_{x\in X}\text{MiRs}\left(x,{\text{miR}}_j\right)+{\sum }_{y\in Y}\text{MiRs}\left(y,{\text{miR}}_j\right)}$$ (2)

where FPM(miR_j) is the miR_j expression level. Since miRNA can bind to any binding sites in the genes connected by the ceRNA relationship (X ∪ Y), both X and Y need to be considered.

The high-dimensional MAT3UTR input data are often highly correlated with each other (for example, 588 miRNAs in equation (2)). Therefore, MAT3UTR employs the ridge regression that is known to address the dimensionality and collinearity^32,33 in biological data. Indeed, the ridge regression yields a remarkably higher prediction power than classical linear regression. For example, MAT3UTR has a much smaller mean square error (0.38) than classical linear regression (mean square error = 10.84) (Supplementary Fig. 3f).

$$\text{MAT}3\text{UTR}\left(y^{\prime }\right)=\sum _{{\text{miR}}_j\in 3^{\prime }\text{UTR}\left(y^{\prime }\right)}{\beta }_{{\text{miR}}_j}\times \log \frac{\text{transE}{\left(y^{\prime },{\text{miR}}_j\right)}_{\text{tumor}}}{\text{transE}{\left(y^{\prime },{\text{miR}}_j\right)}_{\text{normal}}}+{\text{Є}}_{y^{\prime }}$$ (3)

subject to ${\sum }_{{\text{miR}}_j\in 3^{\prime }\text{UTR}\left(y^{\prime }\right)}{\beta }_{{\text{miR}}_j}\le t,$ the ridge regression penalty. MAT3UTR(y′) is the trans effect of 3′ US; β_miRj is the regression coefficient of miR_j; Є_y′ is the gene-specific error term. We use R² to show how much variation in gene expression can be explained by the MAT3UTR model. We also used 10-fold cross-validation (CV) to choose the optimal regularization parameter t with 75% of data for training and the remaining 25% for testing. CV error is measured by mean-squared error. Then, to estimate β, we fit the ridge regression with the entire data set using the selected regularization parameter as chosen by CV.

As a result, y′ would be more repressed following 3′US, if: y′ contains more miRNA-binding sites in its 3′ UTR; X and Y contain fewer miRNA-binding sites; and more transcripts in X undergo 3′ US. The MAT3UTR-control model, which considers miRNA expression but not 3′ US, is defined as:

$$\text{MAT}3\text{UTR}-\text{control}\left(y^{\prime }\right)=\sum _{{\text{miR}}_j\in 3^{\prime }\text{UTR}\left(y^{\prime }\right)}{\beta }_{{\text{miR}}_j}\times \log \frac{\text{FPM}{\left({\text{miR}}_j\right)}_{\text{tumor}}}{\text{FPM}{\left({\text{miR}}_j\right)}_{\text{normal}}}+{\text{Є}}_{y^{\prime }}$$ (4)

where FPM(miR_j) is the miR_j expression level. For model comparison between MAT3UTR and MAT3UTR-control, we randomly selected 75% of data for training and the remaining 25% for testing. We perform random division 100 times to evaluate the performance of the MAT3UTR and MAT3UTR-control models, where 10-fold CV also confirms that MAT3UTR has a 2-fold higher prediction power on gene expression variation than the MAT3UTR-control model (Supplementary Fig. 3e).

Selecting genes for experimental validations.

To test the trans repressive effect of 3’US on PTEN, we chose EPS15 on three grounds. First, its expression is easily detected in MCF-7 cells; second, analysis of RNA-Seq from MCF-7 cells³⁴ indicates distal polyA site usage of the EPS15 transcript; third, the EPS15 3′ UTR contains four microRNA target sites that compete with the PTEN 3′ UTR.

To investigate the tumor-suppressor ceRNA downregulation mechanism, we chose PHF6, because among 57 tumor-suppressor genes in 3′ US ceRNAs, PHF6 was predicted as a strong (sixth highest in MAT3UTR score, Supplementary Table 3) trans-target of 3′ US, was significantly downregulated (second highest in gene expression) and was the most enriched in AGO2 RIP complexes of the ceRNA tested (Fig. 4d).

Statistical analyses.

Differential expression analyses were carried out by edgeR (version 3.8.6)³⁵ (tumor samples versus normal samples) with false discovery rate (FDR) control at 0.05. The significance of observed values for a particular class compared to its control is calculated from one-tailed Pearson’s χ² test. Each variable follows either a binomial or multinomial distribution and each case consists of at least five counts, which meets the assumption of Pearson’s χ² test. To test whether there is a significant enrichment of tumor-suppressor genes or oncogenes among a gene list of our interest, we conducted hypergeometric tests with normalized overlap counts, since assessing overlap between sets meets all criteria to use hypergeometric tests, including trials without replacement. To compare means of two groups that have different variances, we used Welch’s t-test, which does not assume equal population variance. To check the normality assumption for the t-test, we conducted a Shapiro-Wilk normality test for small samples (n < 50). All statistical computations were performed in the Python scipy stats package (version 0.15.1) or R (version 3.1.1).

RNA-Seq for NUDT21 depletion experiment.

We previously sequenced two control and two NUDT21 depletion samples of HeLa cells by HiSeq 2000 (LC Sciences)². After trimming adaptors using Trim Galore (version 0.4.1), paired-end RNA-Seq reads of 101 base pairs in each end were used to reconstruct the transcriptome in the Tuxedo protocol³⁶ (TopHat 2.0.6 and Cufflinks 2.1.1). The resulting FPKM values were normalized for comparison using Cuffdiff 2.2.0. Further analyses are based on 10,681 expressed (FPKM ≥ 1 in >3 samples) RefSeq genes. We sequenced miRNAs from control and NUDT21-knockdown cells to utilize only miRNAs with intermediate expression in ceRNA identification.

CeRNA identification in the NUDT21-knockdown experiment in the HeLa cell line.

Due to the small sample size (two for each condition wild-type and NUDT21 knockdown), ceRNAs were identified solely on the basis of miRNA-binding-site overlap. We considered only binding sites for miRNAs with intermediate expression (between 0.01 and 100 in averaged FPM). A total of 1,450 3′ US genes identified by DaPars had significant miRNA-binding-site overlap with 1,168 ceRNA genes (3′ US ceRNA partners).

MiRNA-Seq for the NUDT21 depletion experiment.

HeLa cells were transfected with control or NUDT21 siRNA. NUDT21 depletion was validated as previously described². Small RNA libraries were generated from one control and one NUDT21 depletion sample using the Illumina Truseq Small RNA Preparation kit, and sequenced on Illumina GAIIx. Raw sequencing reads (40 nucleotides) were obtained using Illumina’s Sequencing Control Studio software following image analysis and base-calling by Illumina’s Real-Time Analysis (v 1.8.70). Then a script ACGT101-miR v 4.2 (LC Sciences) was used for data analysis, where reads are mapped to the reference database (miRbase). The script also normalizes the counts by a library size parameter for comparison.

CeRNA tumor-suppressor repression in HeLa cells with NUDT21 knockdown.

Parental HeLa cells were purchased from ATCC (cat. no. CCL-2) and maintained in Eagle’s minimum essential medium (Lonza, cat. no. 12–604F) with 10% fetal bovine serum. The cells were made mycoplasma free by incubating with Plasmocin (InvivoGen, cat. no. ant-MPT) for two weeks before transfection with three different siRNAs for NUDT21 (Sigma Aldrich, ID: SASI_Hs01_00146875~77) and negative control siRNA (Sigma Aldrich, ID:SIC002) using previously established approaches². Western blotting was also performed as described in our previous work² using antibodies raised against: PHF6 (Santa Cruz, cat. no. sc-271767), YOD1 (abcam, ab178979), NUDT21 (Proteintechlab, cat. no. 10322–1-AP) and GAPDH (Sigma, G9545). To block miRNA function, we selected two miRNAs with a strong trans effect targeting PHF6 (miR-3187–3p and miR-549) and HeLa cells were co-transfected with siRNA for NUDT21 and the two antagomirs, to block the two predicted miRNAs, miR-549 and miR-3187–3p in the PHF6 3′ UTR. The two antagomirs were designed³⁷ and synthesized from Sigma-Genosys: Antagomir-3187–3p: 5′-[mU]s[mU]s[mG]mG] [mC][mC][mA][mU][mG][mG][mG][mG][mC][mU][mG] [mC][mG]s[mC]s[mG]s[mG]s-chol-3′; and Antagomir-549: 5′-[mU]s[mG]s[mA][mC][mA][mA][mC][mU][mA][mU][mG][mG][mA][mU][mG][mA][mG][mC]s[mU]s[mC]s[mU]s-chol-3′. PHF6 and YOD1 expression were detected by western blotting and quantified by Image Lab software (version 5.2.1) from Bio-Rad.

Detection of ceRNA tumor-suppressor gene enrichment by RIP with quantitative PCR.

HeLa cells were seeded in a 6-well plate at 4 × 10⁵ cells per well and transfected with a Cas9 and single-guide RNA (sgRNA) plasmid targeting NUDT21 or with Cas9 and GFP as a control. sgRNAs for NUDT21 (top, ccggccgcccaatcgctcgcagac; bottom, aaacgtctgcgagcgattg ggcgg) were synthesized (Sigma), and the annealing double-stranded DNA was cloned into pGL3-U6-sgRNA-PGK-puromycin. The transfected cells from three wells were combined and then selected with 10 μgml⁻¹ blasticidin for three days. NUDT21-knockdown efficiency was determined by western blot with NUDT21 antibody. RIP was performed with anti-AGO2 antibody, and AGO2-associated RNAs were purified and measured by quantitative real-time PCR³⁸. Briefly, the cells were harvested and lysed with 100 μl polysome lysis buffer (100 mM KCl, 5 mM MgCl₂, 10 mM Hepes pH 7.0, 0.5% NP50, 1 mM DTT and 1×PI cocktail). The cell lysate was centrifuged at 10,000g for 15 min and added to magnetic beads (A+G) with 5 μg anti-Ago2 antibody or normal mouse IgG suspended in 900 μl of NET2 buffer (50 mM Tris-Cl pH 7.4, 150 mM NaCl, 1 mM MgCl₂, 0.05% NP-40, 17.5 mM EDTA pH 8.0, 1 mM DTT and 100 units ml⁻¹ RNaseOUT). The beads were washed six times with NT2 buffer (50 mM Tris-Cl pH 7.4, 150 mM NaCl, 1 mM MgCl₂, 0.05% NP-40). Beads were resuspended in 150 μl proteinase K buffer (50 mM Tris-Cl pH 7.4, 150 mM NaCl, 1 mM MgCl₂, 0.05% NP-40 and 1% SDS) with 9 μl proteinase K. Samples were incubated at 55 °C for 30 min and isolate total RNAs with 150 μl phenol–chloroform. The total RNA was reverse transcribed and the candidate ceRNAs were determined by quantitative real-time PCR using primers described in Supplementary Table 5 (Bio-Rad real-time PCR system).

LightSwitch luciferase reporter assay with PTEN, PHF6 and LARP1 3′ UTR.

LightSwitch luciferase reporter constructs with PTEN, PHF6 and LARP1 3′ UTR were purchased from SWITCHGEAR genomics. Briefly, HeLa cells were seeded in a 96-well white TC plate in 100 μl total volume to yield ≥80% confluence at the time of transfection. For each transfection, the following reagents were combined: 50 nM siRNA and/or miRNAs and/or antagomir RNA, individual GoClone reporter (30 ng μl⁻¹) 3.33 μl and 1 ng Rluc reporter. Lipofectamine 2000 was diluted in OPTI-MEM medium at 1:10 and incubated at room temperature for 5 min and then added to each tube. Following a 20-min incubation at room temperature, 80 μl of pre-warmed (37 °C) OPTI-MEM medium per replicate was added for a total of 100 ul per replicate transfection. All 100 μl of the transfection mixture was added to each well and incubated overnight. The luciferase reporter assays were performed according to the manufacturer’s protocol (Invitrogen).

Immunofluorescence staining for PTEN in MCF7 cells with EPS 3′ UTR.

pLightSwitch-EPS15 3′ UTR construct was purchased from SWITCHGEAR genomics and transfected into MCF7 cells. PTEN expression was detected by immunofluorescence staining with anti-PTEN antibody from Cell Signaling. Briefly, 1 × 10⁵ MCF7 cells were seeded in 4-well chamber slides overnight, and transfected with pLightSwitch-EPS15 3′ UTR/GFP constructs at 10:1 or pLightSwitch-3′ UTR/GFP constructs as a control. One day after transfection, the cells were fixed with 90% cold methanol at −20 °C overnight. The next day, 0.5% Triton X-100 in PBS was added and incubated at room temperature for 30 min. Samples were blocked in 3% BSA in PBS at room temperature for 1 h. PTEN antibody was used at 1:200 dilution in 3% BSA/PBS and 200 μl per well was added to the chamber slides and incubated for 1 h at room temperature. After washing three times, the cells were incubated with Alexa-594-conjugated secondary antibody in 3% BSA/PBS for 1 h at room temperature, in the dark. The cells were rinsed three times with PBS, with the third wash containing DAPI. The coverslips were mounted in anti-fade mounting medium and detected by immunofluorescence microscopy. Both PTEN- and GFP-positive cells were counted in EPS15 3′ UTR/GFP cells and pLightSwitch-3′ UTR/GFP control cells.