RBMS1 Suppresses Colon Cancer Metastasis through Targeted Stabilization of Its mRNA Regulon

PRADA Identifies RBMS1 as a Novel Regulator of Colorectal Cancer Progression

Metastasis is a complex multistep process and requires modulation of many cellular pathways and functions. As such, increased metastatic capacity often involves broad reprogramming of the gene expression landscape in cancer cells. Thus, a mechanistic dissection of cancer progression relies on the systematic identification of the underlying regulatory programs that drive pathologic cellular states. To accomplish this, we developed PRADA, which uses regulatory network predictions to identify the key RNA-binding proteins (RBP) that may act as master regulators of mRNA stability and gene expression. Conceptually, PRADA solves a regression problem to predict changes in gene expression as a function of the expression of RBPs and their known or predicted targets across the transcriptome. In this customized regression analysis, the coefficient assigned to each RBP reflects its strength as a regulator of gene expression and the direction of its effect (i.e., activator or repressor). Given our limited knowledge of post-transcriptional regulatory pathways and their role in human disease, PRADA provides an opportunity for the discovery of previously unknown disease pathways. To assess the contribution of post-transcriptional regulatory programs to colon cancer metastasis, we took advantage of a publicly available set of gene expression profiles from established colon cancer cell lines (GSE59857). We first categorized these cell lines as poorly metastatic or highly metastatic based on their liver colonization capacity in xenograft mouse models (Supplementary Fig. S1A; ref. 5). We then performed differential gene expression analysis to compare gene expression changes across the transcriptomes of the two groups. Finally, we applied PRADA to this dataset to find RBPs whose differential activity is most informative of metastasis-associated gene expression changes. PRADA identified the protein RBMS1 as the RNA-binding protein with the strongest regulatory potential in this dataset (Fig. 1A). The size and the direction of the regulatory coefficient assigned to RBMS1 implies that RBMS1 levels are strongly informative of changes in the expression of transcripts with sequence matches to RBMS1-binding sites. To confirm this, we performed motif enrichment analysis using FIRE, which uses mutual information to capture the association between the presence and absence of a given binding site and genome-wide transcriptomic measurements (6). As shown in Fig. 1B, the RBMS1 consensus binding site is strongly informative of gene expression differences between poorly and highly metastatic cells, showing a highly significant enrichment among the transcripts that have decreased expression in highly metastatic cells. To ensure that this result is not sensitive to the choice of a specific RBMS1 binding motif, we tested two additional sequence models: (i) an independently derived representation of RBMS1-binding preferences (7) and (ii) predictions based on DeepBind models (ref. 8; Supplementary Table S1). Each of these three predicted RBMS1-binding motifs gave largely identical results, and we observed a similar decrease in the expression of the putative RBMS1 regulon derived from each motif (Supplementary Fig. S1B). For consistency, we have used the DeepBind-derived RBMS1 regulon (referred to as the RBMS1 regulon hereafter) in our subsequent analyses, as DeepBind is the state-of-the-art approach for predicting protein–nucleic acid interactions. Importantly, concordant with the results reported by PRADA and the lower expression of its regulon, the expression of RBMS1 was also significantly lower in highly metastatic cells (Fig. 1C). Moreover, consistent with RBMS1 acting as a post-transcriptional regulator of these genes, we observed a significant correlation between the expression of RBMS1 and the average expression of its regulon in multiple independent datasets (Fig. 1D; Supplementary Fig. S1C and S1D).

• Download figure
• Open in new tab
• Download powerpoint

Figure 1.

RBMS1 silencing in metastatic cells is associated with lower expression of RBMS1 targets. A, Regression coefficients set by PRADA as a function of the l1-norm of the coefficient vector. Each line is associated with an RNA-binding protein and the magnitude of its coefficient is a measure of its strength as a putative regulator of gene expression. Here, the first ten nonzero coefficients are shown as a function of l1-norm (i.e., sum of the magnitude of all coefficients). B, Analysis of RBMS1 recognition sites across gene expression changes between poorly and highly metastatic colon cancer lines. In this analysis, transcripts are first ordered on the basis of their log-fold changes from left (lower expression in metastatic cells) to right (higher expression) and then partitioned into equally populated bins (∼1,000 transcripts per expression bin). The red bars on the black background show the range of values in each bin (with the minimum and maximum values, i.e., −1.5 and 2.1, presented on the left). As shown here, genes that are expressed at a lower level in highly metastatic cells were significantly enriched for the RBMS1 binding motif (KAUAUAS; ref. 38). In this heat map, gold represents overrepresentation of putative RBMS1 targets whereas blue indicates underrepresentation. Enrichment and depletions that are statistically significant (based on hypergeometric distribution) are marked with red and dark blue borders, respectively. Also included are the logo representation of the RBMS1-binding motif, its mutual information (MI) value, the associated z-score, and the Bayes factor (BF; for details of this analysis, see ref. 28). C, RBMS1 expression in colon cancer lines grouped on the basis of their metastatic capacity (Supplementary Fig. S1A). P value calculated using a two-tailed Mann–Whitney U test. D, Linear regression analysis of RBMS1 expression versus average normalized expression of its putative regulon in The Cancer Genome Atlas Colon Adenocarcinoma (TCGA-COAD) dataset (cBioPortal; N = 382). Shown are the Spearman correlation coefficient and the associated P value. E, Enrichment and depletion patterns of the RBMS1 regulon in PDX models of colorectal cancer liver metastasis. For this analysis, log-fold changes between parental (CLR-Par) and liver metastatic (CLR-LvM) were averaged across three independent PDX models, CLR4, CLR27, and CLR32. D00198.001 is the unique identifier containing the binding site information of human RBMS1 obtained from DeepBind (8) analysis. The distribution of RBMS1 targets was then assessed using mutual information, its associated z-score and Bayes factor. The enrichment and depletion patterns were visualized as described in B. F, Relative RBMS1 levels in matched poorly metastatic (Par) and highly liver metastatic (LvM) derivatives for three PDX models (CLR4, CLR32, and CLR27). P value was calculated using DESeq2 (39). G and H, Expression of RBMS1 and RBMS1 targets in matched primary tumor and liver metastases from two patients (S1029 and S567). The results are presented as described in B. The P value for RBMS1 silencing was calculated using DESeq2 (controlled for genetic background).

To further assess the association between RBMS1 and colorectal cancer metastasis in more directly disease-relevant models, we took advantage of patient-derived xenograft (PDX) models of colon cancer metastasis to the liver. We used RNA-sequencing (RNA-seq) data from three parental colorectal cancer PDX models (CLR4, CLR27, and CLR32) and their highly liver-metastatic derivatives (CLR-LvM). The highly metastatic CLR-LvM models were derived from the parental PDXs by repeated splenic delivery and growth of cells in the liver of immunocompromised mice (NOD/SCID gamma), recapitulating the major site of colorectal cancer metastasis in humans (9). We observed that the increase in the metastatic capacity of the CLR-LvM PDX models compared with their parental PDXs was accompanied by a significant reduction in the expression of the RBMS1 regulon in the highly metastatic CLR-LvM models (Fig. 1E). More importantly, all three models also showed a significant and concomitant decrease in the expression of RBMS1 (Fig. 1F).

In addition to these cell line and PDX models of colorectal cancer liver metastasis, we also performed RNA-seq on matched primary tumors and liver metastases biopsied and frozen from two patients. As shown in Fig. 1G, in both cases the RBMS1 regulon was enriched among transcripts that were downregulated in the metastases relative to their primary tumors. Interestingly, although the gene expression changes between the primary tumors and metastases were not generally correlated (R = 0.02, P = 0.1), the RBMS1 regulon was independently downregulated in both metastatic tumors (Supplementary Fig. S1E). Consistently, RBMS1 was also strongly silenced in both metastases (Fig. 1H). Collectively, these findings in cell lines, PDXs, and clinical samples implicate RBMS1 silencing in both the downregulation of its putative regulon and in colorectal cancer metastatic progression.

RBMS1 Acts as a Post-Transcriptional Regulator of RNA Stability

The correlated expression of RBMS1 and its putative regulon defined from its binding site preferences implies that RBMS1 modulates gene expression. To further assess the potential role of RBMS1 as a post-transcriptional regulator, we performed RNA-seq of RBMS1 knockdown and control SW480 colon cancer cells. We chose the SW480 cell line for this experiment because: (i) RBMS1 expression in SW480 is among the highest in colon cancer lines tested by us and others and (ii) it is an established xenograft model of liver metastasis and is considered to be poorly metastatic (relative to other colorectal cancer lines listed in Supplementary Fig. S1A). We used two independent shRNAs to silence RBMS1 in SW480 cells and confirmed its knockdown with both qPCR and Western blotting (Supplementary Fig. S2A). A 2.5-fold reduction in RBMS1 protein expression induced major changes in gene expression (Supplementary Fig. S2B) and resulted in a significant decrease in the expression of RBMS1 regulon (Fig. 2A). Because we are focused on the role of RBMS1 as an RNA-binding protein, we reasoned that its effect on gene expression is likely through regulation of the stability of its target RNAs. To test this hypothesis, we took advantage of REMBRANDTS, a computational framework we have developed to estimate RNA stability from RNA-seq data (10). We have previously established that REMBRANDTS accurately recapitulates experimental RNA stability measurements (10). Application of this method to RNA-seq data from RBMS1 knockdown and control cells found a significant enrichment of RBMS1 targets among transcripts destabilized upon RBMS1 silencing (Fig. 2B). REMBRANDTS relies on the comparison of exonic and intronic reads to measure changes in RNA stability. As shown in Supplementary Fig. S2C, RBMS1 silencing results in lower expression of its putative regulon as measured by exonic reads; however, intronic reads, which reflect changes in pre-mRNA levels (and transcription rates), do not significantly change in response to RBMS1 depletion. To further strengthen our claim that RBMS1 functions as a post-transcriptional regulator of RNA stability rather than as a transcriptional activator, we also performed whole-genome RNA stability measurements in control and RBMS1 knockdown cells. For this, we used α-amanitin–mediated inhibition of RNA-polymerase II followed by RNA-seq at two time points (0 and 9 hours). Consistent with our analyses with REMBRANDTS, we observed a similar reduction in the stability of the putative RBMS1 regulon upon RBMS1 silencing (Fig. 2C). We used RNA-seq data from the three matched PDX models (CLR panel) to compare RNA stability in highly liver metastatic models to that of their parental PDXs (using REMBRANDTS). As shown in Fig. 2D, lower RBMS1 expression in the LvM models (Fig. 1F) accompanies a reduction in the stability of its target regulon in all three independently derived models. Together, these findings establish RBMS1 as a post-transcriptional regulator of RNA stability with broad functional consequences for the transcriptome and clear implications for colorectal cancer progression.

• Download figure
• Open in new tab
• Download powerpoint

Figure 2.

RBMS1 post-transcriptionally regulates the stability and expression of its targets. A, Enrichment and depletion patterns of the RBMS1 regulon in RBMS1 knockdown cells relative to control (∼2.5-fold knockdown). B, We used our computational tool, called REMBRANDTS, to estimate changes in RNA stability upon RBMS1 silencing. These differential stability estimates were then used to assess the enrichment patterns of the RBMS1 targets across the changes in RNA decay. C, Experimental RNA stability changes were measured using α-amanitin treatment as described previously (1). The enrichment and depletion patterns of the RBMS1 regulon were then assessed among the transcripts destabilized or stabilized upon RBMS1 knockdown. D, We used REMBRANDTS to measure changes in RNA stability between poorly and highly metastatic PDX models from three independent PDX models (CLR27, CLR32, and CLR4). As shown here, consistent with the silencing of RBMS1 in LvM PDX models (Fig. 1F) and the downregulation of its regulon (Fig. 1E), the RBMS1 regulon is destabilized in these three independent models of colorectal cancer metastasis.

RBMS1 CLIP-seq Reveals 3′ UTR Binding to Target mRNAs

To this point, our analyses were performed using in silico predicted binding targets of RBMS1. To create a transcriptome-wide snapshot of RBMS1-binding sites in colon cancer cells at nucleotide resolution, we performed UV cross-linking immunoprecipitation followed by sequencing (CLIP-seq). We carried out irCLIP (11) for endogenous RBMS1 in SW480 cells and identified hundreds of high-confidence RBMS1-binding sites across the transcriptome (Fig. 3A). Importantly, we observed that 90% of the CLIP-identified RBMS1 targets overlapped with our computationally derived putative RBMS1 regulon (Supplementary Table S1). Consistently, these RBMS1-bound transcripts in our irCLIP data (RBMS1 targets hereafter) follow the same expression patterns as the computationally derived RBMS1 regulon described above, exhibiting both decreased stability and expression upon RBMS1 silencing (Fig. 3B; Supplementary Fig. S3A).

• Download figure
• Open in new tab
• Download powerpoint

Figure 3.

RBMS1 irCLIP identifies direct RBMS1 targets in colon cancer cells. A, 509 RBMS1 binding sites were found using irCLIP, with a significant enrichment of binding to the last exon/3′ UTR (relative to the total length of genomic features). Last exons from LATS2, AKAP12, and SDCBP are shown as examples of RBMS1-binding patterns. B, Enrichment of the RBMS1-bound mRNAs among those that are downregulated in highly metastatic cells (top) and those destabilized upon RBMS1 silencing (bottom). C, qRT-PCR was used to measure changes in GFP mRNA levels upon cloning RBMS1-binding sites of the listed genes downstream of the GFP open reading frame. mCherry was expressed from the same bidirectional promoter as GFP, and mCherry levels were used to normalize GFP measurements. A one-sample Wilcoxon signed rank test was used to assess whether the ratios in siRBMS1 samples were significantly below 1.0. D, Scatter plot of mass spectrometry data showing proteins that coimmunoprecipitate with RBMS1 versus control IgG in SW480 cells. Shown are the average of three replicates across all detected proteins. Proteins enriched in the RBMS1 co-IP samples are shown in pink. RBMS1, PABPC1, and ELAVL1 are highlighted in red, green, and violet, respectively. E, RBMS1, PABPC1 and ELAVL1 were detected by Western blot analysis in input and eluate samples from RBMS1 and IgG immunoprecipitations (SW480 cell lysates). For RBMS1 immunoprecipitation, the lysates were additionally treated by RNaseA. F, Heat map showing the enrichment of poly(U) sites in the 3′ UTRs of RBMS1-bound mRNAs. G, Venn diagram showing the overlap between RBMS1- and ELAVL1-bound mRNA 3′ UTRs. ELAVL1 targets were determined by PAR-CLIP (13). P value calculated using hypergeometric test. H, Density plot showing the log fold-change in RBMS1 target expression (determined by RNA-seq; ref. 14) upon ELAVL1 knockdown in SW620 cells. Median value (mu) is indicated. P value calculated using one-tailed Wilcoxon signed rank test.

In our analysis of RBMS1-binding sites, we noted a strong enrichment of RBMS1 binding to the last exon (on or close to 3′ UTRs; Fig. 3A) of mRNAs. This is consistent with RBMS1 acting as a post-transcriptional regulator of RNA stability, and suggested that RBMS1 3′ UTR binding could result in increased RNA stability. To assess this possibility, we built a bidirectional CMV promoter that drives the expression of both GFP and mCherry. We then cloned nine CLIP-seq–derived RBMS1-binding sites, plus approximately 150 nucleotides flanking the binding sites, downstream of GFP, and asked whether there was a reduction in GFP mRNA relative to mCherry mRNA upon RBMS1 depletion. As shown in Fig. 3C, there was a decrease in GFP reporter transcript levels in almost all instances. To further ensure that this effect was RBMS1-dependent, we also tested the two reporters with the strongest reduction in transcript levels (DAG1 and CD9) in LS174T cells, where RBMS1 is endogenously silenced. In this instance, as expected, there was no response in reporter mRNA levels upon transfection of RBMS1-targeting siRNAs (Supplementary Fig. S3B).

To gain insight into how RBMS1 regulates the stability of its targets, we carried out an unbiased search for its interacting protein partners. We performed immunoprecipitation of RBMS1, along with an IgG control, from SW480 cells and analyzed the resulting protein complexes by mass spectrometry (Fig. 3D). From this data, we found that the Gene Ontology terms RNA binding, RNA processing, and negative regulation of mRNA metabolism were overrepresented among the RBMS1-interacting proteins (Supplementary Table S2). Of these proteins, we have confirmed an RNA-dependent interaction between RBMS1 and two well-characterized RBPs that predominantly bind 3′ UTRs, PABPC1 and ELAVL1 (Fig. 3E). These findings suggest that RBMS1 stabilizes its targets at least partially in concert with ELAVL1 which preferentially binds A/U-rich elements (ARE; ref. 12). In line with this hypothesis, we found a significant enrichment of poly(U) motifs in the 3′ UTRs of RBMS1 targets (Fig. 3F). As shown in Fig. 3G, 72% of RBMS1-bound 3′ UTRs are also bound by ELAVL1 in vivo [as determined by photoactivatable ribonucleoside-enhanced-CLIP (PAR-CLIP); ref. 13]. Furthermore, knockdown of ELAVL1 in SW620 colon cancer cells (of the same genetic background as SW480 cells) resulted in a significant downregulation of RBMS1 targets, as determined by RNA-seq (14; Fig. 3H). Together, these results indicate that direct binding of RBMS1 to mRNA 3′ UTRs, together with other stabilizing factors such as ELAVL1, results in increased mRNA stability.

RBMS1 Is a Suppressor of EMT and Metastatic Liver Colonization in Colon Cancer Cells

To carry out a functional study of RBMS1 and its downstream regulon in colorectal cancer metastasis, we picked an additional cell line model to complement the SW480 line. We chose the LS174T line because RBMS1 is almost completely silenced in this line as measured by qPCR and Western blotting (Supplementary Fig. S4A and S4B). LS174T is also an established xenograft model and is considered a highly liver metastatic cell line, approximately 100 times more metastatic than the SW480 line as measured by liver colonization assays (15). We first performed RNA-seq of both SW480 and LS174T lines and, consistent with RBMS1 silencing in the LS174T cells, we observed a significant reduction in the expression of the RBMS1 regulon in this line (Supplementary Fig. S4C).

To establish a causal link between RBMS1 silencing and higher metastatic capacity, we performed liver colonization assays by splenically injecting RBMS1 knockdown and control SW480 cells and measuring metastatic burden in the livers of mice over time. Although RBMS1 silencing did not have a strong effect on in vitro cell proliferation (Supplementary Fig. S4D), we observed a significant increase in liver colonization upon silencing RBMS1, based on in vivo bioluminescence measurements and gross liver mass at the experimental endpoint (Fig. 4A; Supplementary Fig. S4E). To control for possible off-target effects of the shRNAs, we repeated the experiment with an additional RBMS1-targeting hairpin and observed a similarly significant increase in metastatic liver colonization (Supplementary Fig. S4F). We also performed a gain-of-function experiment by expressing RBMS1 in the LS174T line, where RBMS1 is endogenously silenced. Consistent with our earlier findings, we observed that expressing RBMS1 in LS174T cells results in a marked reduction in their liver colonization capacity (Fig. 4B; Supplementary Fig. S4G). It should be noted that, as expected, expressing shRNAs targeting RBMS1 in highly metastatic cells in which RBMS1 is already silenced, namely LS174T, HCT116, and COLO320, did not have an impact on their liver colonization capacity in xenograft mouse assays, further supporting the on-target effect of the shRNAs used in these experiments (Supplementary Fig. S4H). In contrast, knockdown of RBMS1 in WiDr cells, in which RBMS1 is endogenously expressed, resulted in a significant increase in liver colonization in xenograft mouse models without an overt effect on in vitro cell proliferation (Supplementary Fig. S4I and S4J).

• Download figure
• Open in new tab
• Download powerpoint

Figure 4.

RBMS1 is a suppressor of epithelial-to-mesenchymal transition (EMT) and metastatic liver colonization. A, Bioluminescence imaging plot of liver colonization by RBMS1 knockdown or control cells; N = 5 in each cohort. Two-way ANOVA was used forstatistical testing. Ex vivo liver signal was also measured and comparedusing a one-tailed Mann–Whitney U test. Also shown are representative (median signal) mice and livers. B, Splenic injection of LS174T highly metastatic colon cancer cells overexpressing RBMS1 (RBMS1-OE) and those expressing mCherry as a control. Day 21 signal (normalized to day 0) was plotted and compared for both in vivo and ex vivo signal (N = 4–5; Mann–Whitney U test). C, Downregulation of EMT(−) signature genes upon RBMS1 knockdown in SW480 cells. 160-gene EMT(−) signature set (17) was compared with the rest of the transcriptome using a Mann–Whitney U test. D, Immunofluorescence staining for E-cadherin (red) in RBMS1 knockdown and control cells (SW480 background). Note the lower expression of E-cadherin and the more spindle-like cellular morphology in RBMS1 knockdown cells (top panels show DAPI signal). E-cadherin intensity and maximum Feret diameter (a measure of length of the cell) for cells in control and knockdown samples (N = 64 and 37, respectively). Two-tailed Mann–Whitney U test was used to compare measurements.

The metastatic cascade is a complex multistep process, involving cellular processes that affect not only cancer cell growth and survival, but also cell migration and invasion. As mentioned earlier, we did not observe a significant change in in vitro cell proliferation rates upon RBMS1 knockdown. In addition, transwell invasion assays of RBMS1 knockdown and control cells did not detect a significant role for RBMS1 in cancer cell invasion (Supplementary Fig. S4K). Therefore, to gain a better understanding of the key step(s) in the metastatic cascade that are regulated by RBMS1, we performed a systematic analysis of known and predicted gene sets to identify those that may be modulated upon RBMS1 silencing. We performed this analysis using iPAGE, an information-theoretic framework we have developed for this type of analysis (ref. 16; Supplementary Fig. S4L). Downregulation of genes associated with negative regulation of epithelial–mesenchymal transition [EMT(−) signature gene set defined by various studies; refs. 17–19] was the most significant pathway identified in RBMS1 knockdown cells (Fig. 4C; Supplementary Fig. S4L and S4M). For example, the canonical EMT marker E-cadherin (CDH1) is significantly downregulated by approximately 2-fold in RBMS1 knockdown cells based on RNA-seq data (Supplementary Fig. S2B). We confirmed this by performing immunofluorescence staining for E-cadherin in RBMS1 knockdown and control cells. We observed a reduction in E-cadherin expression in RBMS1 knockdown cells and also noted the appearance of spindle-like cell morphology that is associated with EMT (Fig. 4D). Moreover, the EMT(−) signature genes were also expressed at relatively lower levels in the LS174T cell line, where RBMS1 is endogenously silenced (Supplementary Fig. S4M). Finally, analysis of The Cancer Genome Atlas Colon Adenocarcinoma (TCGA-COAD) dataset revealed significantly lower E-cadherin expression in tumor samples with low RBMS1 expression (∼20% reduction, P = 0.006) and a significant general correlation between RBMS1 and E-cadherin expression (ρ = 0.1, P = 0.03), which is indicative of the clinical relevance of this regulatory axis in colorectal cancer.

AKAP12 and SDCBP Are Functional Downstream Targets of RBMS1

To identify genes that are regulated by RBMS1 and act as suppressors of metastasis in colorectal cancer, we used an integrated analytic approach, which relies on mining relevant datasets to prioritize target genes based on their direct interaction with RBMS1 (3′ UTR binding), their RBMS1-dependent magnitude of change in expression, the robustness of their response to RBMS1 silencing, as well as their lower expression in metastatic clinical samples. Using this approach, we prioritized two target genes, AKAP12 and SDCBP, that satisfy these criteria: (i) destabilized and downregulated in highly metastatic colorectal cancer cells and PDXs, (ii) destabilized and downregulated in RBMS1 knockdown cell lines, (iii) directly bound by RBMS1 based on our irCLIP data (Fig. 3A), and (iv) downregulated in liver metastases relative to primary colon cancers in a publicly available dataset (20). Consistently, silencing RBMS1 in SW480 cells resulted in the downregulation and destabilization of AKAP12 and SDCBP mRNAs (Fig. 5A and B). Conversely, overexpression of RBMS1 in LS174T cells resulted in upregulation and stabilization of these targets (Fig. 5A and B). In addition, in line with the concerted action of RBMS1 and ELAVL1, both AKAP12 and SDCBP 3′ UTRs are bound by ELAVL1 in vivo (13). These observations establish AKAP12 and SDCBP as direct targets of RBMS1 that are post-transcriptionally regulated by this RBP.

• Download figure
• Open in new tab
• Download powerpoint

Figure 5.

AKAP12 and SDCBP act downstream of RBMS1 to suppress colorectal cancer metastasis. A, Changes in the expression of AKAP12 and SDCBP mRNAs upon RBMS1 knockdown in SW480 cells and RBMS1 overexpression (OE) in LS174T cells. The expression of target genes was determined by qRT-PCR, normalized to HPRT internal control and shown as relative fold change over shControl or OE-Control. P value was calculated using one-tailed Mann–Whitney U test. B, Changes in the stability of AKAP12 and SDCBP mRNAs upon RBMS1 knockdown in SW480 cells and RBMS1 overexpression in LS174T cells. RNA stability was calculated by comparing mRNA levels with and without treatment with α-amanitin, and shown as relative fold change over shControl or OE-Control. The relative abundance of target genes was determined by qRT-PCR, normalized to 18S RNA (RNA Pol I transcript insensitive to α-amanitin). P value was calculated using one-tailed Mann–Whitney U test. C, In vivo liver colonization assays were used to measure the impact of CRISPRi-mediated silencing of the RBMS1 targets AKAP12 and SDCBP on liver metastasis (N = 6). Also shown are representative mice from each cohort. Two-way ANOVA was used to compare cohorts with control (P = 0.01 and 0.02 for sgAKAP12 and sgSDCBP, respectively). Livers were also extracted and their tumor burden was measured ex vivo. Mann–Whitney U test was used to compare measurements. D, The expression of EMT(−) and EMT(+) signature genes relative to background (BG) in AKAP12 knockdown (CRISPRi) and control cells (measurements using 3′-end RNA-seq). Shown are the ANOVA P value and a Mann–Whitney comparison between EMT(−) and background genes.

To test the possible role of AKAP12 and SDCBP in driving colorectal cancer metastatic progression, we silenced them individually in SW480 cells using CRISPR interference (CRISPRi; ref. 21) and performed liver colonization assays. As shown in Supplementary Fig. S5A, silencing these genes did not have a significant impact on in vitro proliferation of SW480 cells. However, reduced expression of these genes resulted in increased metastatic liver colonization in mice (Fig. 5C). To confirm that AKAP12, which, of the two targets tested, elicited a stronger phenotype when silenced, indeed functions downstream of RBMS1, we performed an in vivo epistasis experiment by generating cells with knockdown of both AKAP12 and RBMS1 in SW480 cells. As shown in Supplementary Fig. S5B, unlike in Fig. 5C, further silencing of AKAP12 failed to increase metastatic liver colonization in RBMS1 knockdown cells, in which AKAP12 is already downregulated. To evaluate the contribution of AKAP12 to the phenotype of RBMS1 silencing, we compared AKAP12 knockdown and double knockdown cells. As shown in Supplementary Fig. S5C, we observed a slight additional increase (albeit not statistically significant) in metastatic liver colonization upon RBMS1 silencing in AKAP12 knockdown cells. This suggests that AKAP12 is a major effector downstream of RBMS1, and that other downstream targets, such as SDCBP, may also play minor roles in this process. Importantly, and consistent with AKAP12 acting downstream of RBMS1, gene expression profiling of AKAP12 knockdown and control cells revealed a significant induction of an EMT signature (Fig. 5D). Notably, this included the upregulation of canonical EMT transcriptional repressors SNAI1 and SNAI2 (Supplementary Fig. S5D).

As an orthogonal approach, to assess whether the effects of RBMS1 silencing on the transcriptome are reflected in the proteome, we used tandem mass tag labeling and tandem mass spectrometry to compare the global proteomes of SW480 cells with shRNA-mediated RBMS1 knockdown and control cells (Supplementary Fig. S5E; Supplementary Table S2). As shown in Supplementary Fig. S5F, we observed a positive and significant correlation between changes in mRNA abundance (as determined by RNA-seq) and changes in protein abundance. We found that RBMS1 targets, as determined by irCLIP, were enriched among the downregulated proteins in the RBMS1 knockdown cells (Supplementary Fig. S5G). Most notably, AKAP12, which we have identified as the downstream effector of RBMS1 in EMT regulation, was significantly downregulated at the protein level as well, mirroring our transcriptomic results (Supplementary Fig. S5H). Taken together, our findings demonstrate that the RBMS1–AKAP12 regulatory axis acts as a suppressor of EMT and liver metastasis in models of colorectal cancer progression.

RBMS1 Silencing and the Downregulation of Its Targets Is Associated with Colorectal Cancer Progression

To further assess the clinical relevance of this previously unknown regulatory pathway, we performed a variety of measurements in clinical samples as well as analysis of clinical data to evaluate RBMS1 activity in colorectal cancer metastasis. First, we performed qRT-PCR for RBMS1 in two independent clinical cohorts, one cohort stratifying patients based on their tumor stage (n = 96) and another comparing samples from normal mucosa, primary colorectal cancer tumors, and liver metastases (n = 91). As shown in Fig. 6A and B, we observed a significant reduction in RBMS1 expression as the disease progresses, with stage IV clinical samples showing the lowest RBMS1 expression levels. We also carried out survival analysis in a large clinical cohort with publicly available expression data and clinical outcomes (22). We observed a significant association between RBMS1 silencing and reduced relapse-free survival as well as overall survival in patients with colon cancer (Fig. 6C). A multivariate Cox proportional-hazards model revealed that this association with RBMS1 silencing remains statistically significant even when controlled for other known clinical factors that may contribute to patient survival (Supplementary Fig. S6A). This observation emphasizes the relevance of RBMS1 silencing as an effective readout for risk stratification of patients based on samples collected from their primary tumors.

• Download figure
• Open in new tab
• Download powerpoint

Figure 6.

RBMS1 and its target gene signature score are associated with colon cancer metastasis and reduced survival in patients with colorectal cancer (CRC). A, RBMS1 qPCR (relative to HPRT) in 96 colon cancer samples stratified on the basis of tumor stage. B, RBMS1 qPCR (relative to HPRT) in 29 normal mucosa (NM), 25 primary colon cancer (Pri), and 37 liver metastases (Met). C, RBMS1 silencing was observed in approximately 5% of primary colorectal cancer tumors (see Methods), and these patients showed substantially lower relapse-free and overall survival. Reported are Mantel–Haenszel HRs and P values from Gehan–Breslow–Wilcoxon tests. D and E, AKAP12 expression in clinical samples (similar to A and B). F, Patient primary tumors were scored on the basis of aggregate expression of RBMS1 signature genes and the resulting values were used to perform survival analyses similar to those in C. As shown here, lower RBMS1 signature score was significantly associated with lower relapse-free and overall survival. Also reported are Mantel–Haenszel HR and P values from Gehan–Breslow–Wilcoxon tests. G and H, Regression analysis comparing the expression of RBMS1 (G) or ELAVL1 (H) and RBMS1 80-gene signature set in TCGA pan-cancer dataset. Shown are the Spearman correlation coefficient and the associated P value.

Our results indicate that AKAP12 acts as a suppressor of metastasis downstream of RBMS1 and therefore is expected to show a similar association with metastatic disease. To test this hypothesis, we performed qRT-PCR measurements in the clinical cohorts mentioned above, and we observed a significant reduction in AKAP12 expression as a function of disease progression (Fig. 6D and E). Consistent with RBMS1 acting as a regulator of AKAP12, we observed a highly positive and significant correlation between the expression of these two genes (Supplementary Fig. S6B). Interestingly, our analyses indicate that the identified RBMS1 targets provide a robust gene signature that, similar to RBMS1, is highly informative of clinical outcomes. For this analysis, we defined an RBMS1 target score as an aggregate measure of expression for the RBMS1 80-gene signature set [average normalized expression of target mRNAs that (i) interact in vivo with RBMS1 on or within 200 nucleotides of their 3′ UTRs and (ii) are downregulated upon RBMS1 knockdown; this is referred to as the RBMS1 80-gene signature hereafter (Supplementary Table S1)]. We then stratified patients based on this score and we observed that lower expression of the RBMS1 80-gene signature is associated with lower relapse-free and overall survival in patients with colon cancer (Fig. 6F). Consistently, the RBMS1 signature score remained a significant covariate in a multivariate Cox proportional-hazards model of relapse-free survival (Supplementary Fig. S6C). Moreover, we have not observed significant association between the RBMS1 signature score and validated clinicopathologic features or other molecular markers (such as microsatellite instability status, left- or right-sidedness, or colorectal cancer consensus molecular subtypes). Importantly, similar to our initial observation with the RBMS1 regulon (Fig. 1D), the RBMS1 signature score was significantly correlated with RBMS1 expression in a large TCGA pan-cancer dataset (Fig. 6G). We also observed a high correlation between the RBMS1 signature score and ELAVL1 expression (Fig. 6H), consistent with RBMS1 and ELAVL1 acting in concert to promote mRNA stabilization. The association between the RBMS1 regulon and colorectal cancer metastasis is not limited to a single dataset. Consistent with our initial observations in matched samples (Fig. 1G), comparison of gene expression changes in liver metastases relative to colorectal cancer primary tumors in a publicly available dataset (23) revealed a significant reduction in the expression of the RBMS1 regulon (Supplementary Fig. S6D and S6E). Furthermore, analysis of publicly available RNA-seq data using REMBRANDTS to infer changes in mRNA stability from a set of matched primary and metastatic samples (24) also confirmed (i) RBMS1 silencing in multiple liver metastases and (ii) a concordant reduction in the stability of the RBMS1 regulon (Supplementary Fig. S6F). Together, these results establish the clinical relevance of our findings and the importance of RBMS1 silencing both as a prognostic factor and as a suppressor of liver metastasis in patients.

HDAC1-Mediated Promoter Deacetylation Leads to RBMS1 Silencing in LS174T Cells

Our findings described above establish a previously unknown regulatory pathway driven by the RNA-binding protein RBMS1 that plays a functional role in colorectal cancer metastasis to liver. Regulatory pathways can be expanded by uncovering the upstream pathways that influence their activity. We have developed computational and data-analytic tools designed to integrate publicly available datasets from a variety of biological sources to identify such upstream regulatory mechanisms (16). Using these tools, we sought to extend the RBMS1 regulatory pathway by exploring the mechanisms for RBMS1 silencing observed in the highly metastatic LS174T colon cancer cells. First, we found RBMS1 expression to be strongly associated with promoter acetylation, but did not identify any specific transcription factors associated with RBMS1 expression (Fig. 7A). Consistent with this observation, analysis of the connectivity map dataset (25), which reports the impact of hundreds of small-molecule treatments on gene expression, identified the histone deacetylase inhibitor trichostatin A (TSA) as an activator of RBMS1 expression and of the RBMS1 regulon (Supplementary Fig. S7A and S7B). As RBMS1 is endogenously silenced in LS174T cells and expressed in the SW480 line, we tested the relative acetylation levels of the RBMS1 promoter in these two lines by performing H3K27Ac chromatin immunoprecipitation (ChIP)–qPCR. These data showed relatively low levels of H3K27 acetylation at the RBMS1 promoter in the LS174T line compared with the SW480 line, consistent with the differences in the level of RBMS1 in the two lines (Fig. 7B). TSA simultaneously inhibits multiple histone deacetylases (HDAC); however, we noted that among this group, HDAC1 is the most significantly upregulated in highly metastatic cells generally, and in LS174T cells compared with SW480 cells in particular (Fig. 7C and D; Supplementary Fig. S7C). Moreover, we observed that RBMS1 and HDAC1 expression are negatively correlated in colon cancer clinical samples (Supplementary Fig. S7D). Furthermore, we observed that silencing HDAC1 increases RBMS1 expression (Fig. 7E; Supplementary Fig. S7E and S7F). Increased expression of HDAC1, and other class I HDAC genes, has been reported as a strong predictor of survival in patients with colon cancer (26), and modulations in HDAC activity result in widespread gene expression reprogramming, including repression of tumor suppressor genes (27). Our observations here indicate that HDAC1-mediated transcriptional repression may be one possible mechanism for RBMS1 silencing in colon cancer cells.

• Download figure
• Open in new tab
• Download powerpoint

Figure 7.

HDAC-mediated promoter deacetylation results in RBMS1 silencing. A, RBMS1 shows dynamic expression and acetylation changes across different cell types. Also shown here is the association between RBMS1 expression and its promoter acetylation (source from ENCODE). B, RBMS1 promoter acetylation levels were measured in SW480 and LS174T cells using H3K27Ac ChIP-qPCR. An unacetylated region approximately 40 kb away from the promoter was used as control (N = 3). C, HDAC1 expression in colon cancer lines stratified on the basis of their metastatic capacity. One-tailed U test was used to compare the two groups. Cell line names are listed in Supplementary Fig. S1A. D, Quantitative PCR to compare HDAC1 levels in the highly metastatic LS174T cells (with silenced RBMS1) relative to poorly metastatic SW480 cells. E, qRT-PCR was used to measure RBMS1 mRNA levels in LS174T cells with RNAi-mediated HDAC1 silencing. F, A schematic model of RBMS1 silencing and its role in suppressing colon cancer metastasis. One-tailed Mann–Whitney U tests were used to assess statistical significance for all panels.