Identifying novel susceptibility genes for colorectal cancer risk from a transcriptome-wide association study of 125,478 subjects

Abstract

Background and Aims:

Susceptibility genes and the underlying mechanisms for the majority of risk loci identified by genome-wide association studies (GWAS) for colorectal cancer (CRC) risk remain largely unknown. We conducted a transcriptome-wide association study (TWAS) to identify putative susceptibility genes.

Methods:

Gene-expression prediction models were built using transcriptome and genetic data from the 284 normal transverse colon tissues of European descendants from the Genotype-Tissue Expression (GTEx), and model performance was evaluated using data from The Cancer Genome Atlas (TCGA, n = 355). We applied the gene-expression prediction models and GWAS data to evaluate associations of genetically predicted gene-expression with CRC risk in 58,131 CRC cases and 67,347 controls of European ancestry. Dual-luciferase reporter assays and knockdown experiments in CRC cells and tumor xenografts were conducted.

Results:

We identified 25 genes associated with CRC risk at a Bonferroni-corrected threshold of P < 9.1 × 10⁻⁶, including genes in four novel loci, PYGL (14q22.1), RPL28 (19q13.42), CAPN12 (19q13.2), MYH7B (20q11.22), and MAP1L3CA (20q11.22). In nine known GWAS-identified loci, we uncovered nine genes that have not been previously reported, whereas four genes remained statistically significant after adjusting for the lead risk variant of the locus. Through colocalization analysis in GWAS loci, we additionally identified 12 putative susceptibility genes that were supported by TWAS analysis at P < 0.01. We showed that risk allele of the lead risk variant rs1741640 affected the promoter activity of CABLES2. Knockdown experiments confirmed that CABLES2 plays a vital role in colorectal carcinogenesis.

Conclusion:

Our study reveals new putative susceptibility genes and provides new insight into the biological mechanisms underlying CRC development.

Lay Summary

This large-scale transcriptome-wide association study (TWAS) has revealed 25 putative susceptibility genes, including five in novel loci and an additional nine in GWAS loci that have not been previously reported.

Introduction

Genetic factors play an important role in the etiology of both sporadic and familial colorectal cancer (CRC). Multiple CRC susceptibility genes, including APC, MUTYH, MLH1, PMS2, MSH2, MSH6, PTEN, SMAD4, BMPR1A, and MYH, have been identified as being responsible for family CRC syndromes, such as Lynch Syndrome^1–3. In addition to these known high-penetrance genes, approximately 150 common genetic variants have been found to be associated with CRC risk through genome-wide association studies (GWAS)^4–21. However, together, these variants explain only a small proportion of the familial relative risk of CRC^18–21. Thus, additional loci for genetic susceptibility for CRC remain to be identified. Furthermore, the target genes at most of the GWAS-identified risk loci remain largely unclear.

Most GWAS-identified risk variants are located in non-coding or intergenic regions without clear evidence of regulatory epigenetic signals. However, it has been shown that many functional variants in strong linkage disequilibrium (LD) with these variants play regulatory roles in gene expression^22–26. Large genomics consortia, including the Genotype-Tissue Expression (GETx, see URLs) and The Cancer Genome Atlas (TCGA, see URLs) projects, have generated massive quantities of high-dimensional genomics data, including data from matched genetic and transcriptome profiles from hundreds of normal and tumor tissues of the colon and rectum. These data are valuable for expression quantitative trait loci (eQTL) analyses, which evaluate the association of SNP genotypes with gene-expression levels. Using an eQTL approach, previous studies, together with our work, have revealed approximately 30 putative susceptibility genes linked to GWAS-identified SNPs (refer to index SNPs) or lead SNPs (refer to SNPs with the best association signals)^27–31. In addition to the eQTL approach, the advance of chromatin-chromatin interaction technology has produced large amounts of chromatin-chromatin interaction data in various normal and cancer-cell lines, including CRC cell lines^32–35.

Recently, transcriptome-wide association studies (TWAS)³⁶ have been initiated to systematically investigate the transcriptome for disease risk. First, prediction models are built to predict gene expression with cis-SNPs using a reference transcriptome. Then they are applied to GWAS data to evaluate the association of the predicted gene expression with disease risk. A recently performed TWAS, using approximately 27,000 cases and controls, reported several genes of which the predicted expressions were associated with CRC risk^{37, 38}. However, the sample size of that study was relatively small, and several reported associations were not statistically significant after Bonferroni-correction. Herein, we report results from a large CRC TWAS, conducted among 125,478 subjects of European descent, to comprehensively search for susceptibility genes.

Methods

GWAS data from study cohorts

The study utilized GWAS summary statistics data from 125,478 individuals of European descent (58,131 CRC cases and 67,347 controls) from six cohort studies. The cohort studies include GECCO (cases: 11,958, controls: 14,740), CORECT (cases: 22,911, controls: 14,311), CROC (cases: 12,007, controls: 12,000), CONE (cases: 4,439, controls: 4,115), CORSA (cases: 1,460, controls: 774), and UKBio (cases: 5,356, controls: 21,407). All participants provided written informed consent, and each study was approved by the relevant research ethics committee or institutional review boards. Details on sample selection and matching, sample numbers and demographic characteristics of study participants have been described previously¹⁸.

Gene-expression model building using GTEx

The recent GTEx release 8 includes RNA sequencing data of transverse colon samples obtained from 284 subjects and whole genome sequencing data from these subjects. Gene-expression and WGS data from these samples were processed according to the GTEx protocol, as described in our previous work^39–41. For gene-expression data, expression levels of each gene have been measured using reads per kilobase per million (RPKM) units from RNA-SeQC. We performed data quality control (QC) and normalization processing by filtering lowly-expressed genes, log2 transforming and Robust Multichip Average (RMA). We further performed rank-based inverse normal transformation for gene expression across all samples. We performed a probabilistic estimation of expression residuals (PEER) analysis to generate the top PEER factors (top 15 PEERs used in this study) to adjust batch and other potential confounding factors⁴² for downstream prediction model building. Only common SNPs (minor allele frequency > 0.05) with ‘PASS’ tags were included. SNPs with a call rate < 98%, with a Hardy-Weinberg equilibrium P value < 10⁻⁶ (among subjects of European ancestry) or showing batch effects were excluded. Principal component analysis (PCA) was conducted using EIGENSTRAT⁴³ (see URLs) to generate top PCs from the genotype data.

Genetic and transcriptome data from the 284 transverse colon samples from GTEx were used to build gene-expression prediction models for this study. We trained the gene-expression prediction model by flanking genetic variants (flanking ±1Mb region) using an elastic-net approach, implemented in the MetaXcan tool (Formula 1).

$$\text{Y=}\text{G}\text{+}\text{PC}\text{+}\text{PEERs}\text{+}\epsilon$$ (1)

In Formula 1, for each gene, Y is the expression level, G is the number of effect alleles (0–2) for each genetic variant, with an adjustment for top PCs, sex, age, potential batch effects, and other potential confounding factors (PEERs), and ε is the random error. We focused on the cis-regulation of a gene predicted by local genetic variants within 2 Mb flanking the gene region. The parameters of the model for each gene were assessed using tenfold cross-validation, and the correlation (R²) between predicted and observed gene-expression levels was used to evaluate the prediction performance.

Evaluating prediction performance of gene-expression models using TCGA

To further evaluate the prediction performance of these genes in GTEx externally, we downloaded TCGA Colorectal Adenocarcinoma data, including data from gene expressions (RNA-seq V2, measured by median Zscore), DNA methylations (Level 3, Infinium HumanMethylation450), and SCNAs (data_linear_CNA.txt) from cBioPortal (see URLs). We also downloaded Level 3 SNP data (COAD and READ) that were genotyped using the Affymetrix SNP 6.0 array from TCGA’s data portal. A total of 355 tumor samples with gene expressions, DNA methylations, SCNA and SNP data from TCGA were used to evaluate the prediction model. The processing protocol for genotype and gene-expression data was performed for the data from TCGA as described in our previous work^{25, 31}. To evaluate the prediction performance for each gene predicted by a set of SNPs in GTEx, we first constructed a residual model for gene expression as a dependent variable by adjusting DNA methylation, SCNA and the first five PCs derived from the genotype data. We further used the derived residuals to construct a linear regression model with the genotype data from the set of SNPs. The prediction performance of the model for each gene was assessed by the adjusted R² for the linear model, including all genetic predictors determined in GTEx data. The prediction performance of gene-expression models has been summarized in Supplementary Table 1.

Association analyses between predicted gene expression and CRC risk

To identify susceptibility genes for CRC risk, we applied the weight matrix to the summary statistics data on SNPs from CRC GWAS datasets using the MetaXcan tool⁴⁴. The MetaXcan method, which has been described elsewhere^{45, 46}, was used for the association analyses.

$$Z_g\approx {\sum }_{l\in {\text{Model}}_g}{w}_{lg}\frac{{\widehat{\sigma }}_l}{{\widehat{\sigma }}_g}\frac{{\widehat{\beta }}_l}{\text{se}\left({\widehat{\beta }}_l\right)}$$ (2)

In Formula 2, the Z-score was used to estimate the association between predicted gene expression CRC risk. Here, w_lg is the weight of SNP l for predicting the expression of gene g. ${\widehat{\beta }}_l$ and se $\left({\widehat{\beta }}_l\right)$ are the association regression coefficients, and its standard error for SNP l in GWAS, and ${\widehat{\sigma }}_l$ and ${\widehat{\sigma }}_g$ are the estimated variances of SNP l and the predicted expression of gene g, respectively. For this study, we estimated the correlations between SNPs included in the prediction models using the phase 3, 1000 Genomes Project data of European populations. The summary statistics were analyzed based on a meta-analysis of GWAS data among 125,478 individuals of European descent from six cohort studies. An association result was included in the meta-analysis only if the estimated minor allele count (MAC) was > 50 and the imputation quality metric Rsq >0.3. Notably, additional QC steps for SNPs and samples have been performed in previous GWAS by the CRC consortia¹⁸.

For our TWAS analyses, a Bonferroni-corrected P < 9.1 ×10⁻⁶ (0.05/5,491) was used to identify a statistically significant association.

Conditional and colocalization analyses of GWAS association signals

To determine whether the identified associations between genetically predicted gene expression and CRC risk were influenced by association signals identified in GWAS, we conducted conditional analyses by adjusting for index SNPs. We followed a previously reported approach⁴⁷ to estimate adjusted odds ratios and their standard errors for the association between selected SNPs and CRC risk, given the reported index SNP. Then we re-ran the MetaXcan analyses using the adjusted summary statistics.

We downloaded cis-eQTL results from normal transverse tissues in the GTEx database (Version 8). Colocalization analysis of GWAS association signals and expression of TWAS-identified genes was conducted using summary data–based Mendelian randomization (SMR)⁴⁸ based on the eQTL results from the GTEx and the GWAS summary statistics data. The genotype data from the European populations of the 1000 Genomes Project Phase 3 was used for the LD estimation. The significant colocalized signals was determined based on the nominal threshold of P_SMR < 0.05.

Results

Gene expression predicted by flanking genetic variants

We used transcriptome and genotype data from the GTEx to build gene-expression prediction models for normal transverse colon tissues from European descendants (N = 284) (see Methods). A total of 16,082 models were built for genes predicted by flanking genetic variants (flanking ±1Mb region) using the elastic net approach^{36, 49}. Our results showed that the expression levels of 9,503 genes could be predicted by local genetic variants at R² > 0.01 (10% correlation), with a median of 27 variants per gene (See Supplementary Table 1). To evaluate the prediction performances for each gene predicted by a set of SNPs in GTEx, we constructed a residual model for gene expression as a dependent variable by adjusting DNA methylation, somatic copy number alteration (SCNA) and the first five principal components (PCs) using data from 355 primary CRC tissues from TCGA. We further used the derived residuals to construct a linear regression model with the genotype data from the same set of SNPs (see Methods). After removing poorly predicted genes, we focused on 5,491 genes for a downstream association analysis between predicted gene expression and CRC risk. These genes had either 1) a model prediction R² > 0.01 in the GTEx and replicated by TCGA, or 2) a R² > 0.0625 (25% correlation) in the GTEx regardless of TCGA; this second set included additional genes that could not be evaluated in TCGA due to a lack of data (See Supplementary Table 1).

Identified CRC susceptibility genes in TWAS

We evaluated the associations of the predicted genes with CRC risk using the MetaXcan tool⁴⁴, based on the weights of the SNPs in the expression-prediction models from GTEx, and their summary statistics from GWAS data (N = 125,487; see Methods). In total, we identified 25 genes with genetically predicted expressions associated with CRC risk at a Bonferroni-corrected threshold of P < 9.1 × 10⁻⁶ (Fig. 1 and Supplementary Table 4). Of those, five genes, PYGL (14q22.1), RPL28 (19q13.42), CAPN12 (19q13.2), MYH7B (20q11.22), and MAP1L3CA (20q11.22) are located in four independent loci at least 3 Mb away from any previously reported GWAS-identified locus for CRC risk, suggesting that they may be novel CRC susceptibility loci. Specifically, low predicted expression levels of all of these genes were associated with an increased risk of CRC (Table 1). Using a less stringent threshold at a P < 6.6 × 10⁻⁴, a false discovery rate (FDR)-corrected significance level, we identified an additional 48 genes with genetically predicted expressions associated with CRC risk (Fig. 1 and Supplementary Table 5).

Figure 1. Manhattan plot of the association results from the TWAS of colorectal cancer (N = 125, 478).

The blue and red lines represent a false discovery rate (FDR)-corrected significance level of P < 6.6 × 10⁻⁴ and a Bonferroni-corrected threshold of P < 9.1 × 10⁻⁶, respectively.

Table 1.

Association of colorectal cancer risk with predicted expression for six genes located at least 1MB away from any GWAS reported colorectal cancer risk variants.

Locus	Gene	Z score	P valuea	R2 (GTE×)b	R2 (TCGA)b	Lead SNPc	Distance (Mb)c
14q22.1	PYGL	−4.60	4.20 × 10−6	0.20	0.07	rs35107139	3.0
19q13.2	CAPN12	−4.51	6.24 × 10−6	0.11	0.04	rs28840750	5.7
19q13.4	RPL28	−5.03	4.86 × 10−7	0.23	0.04	rs73068325	3.2
20q11.22d	MYH7B	−4.78	1.72 × 10−6	0.13	0.02	rs6031311	9.1
20q11.22d	MAP1LC3A	−4.94	7.66 × 10−7	0.13	<0.01	rs6031311	9.5

^“a”P value derived from association analyses in TWAS; Statistically significant based on a Bonferroni-corrected threshold of P < 9.1 × 10⁻⁶ from 5,491 tests (0.05/5,491);

^“b”Prediction performance (R²) was derived using data from GTEx, while adjusted R² was derived from TCGA data for validation of a gene expression prediction model. “−” refers to a model in GTEx that could not be validated in TCGA due to a lack of data.

^“c”Distance between a gene with the closest lead SNP identified from previous CRC GWAS.

^“d”The initial GWAS-identified locus (rs6058093) was excluded in the recent meta-analysis among 125,478 subjects of European descent¹⁸.

To determine whether the genes identified were implicated in the associations of established GWAS-identified risk variants, we investigated 20 (of the total of 25) genes located in 16 established independent CRC risk loci (Table 2). We performed conditional analyses for the associations between CRC risk and these genes, adjusting for the associations with the closest lead SNP for each locus (see Methods). We showed that an association with genesCCHCR1 (rs2516420, in a distance of 323kb), LRP1 (rs4759277, 0 kb), RGS9BP (rs28840750, 350kb), and PARD6B (rs6063514, 293kb) remained statistically significant at a P < 2.5 × 10⁻³ (with multiple comparisons corrections of 0.05/20) after adjusting for the closest risk SNP (Table 2). Of the remaining 16 genes, we observed that the expression prediction models for five of these, SFMBT1 (rs9831861), COLCA1 (rs3087967), COLCA2 (rs3087967), AL121832.2 (rs1741640), and AL121832.3 (rs1741640), included the lead SNPs, or SNPs in strong LD (R² >0.95) with them. This supports the hypothesis that the observed association for these genes may be driven by these lead SNPs (Table 2). Further colocalization analysis between the expression of all TWAS-identified genes and previous GWAS association signals using the SMR approach showed that the effects of eight lead SNPs on cancer risk may be mediated by expression of TWAS-identified genes, including SF3A3, ACTR1B, AC021016.1, SFMTB1, AC004847.1, C11orf53, COLCA1, COLCA2, DACT1, AL121832.2 and AL121832.3 (Supplementary Table 5). The remaining nine genes have not been previously reported.

Table 2.

Association of colorectal cancer risk with predicted expression for 15 genes located in the regions within 500kb from GWAS reported colorectal cancer risk variants.

Locus	Gene	Z score	P valuea	R2 (GTE×)b	R2 (TCGA)b	Lead SNPc	Distance (kb)c	P value after adjusting for lead SNPd
1p34.3	SF3A3	4.99	5.99 × 10−7	0.13	0.01	rs4360494	144	0.94
2q11.2	ACTR1B	−4.8	1.62 × 10−6	0.28	0.02	rs11692435	0	0.03
2q35	AC021016.1	5.77	7.90 × 10−9	0.52	-	rs35470271	27,542	0.28
2q35	ARPC2	4.49	7.10 × 10−6	0.02	0.01	rs3731861	72,177	0.90
3p21.1	SFMBT1	4.61	4.03 × 10−6	0.20	0.15	rs9831861	7,519	-
6p21.3	ATP6V1G2	−4.75	2.00 × 10−6	0.06	0.08	rs2516420	62,619	7.6 × 10−3
6p21.33	CCHCR1	5.24	1.61 × 10−7	0.46	<0.01	rs2516420	323,605	8.78 × 10−7
7p13	AC004847.1	−4.46	8.16 × 10−6	-	-	rs12672022	135,915	0.65
11q23.1	C11orf53	−11.66	1.94 × 10−31	0.21	0.06	rs3087967	0	7.4 × 10−3
11q23.1	COLCA1	−10.16	2.94 × 10−24	0.24	0.01	rs3087967	4,676	-
11q23.1	COLCA2	−11.25	2.31 × 10−29	0.36	0.06	rs3087967	12,444	-
12q13.12	AC074032.1	−5.35	8.82 × 10−8	0.19	-	rs12372718	611,689	0.86
12q13.3	LRP1	5.25	1.50 × 10−7	0.16	0.01	rs4759277	0	4.47 × 10−4
12q24.12	ALDH2	−4.48	7.47 × 10−6	0.04	0.02	rs597808	231,333	0.40
14q23.1	DACT1	−5.42	6.04 × 10−8	0.06	<0.01	rs17094983	74,322	0.09
19q13.11	RGS9BP	−5.02	5.22 × 10−7	0.02	0.02	rs28840750	350,721	1.12 × 10−5
19q13.43	AC016629.3	−4.82	1.45 × 10−6	0.13	-	rs73068325	7,716	0.13
20q13.13	PARD6B	5.07	4.02 × 10−7	0.07	0.01	rs6063514	292,799	1.93 × 10−6
20q13.33	AL121832.2	9.68	3.84 × 10−22	0.23	-	rs1741640	28,945	-
20q13.33	AL121832.3	−8.59	8.99 × 10−18	0.11	-	rs1741640	44,878	-

^“a”P value derived from association analyses in TWAS; Statistically significant based on a Bonferroni-corrected threshold of p < 9.1 × 10⁻⁶ from,5,491 tests (0.05/5,491).

^“b”prediction performance (R²) was derived using data from GTEx, while adjusted R² was derived from TCGA data for validation of a gene-expression prediction model. “−” refers to a model in GTEx that could not be validated in TCGA due to a lack of data.

^“c”Distance between a gene with the closest lead SNP identified from previous CRC GWAS.

^“d”P value derived from association analyses in TWAS after adjusting for the association of the closest lead SNP. “−” refers to the lead SNP that has been included in the model or is in strong LD (R² > 0.95) with a SNP in the model that showed the best association signal with CRC risk.

^“e”The genes in bold refer to those reported from the previous eQTL analysis in CRC and the colocalization analysis in this study.

In addition, we systemically evaluated eQTL results in the established GWAS loci^18,23 using data from the GTEx and identified 42 eQTL genes for 19 lead variants at a Bonferroni-corrected threshold of P < 2.4 × 10⁻⁵ (Supplementary table 6). Further colocalization analysis showed that the CRC susceptibility for 18 leading variants may be mediated by cis-effects on gene regulation, whereas 38 genes were involved (Supplementary Table 6). In addition to the genes identified in the above TWAS, we observed that an additional 12 genes, including ASAH2B, ATF1, CABLES2, HCG20, HLA-DQA1, HLA-DQA2, HLA-DQB1, HLA-DRB5, PRRT1, TRPC6, ZNF132, and ZNF584 were supported by TWAS analysis at a less stringent threshold of P < 0.01 (Supplementary Table 6).

Putative susceptibility genes supported by functional genomic data analysis

To search for additional evidence for the 25 putative susceptibility genes identified from TWAS, we explored regulatory mechanisms for putative functional variants correlated with the SNP that has the strongest association with CRC risk in the prediction model. We performed functional annotations of the variants in strong LD (R² > 0.8) with these SNPs (Supplementary Table 7; see Supplemental Notes). We analyzed 349 putative functional variants (in strong LD with these SNPs), which showed evidence of the epigenetic signals from a data analysis of Roadmap (see URLs, Supplementary Table 7; see Supplemental Notes). Specifically, a total of 118 variants were mapped to promoter regions, whereas the remaining 231 variants were mapped to enhancer regions. To search for direct evidence that variants regulate the putative susceptibility genes identified from our TWAS analysis, we first examined if the potential functional variants were positioned in proximal promoter regions, as such variants would most likely play a regulatory role in relation to the closest genes. The 13 genes, C11orf53, COLCA1, COLCA2, AC074032, ALDH2, ARPC2, CAPN12, LRP1, PARD6B, PYGL, RPL28, SF3A3, and SFMBT1, were likely regulated by nearby putative functional SNPs with promoter activities (Supplementary Table 7). We further examined whether the genes could be regulated by putative functional variants located in enhancer regions via long-distance promoter-enhancer interactions by analyzing chromatin-chromatin interaction data. An additional six genes, AC004847.1, AC074032.1, MAP1LC3A, ACTR1B, AL121832.2, and DACT1, showed evidence of potential regulation via distal promoter-enhancer interaction. Taken together, a total of 19 of the original 25 identified genes (76%) showed evidence of regulation by putative functional variants via proximal promoter or distal enhancer-promoter interactions. All of them were further supported with the evidence of their potential functional variants located in regions of histone modifications, DNase I hypersensitive or TF binding sites in normal colorectal epithelium and CRC cell lines (Supplementary Table 8).

Functional assays for putative tumor suppressor CABLES2

Previous studies have indicated that the CABLES1 gene may be involved in cell growth and differentiation^{50, 51}, and it appears to function as a tumor suppressor by inhibiting CRC formation and growth^51–55. In line with these previous findings, our results showed that a lower predicted expression of CABLES2 may mediate the effect of the lead SNP rs1741640 on an increased risk of CRC. A functional annotation of rs1741640 indicates that this SNP is located in a region with promoter activity. To confirm whether the rs1741640 can regulate CABLES2 expression, we conducted luciferase reporter assays for rs1741640 and another SNP, rs477859, in strong LD with rs1741640 in both HCT116 and RKO cell lines (see Supplemental Notes). Our results showed that the fragment containing the risk allele (C, rs1741640) significantly decreased the promoter activity of CABLES2 compared to the reference allele T in both cell lines (Fig. 2A). The observation was in line with the results from both colocalization and TWAS analyses (Supplementary Table 6). For rs477859, we did not observe that the fragment containing the alternative allele significantly affect the promoter activity compared to the reference allele (Data not shown).

Figure 2. Dual-Luciferase reporter assay and CRC cell proliferation, colony formation, migration and invasion for the CABLES2 gene.

A) Boxplots showing that alternative alleles (CC) of functional SNP rs1741640 can affect promoter activity of CABLES2 compared to reference alleles (TT), based on the results using luciferase reporter assays in both the HCT116 and RKO cell lines. The yellow and black boxes represent the sequence of CABLES2 promoter and the enhancer elements containing the candidate functional SNP individually. The error bars represent the standard deviation of promoter activities. A paired t test was performed to derive p value. B) The virus infection efficiency and C) the knockdown efficiency were, respectively, verified by RFP fluorescence and qPCR in CABLES2 shRNA treated (sh1 and sh2) and vehicle control cells. After SW480 cells were transfected with CABLES2 shRNA, the cell colonies were D) imaged, E) quantified and F) given a ratio using a colony formation assay for the CABLES2 shRNA and vehicle control cells. G) Cell viabilities were measured using CCK8 assays at different time points (Days 1, 2, 3, 4, 5) and H) migration and invasion were measured using Transwell assays for CABLES2 shRNA and vehicle cells. P-values were determined by a t-test from the comparison of knockdown and control cells. “*”, P < 0.05; “**”, P < 0.01; “***”, P < 0.001.

We further performed in vitro functional assays in CRC cell lines to investigate the cellular function of CABLES2. qPCR experiments were conducted to compare its relative expression level in three CRC cell lines (SW480, RKO and HCT116). CABLES2 showed higher expression levels in both the SW480 and HCT116 cell lines when compared to the RKO cell line (Supplementary Fig. 1). We designed knockdown experiments in the SW480 and HCT116 cell lines using short hairpin RNA (shRNA). We employed packaged lentivirus from lentiviral expression vectors containing the Red Fluorescent Protein (RFP) reporter gene with open reading frames (ORF) and a fragment from CABLES2 shRNA-1, shRNA-2, or a control vehicle shRNA. Using the RFP fluorescence reporter system, we observed high transfection efficiencies for the control vehicles, shRNA-1 and shRNA-2, in the SW480 cells (Fig. 2B). In SW480 cells, the shRNA-1 virus infection was able to reduce the endogenous mRNA level by approximately 90%, whereas shRNA-2 decreased the mRNA level by 40% in SW480 cells (Fig. 2C). After performing the CABLES2 shRNA-1 treatment, colony formation was significantly enhanced over that of the control vehicle; CABLES2 shRNA-2 also led to a mild increase in colony formation (Fig. 2D–F; P < 0.001). Cell viabilities after the CABLES2 shRNA-1 or shRNA-2 treatment were significantly increased over those of the control vehicle in the CCK8 assay (P < 0.01 for both; Fig. 2G). Further, we observed that shRNA-1 significantly increased cell migration and invasion when compared to the control vehicle; shRNA-2 also led to an increase of cell invasion in the Transwell assay in the SW480 and HCT116 cell lines (Fig. 2H and Supplementary Fig. 2).

To explore potential genes and pathways regulated by CABLES2, we conducted RNA-sequencing in the CABLES2 shRNA-1 and control vehicle SW480 cells (see Supplemental Notes)). We identified 169 differentially expressed genes, which included 85 up-regulated and 84 down-regulated genes, with CABLES2 shRNA-1, compared to control vehicle cells (Supplementary Fig. 3A; see Supplemental Notes). An enrichment analysis using the Kyoto Encyclopedia of Genes and Genomes (KEGG, see URLs) Pathways revealed that these differentially-expressed genes were significantly enriched in thePI3K/AKT cancer pathway (P = 7.2 × 10⁻³). A total of 10 genes were involved in that pathway, including nine up-regulated genes: TNC, ANTXR2, IL7R, CHN2, SGK1, TLR4, FN1, MET and ARHGAP29. The presence of only one down-regulated gene, IL2RG, suggests that the silenced CABLES2 can promote the PI3K-AKT signaling pathway (Supplementary Fig. 3B). We further validated these expression changes for the top seven genes using qPCR in the cells (Supplementary Fig. 3C).

To confirm that CABLES2 is capable of inhibiting the growth of colonic carcinoma, we developed tumor xenografts by the subcutaneous injection of SW480 CABLES2 knockdown and control shRNA cells in nude mice (see Supplemental Notes). We showed that the injected CABLES2 knockdown cells markedly increases tumor volume and weight when compared with control cells (Fig. 3A–C). To analyze cell proliferation in tumor tissue, Ki67 staining was conducted in tumor tissue from nude mice that had been injected with knockdown and control cells. We observed that CABLES2 knockdown cells increase Ki67 foci when compared with control cells (Fig. 3D); 77% of Ki67-positive cells were observed in knockdown cells compared to 35% in control cells (Fig. 3E).

Figure 3. CABLES2 inhibiting growth of colonic carcinoma in nude mice.

A) Photograph of tumor formation in nude mice after injection of cells at day 48. B) Tumor volume was recorded by a measurement of tumor length, height and width every 3–5 days during tumor proliferation, and calculated as length × width × height/2. C) Quantification of tumor weight at day 48 after the anesthetic execution of mice; D-E) Ki67 staining of tumor tissue and quantification of Ki67-positive cells in knockdown and control cells.

Discussion

Utilizing data from a large-scale GWAS collaboration of GTEx and TCGA, we conducted a TWAS analysis to search for susceptibility genes for CRC risk. We disclosed 25 putative CRC predisposition genes, including six novel genes located in the regions far away from any previously reported susceptibility loci for CRC risk, and nine novel genes as targets for established CRC GWAS loci. We have also identified an additional 48 genes with genetically predicted expressions associated with CRC risk using a relax threshold at FDR < 0.05. These genes can be further verified by additional future large-scale genetic studies. These findings greatly increase the number of potential susceptibility genes identified for CRC risk. On the other hand, we also identified 38 putative target genes, including CABLES2, for the established GWAS loci through the colocalization analysis, whereas a majority of genes were supported by TWAS analysis. Using both in vitro and in vivo functional assays, we further showed that the putative tumor suppressor CABLES2 plays a vital role in CRC tumorigenesis. Our findings provide novel insight into the genetic and biological basis for CRC development.

Our identification of these 25 putative susceptibility genes is supported by several lines of additional evidence. For example, for 76% of them, we observed evidence of cis-regulation by putative functional risk SNPs via proximal promoter or distal enhancer-promoter interactions. These data suggest a possible link of a gene with a potential regulatory SNP. Functional assays, such as luciferase reporter assays in CRC cell lines are needed to establish the underlying regulatory mechanisms, as demonstrated by CABLES2 (rs1741640) in our study. Furthermore, 11 of these genes have been implicated as putative target genes for lead SNPs based on previous eQTL and our current colocalization analysis analyses^27–31 (Table 2). A particular lead SNP may be a surrogate for multiple variants for cancer risk in the locus, while some target genes for potential causal variants, which are in weak LD with index SNPs, may not be detected by an eQTL analysis. In addition, we conducted a differential gene expression analysis for the genes identified in TWAS among normal colon mucosa, adenoma and adenocarcinoma using publicly available gene expression data, including 135 normal colon mucosa, 363 colon adenoma and 2,760 colon adenocarcinoma (Supplementary Table 5 and 6). Of the 25 genes identified in TWAS, we observed seven genes, including DACT1, CCHCR1, PARD6B, MYH7B, SFMBT1, PYGL, and ALDH2, that showed significant differential expression between carcinoma and adenoma tissues at P < 0.05. For three genes, MYH7B, PYGL, and ALDH2, we observed that low predicted expression levels were associated with an increased risk of CRC and showed lower expression in carcinoma than in adenoma. For the remaining four genes where high predicted expression levels were associated with an increased risk of CRC, they showed higher expression in carcinoma than in adenoma, except for the gene DCAT1 (Supplementary Table 5). The results provide additional evidence to support a possible role of these genes in tumorigenesis. However, we did not observe the same pattern when we compared adenoma with mucosa. This may be due to their potential oncogenic role in the initiation of carcinogenesis in adenoma stage. It should be noted that more than half of our identified genes, including, RPL28⁵⁶, CCHCR1^{57, 58}, PARD6B⁵⁹, DACT1^{60, 61}, LRP1^62–64, SF3A3⁶⁵, MAP1LC3A⁶⁶, ACTR1B⁶⁷, ATP6V1G2⁶⁸, PYGL^{69, 70}, ARPC2^{71, 72}, ALDH2^{73, 74}, KCNQ1⁷⁵, SFMBT1⁷⁶, have been implicated in cancer-driver events or cancer-related functions in in vitro or in vivo functional experimental studies in CRC or other cancer types.

In our gene-expression prediction model building from GTEx, although we used a relatively low cutoff of prediction performance at R² > 0.01 (10% correlation) for downstream TWAS analysis, all of the reported genes can still be well predicted with a minimal prediction performance at R² > 0.04. In addition, we used tumor tissues from TCGA to evaluate the prediction performance of gene-expression built in GTEx. Although we performed additional analyses to adjust for possible confounders of gene expression in tumor tissues of TCGA, large independent normal tissues samples are desirable for the evaluation of the prediction performance of gene-expression. For our prediction model, the reference panels were built using only the data from European populations. Both GWAS and transcriptome data from GTEx are currently limited for other racial groups, preventing us from evaluating these susceptibility genes in non-European populations. Additional efforts are still needed to generate transcriptome data for TWAS in those populations.

Although many of the genes we identified have been implicated previously in CRC susceptibility genes, we provided additional evidence to support their role. Further, we have identified 15 genes not reported previously for CRC susceptibility, including six genes located in the regions not yet identified by GWAS. We performed functional genomic experiments for one particular gene, CABLES2, to demonstrate that some genes identified in TWAS are functionally important in colorectal carcinogenesis. Further in vitro and in vivo experiments will be needed to firmly establish a causal association between the other reported genes and CRC risk.

Background and Context.

Large-scale transcriptome-wide association studies (TWAS) to identify genetic risk loci and susceptibility genes for colorectal cancer (CRC) are lacking.

New Findings.

Five genes in four novel loci were identified to be associated with CRC risk in individuals of European descent, along with an additional nine genes in known GWAS loci that have not been previously reported.

Limitations.

Further functional assays are requested to establish the underlying regulatory mechanisms for the risk genetic variants and susceptibility genes identified in this study.

Impact.

Identification of CRC risk loci and susceptibility genes can help identify individuals at high risk. These findings have significant implications for genetic screening.