Using DEPendency of association on the number of Top Hits (DEPTH) as a complementary tool to identify novel colorectal cancer susceptibility loci

Abstract

Background:

DEPendency of association on the number of Top Hits (DEPTH) is an approach to identify candidate susceptibility regions by considering the risk signals from overlapping groups of sequential variants across the genome.

Methods:

We conducted a DEPTH analysis using a sliding window of 200 SNPs to colorectal cancer (CRC) data from the Colon Cancer Family Registry (CCFR) (5,735 cases and 3,688 controls), and GECCO (8,865 cases and 10,285 controls) studies. A DEPTH score >1 was used to identify candidate susceptibility regions common to both studies. We compared DEPTH results against those from conventional GWAS analyses of these two studies as well as against 132 published susceptibility regions.

Results:

Initial DEPTH analysis revealed 2,622 (CCFR) and 3,686 (GECCO) candidate susceptibility regions, of which 569 were common to both studies. Bootstrapping revealed 40 and 49 candidate susceptibility regions in the CCFR and GECCO data sets, respectively. Notably, DEPTH identified at least 82 regions that would not be detected using conventional GWAS methods, nor had they been identified by previous CRC GWASs. We found four reproducible candidate susceptibility regions (2q22.2, 2q33.1, 6p21.32, 13q14.3). The highest DEPTH scores were in the HLA locus at 6p21 where the strongest associated SNPs were rs762216297, rs149490268, rs114741460, and rs199707618 for the CCFR data, and rs9270761 for the GECCO data.

Conclusion:

DEPTH can identify candidate susceptibility regions for CRC not identified using conventional analyses of larger datasets.

Impact:

DEPTH has potential as a powerful complementary tool to conventional GWAS analyses for discovering susceptibility regions within the genome.

Introduction:

Colorectal cancer (CRC) is the third most diagnosed cancer in the Western World for both men and women, accounting for the third highest number of cancer-related deaths (1,2). Notably, CRC mortality has been decreasing for the past few decades (1,2) driven in part by higher uptake of screening and the removal of polyps (3–6). Because early intervention is an effective tool, many national guidelines recommend that CRC screening start from age 45-50 years. Identifying individuals with a high risk of developing CRC could further reduce CRC mortality as these individuals might benefit from having earlier or more frequent CRC screening.

One CRC risk factor of increasing interest is genetic susceptibility. There is a strong familial component underlying CRC risk which, if genetic, appears to be driven by major susceptibility genes and polygenes (7). Genome wide association studies (GWASs) have now identified 132 loci that are reliably ‘known’ to be associated with CRC risk for individuals of European ancestry (8–12). Conventional GWASs typically assess the correlations of a single nucleotide polymorphisms (SNPs) with a phenotype independently, and then adjust for multiple testing to identify SNPs for which their association with the phenotype is considered to be ‘genome-wide significant’.

We developed a method called DEPendency of association on the number of Top Hits (DEPTH) to identify genomic susceptibility regions (loci). Risk associations are calculated for overlapping regions that contain multiple SNPs, rather than for each SNP individually (13). Notably, we have applied DEPTH analyses to a prostate cancer GWAS data set and identified 112 novel candidate prostate cancer susceptibility regions (14).

Here, we have applied DEPTH to try to identify novel CRC susceptibility regions that might not be detected by conventional GWAS approaches. We compare our DEPTH results with the results from conventional GWAS analyses of the same data sets, as well with the known CRC susceptibility loci. We report in this study how DEPTH can identify, with high confidence, novel candidate susceptibly loci, and propose that this method could be a useful complementary tool for the discovery of true susceptibility regions.

Materials and Methods:

Study subjects.

DEPTH analyses were carried out on two independent CRC data sets—the Colon Cancer Family Registry (CCFR) data set, and the Genetics and Epidemiology of CRC Consortium (GECCO) data set. The initial CCFR data set comprises 11,489 participants (7,151 cases and 4,338 controls) from five separate GWAS data sets, and the initial GECCO data set comprises 20,320 participants (9,498 cases and 10,822 controls) (Supplementary Table S1). Further information about CCFR and GECCO participants can be found in previous publications (11,12,15). Participants provided written informed consent by their respective Institutional Review Boards, and self-reported race as being White.

Genotyping quality control (QC) and data harmonisation.

Supplementary Figure S1, S2, and Supplementary Table S2 details the QC, harmonisation, and data analyses in this study. The initial imputed genotype data sets used in this study comprised of between 39,127,678 to 39,222,163 SNPs (Supplementary Table S3). QC and imputation had already been carried out by the respective consortia for these data sets (11,12,15,16). However, further QC was applied to ensure there is a baseline level of consistency between the data sets (data harmonisation) – see Supplementary Table S2. QC and data harmonisation were primarily carried out using PLINK v1.9 (17) and custom Perl scripts (Supplementary Table S2). All chromosomal positions/regions specified in this study refer to genome build GRCh37/hg19.

DEPTH SNPs and participants.

Typical GWAS QC was applied to the CCFR and GECCO datasets, plus an extra QC step that removed SNPs that had an odds ratio or P value that differed by more than 20% when comparing logistic regression results that adjusted for sex and the first four principal components, against logistic regression results that adjusted for sex but not the first four principal components (Supplementary Figure S3). This extra QC step was performed to remove SNPs that may be affected by principal components, as DEPTH analyses cannot adjust for principal components. The number of SNPs and patients that were removed after each step are detailed in Supplementary Tables S3 and S4. In total 7,234,408 SNPs on 10,846 CCFR participants (5,735 cases and 3,688 controls), and 19,150 GECCO participants (8,865 cases and 10,285 controls) were retained for DEPTH analyses following genotyping QC and data harmonisation.

DEPTH analyses.

All DEPTH analyses were carried out using the University of Melbourne’s high-performance computing system (18). Risk regions were defined as having a posterior log-odds in favour of association above 1.0. This risk value was based on a previous DEPTH study which suggests that the maximum 95^th percentile of the null distribution equals approximately 1.0 across the genome (14). DEPTH analyses were carried out in three stages, whereby the aim of the first stage was to identify potential common regions of CRC risk between the CCFR and GECCO data sets. – see Supplementary Figure S4 for a description of common regions. The second stage then evaluated candidate higher-confidence regions (95% confidence interval) using 100 simulations on the common risk regions that were identified in the first stage. The third stage entailed carrying out a more rigorous DEPTH simulation analysis using 1,000 bootstrap iterations on risk regions identified in stage 2 that was supported at the 95% confidence level. DEPTH was carried out using a sliding window of 200 SNPs for all stages. Genes that are located within DEPTH risk regions were identified by scanning DEPTH regions against the UCSC Genes table (GRCh37/hg19) (19) using a custom Perl script.

Comparison of DEPTH results with conventional GWAS.

Conventional GWAS logistic regression analysis of both the CCFR and GECCO data sets were carried out on 14,285,390 SNPs that were common to both data sets. Apart from the last QC step, these SNPs underwent the same QC and data harmonisation as the DEPTH data sets (data sets from Step 12 of Supplementary Table S2). Logistic regression was carried out using PLINK v1.9 (17), adjusting for sex and the first four principal components. A P value less than 5 x 10⁻⁸ was used to define a significantly associated SNP. GWAS summary statistics were uploaded to the Functional Mapping and Annotation of Genome-Wide Association Studies (FUMAR) platform (20) to identify and annotate SNPs that are associated with CRC risk.

The 50 CCFR DEPTH risk regions and the 66 DEPTH risk regions in the GECCO data sets that were identified at the 95^th percentile using 100 simulations were then compared with conventional GWAS analysis results using a custom Perl script. A comparison of the 40 CCFR and 49 GECCO DEPTH risk regions at the 95^th percentile using 1,000 simulations with conventional GWAS results was also carried out.

A comparison of DEPTH risk regions against previously published risk loci from conventional CRC GWASs was also carried out. This was performed as above but using the 132 risk loci (Supplementary Table S5) from the Law et. al. and Huyghe et. al. studies (10,11). However, of the 132 known CRC risk SNPs, only 98 risk SNPs are found in our DEPTH data sets comprising 7,234,408 SNPs and thus, only these 98 SNPs were assessed when comparing logistic regression risk SNPs against DEPTH peaks and the 98 known CRC risk SNPs (Supplementary Table S5). It was not possible to use LD SNPs for the remaining 34 known CRC risk SNPs because LD SNPs were removed during the QC process to facilitate feasible computational time for the bootstrap iteration step.

Ranking SNPs within the HLA risk locus using penalised regression.

The HLA region was chosen for further analysis given it was the most reproducible common loci detected between the CCFR and GECCO datasets after Stage 3 analysis. Penalized logistic regression with the lasso penalty and the random forest supervised machine learning algorithm were used to locate genomic variants most associated with CRC risk in the HLA region. Lasso regression was fitted using the cyclical coordinate descent algorithm with cross validation to select a model in the lasso solution path as implemented in the R package glmnet. The order in which the SNPs enter the lasso solution path via the cyclical coordinate descent algorithm was taken as a surrogate for the rank of each SNP. For example, the first SNP that entered the model was ranked top (ie, most associated with CRC risk), the second SNP that entered the regression model was ranked second, etc. Random forests were fitted using the randomForest R package for supervised classification and regression with default values for the number of trees and the number of candidates variables for each split in a tree. Further tuning of these parameters did not change the final ranking of SNPs significantly. Importance of each SNP in the HLA locus was determined automatically by the random forest algorithm using two default criteria (prediction error on the out-of-bag portion of the data and the total decrease in node impurities from splitting on the variable). In total, 381 SNPs located within the HLA risk locus were ranked in order of strength of association with CRC risk in both the CCFR and GECCO data sets.

Data Availability.

Summary-level data for the full set of Asian and European GWASs are available through the GWAS catalog (accession no. GCST90129505). For individual-level data, CCFR, CORECT, CORSA_2 and GECCO are deposited in dbGaP (accession nos. phs001415.v1.p1, phs001315.v1.p1, phs001078.v1.p1, phs001903.v1.p1, phs001856.v1.p1 and phs001045.v1.p1).

Results:

Identification of CRC risk regions by DEPTH analyses.

The first stage of DEPTH analysis identified 2,622 risk regions in the CCFR data set, 3,686 risk regions in the GECCO data set (Table 1 and Supplementary Table S6), and 569 risk regions that were common to both the CCFR and GECCO data sets (Supplementary Table S7). The second stage of DEPTH analysis indicated that of these 569 common risk regions, 50 (8.8%) and 66 (11.6%) remained significantly associated with risk after 100 bootstrap iterations at the 95% confidence level in the CCFR and GECCO data sets, respectively, with 5 regions that were common to both data sets (Table 1 and Supplementary Table S8). The third stage of DEPTH analysis indicated that of the 50 CCFR and 66 GECCO risk regions from stage 2, 40 (80.0%) and 49 (76.6%) remained significantly associated with risk after 1,000 bootstrap iterations at the 95% confidence level in the CCFR and GECCO data sets, respectively, with four regions (chr2:143,560,338 – 143,668,969 (2q22.2), chr2:201,851,477 – 201,991,983 (2q33.1), chr6:32,421,184 – 32,682,590 (6p21.32), and chr13:52,434,734 – 52,788,187 (13q14.3)) that were common to both data sets (Table 1 and Supplementary Table S9). Notably, three ((chr2:143,560,338 – 143,668,969 (2q22.2), chr2:201,851,477 – 201,991,983 (2q33.1), and chr13:52,434,734 – 52,788,187 (13q14.3)) of the four regions were not discovered by previous GWAS studies, and none were detected using logistic regression analysis using the same datasets (Supplementary Table S9).

Table 1:

Number of risk regions (posterior log-odds > 1) identified by DEPTH analyses.

DEPTH stage	CCFR	GECCO	Common regions
aFirst stage	2,622	3,686	569
bSecond stage	50	66	5
cThird stage	40	49	4

^aUsing raw DEPTH results.

^bRisk at 95^th percentile results using 100 bootstrap iterations.

^cRisk at 95^th percentile results using 1,000 bootstrap iterations.

DEPTH analyses identify risk regions that are not detected by conventional GWAS.

Conventional GWAS using logistic regression analyses revealed four (CCFR) and nine (GECCO) SNPs that are associated with CRC risk (Supplementary Table S10 and Supplementary Figure S5), and FUMAR analysis indicates that these SNPs map to six independent risk loci (CCFR: rs13200613 at 6q27, rs2391397 at 7p15.2, rs4325396 at 13q11, rs4553118 at 1p36.33; GECCO: rs6983267 at 8q24.21, rs79207432 at 15q13.3) (Table 2). A comparison of these six lead conventional GWAS logistic regression risk SNPs from our DEPTH data sets against 98 of the 132 known CRC risk loci in our DEPTH dataset indicates that only two (rs6983267 and rs79207432) of the 98 loci (1.52%) could be detected in our CCFR and GECCO data sets (Table 2 and Supplementary Table S5).

Table 2:

Risk loci identified by conventional GWAS analyses.

Data Set	Lead SNP	Chr	aPos.	loci	bEffect allele	cMAF	Odds ratio	Pvalue	Novel loci
GECCO	rs4553118	1	832398	1p36.33	T	0.270	1.149	9.541e−9	yes
GECCO	rs6983267	8	128413305	8q24.21	T	0.500	0.893	4.193e−8	no
GECCO	rs79207432	15	33011697	15q13.3	C	0.215	1.151	2.873e−8	no
CCFR	rs13200613	6	170529391	6q27	G	0.026	0.809	1.118e−12	yes
CCFR	rs2391397	7	27236558	7p15.2	G	0.078	0.848	8.523e−9	yes
CCFR	rs4325396	13	19348805	13q11	C	0.132	0.791	9.953e−15	yes

^aGRCh37/hg19.

^bEffect allele refers to the allele that is associated with the odds ratio value.

^cMAF = minor allele frequency.

Similar comparative analysis of our CCFR and GECCO conventional GWAS results against DEPTH first stage results indicates that only one (rs13200613) of the six lead logistic regression risk SNPs are located within the 569 common DEPTH risk regions (Supplementary Table S7). Notably, these three risk loci were also detected in DEPTH stage 2 and 3 results (Supplementary Tables S8 and S9). Thus, most DEPTH risk loci (568 of 569 in the first stage, 110 of 111 in the second stage, and 84 of 85 in the third stage) cannot be detected by conventional GWAS analyses in our CCFR and GECCO data sets (Figure 1 and Supplementary Tables S7–S9).

Figure 1.

Venn diagram showing how many of the DEPTH risk regions (red circles) were detected by conventional GWAS analyses using the same data set (green circles) or were found in previous CRC GWASs (blue circles). Analysis was compared across results from the first (a), second (b), and third (c) stages of DEPTH analyses.

A comparative analysis was also carried out to assess how many of the 98 known CRC risk loci that are in our DEPTH data set are also detected by DEPTH analysis. Our results indicate that four (4.1%), two (2.0%), and two (2.0%) of the 98 known risk loci was detected by the first, second, and third stages of DEPTH analysis, respectively (Figure 1 and Supplementary Table S5). Thus, most of the DEPTH risk regions have not been detected by previous studies that use conventional GWAS approaches. Collectively, DEPTH identified 564 (first stage), 108 (second stage) and 82 (third stage) novel risk regions that could not be detected using conventional GWAS approaches in our CCFR and GECCO data sets, nor were they identified in previous CRC GWASs (Figure 1).

Analysis of the Human Leukocyte Antigen (HLA) locus.

The HLA risk locus spanning chr6:32,421,184-32,682,590 at 6p21 (Figure 2) was found to be most significantly associated with CRC risk at all stages of DEPTH analyses and was consistently significant in both CCFR and GECCO datasets (Table 3). As such, this region was selected for penalised regression analysis to determine the main SNPs that may confer risk in this region. Our results indicate that the most associated SNPs that were identified by all three methods were rs762216297, rs149490268, rs114741460, and rs199707618 for the CCFR data set, and rs9270761 for the GECCO data set, all of which cover the same HLA region (Figure 2 and Supplementary Table S11).

Figure 2.

Diagram of the HLA locus (bordered by the GABBR1 and KIFC1 genes) showing a common DEPTH risk region. Loci with a log-posterior odds (LPO) above one are classified as risk regions. Also shown is a detailed view of the common DEPTH risk region that encompasses the HLA-DRB5, HLA-DRB6, AK293020, HLA-DQA1, and HLA-DQB1 genes, as well as the five main SNPs (rs199707618, rs9270761, rs149490268, rs114741460, and rs762216297) driving risk in this region. Chromosomal positions above the genes refer to GRCh37/hg19 genome build. The density of common SNPs in the common risk region is shown by the green blocks.

Table 3:

DEPTH analyses¹ at the HLA locus (chr6:32,421,184-32,682,590).

DEPTH stage	CCFR DEPTH PLO	GECCO DEPTH PLO
aFirst stage	3.60	6.63
bSecond stage	2.94	12.27
cThird stage	3.00	12.23

¹Significant regions were defined as having a Posterior log-odds (PLO) > 1.

^aPosterior log-odds (PLO) risk using raw DEPTH results.

^bPosterior log-odds (PLO) risk at 95^th percentile using 100 bootstrap iterations.

^cPosterior log-odds (PLO) risk at 95^th percentile using 1,000 bootstrap iterations.

Discussion:

Since the first GWAS was proposed in 1996 (21), this method has been used extensively to identify low penetrant but common SNPs that are associated with a wide range of complex phenotypes. For CRC, 132 SNPs have been found to be associated with the risk of developing this disease in individuals of European ancestry (8,10–12), of which 98 SNPs could be tested in this study.

Novel SNPs that are associated with CRC risk are increasingly more difficult to identify using conventional GWAS approaches. As such, conventional GWASs have adopted various techniques such as increasing sample sizes, meta-analysis, and using imputed SNPs, to identify novel lower penetrant SNPs in subsequent GWASs. Nevertheless, these approaches still assess for association by using one SNP at a time. However, this approach is overly conservative because it uses excessively stringent Bonferroni thresholds because they assume statistical independence of SNPs, even though there is linkage disequilibrium. In this study, we have applied a novel alternative approach called (DEPTH), whereby we have investigated CRC risk regions by using a sliding window of 200 SNPs to collectively identify risk regions in two CRC GWAS data sets (CCFR and GECCO). DEPTH is a hypothesis generation algorithm based on a novel feature ranking algorithm incorporating data re-sampling and stability selection (13). The utility of DEPTH in discovering novel risk regions using GWAS data sets is highlighted by our identification of 112 novel prostate cancer risk regions, as well as five known risk regions that have only been identified by much larger conventional GWAS analyses (14). Had we limited our analysis to the conventional GWAS method applied to both data sets, we would have only found 6 independent loci.

Here, we performed DEPTH analysis in three stages due to the computational demands of DEPTH. This approach ultimately led to the identification of 37 (CCFR) and 48 (GECCO) risk regions that were not detected by conventional GWAS analyses, nor were they discovered in previous CRC GWASs, demonstrating the novelty of our approach.

Interestingly, the most highly associated DEPTH risk region (6p21.32 at chr6:32,421,184-32,682,590) was located within the human leukocyte antigen (HLA) locus, which is defined as being between the GABBR1 and KIFC1 genes (22). Notably, this CRC risk region could not be detected until only recently when Huyghe et al and Law et. al. carried out a meta-analysis that includes new data sets that had not been used in previous GWASs (10,11). However, we were able to identify the HLA risk loci despite only having 5,735 cases and 3,688 controls for the CCFR data set, and 8,865 cases and 10,285 controls for the GECCO data set, whereas Law et al. had 34,627 cases and 71,379 controls. Thus, DEPTH has a sensitivity that is much greater than conventional GWAS approaches for the same sample size. Further, the SNP (rs9271770) at the HLA risk region in Law et. al. is located directly within our DEPTH 6p21.32 risk region, even though this SNP is not in our data set. Thus, we were able to identify this risk loci without using rs9271770. The other SNP (rs3131043) at the HLA risk region in the Law et. al. study was not located within any of our DEPTH risk regions. Interestingly, the main SNPs driving risk at this locus are different between the two data sets (rs762216297, rs149490268, rs114741460, rs199707618 for CCFR, and rs9270761 for GECCO). The correlation (r²) and linkage disequilibrium (D prime) between rs199707618 and rs9270761 is 0.02 and 1—the discrepancy possibly due to rs9270761 having a very low minor allele frequency. This finding further highlights the utility of DEPTH in that it is not solely reliant on one SNP as it might be due to non-linear associations and epistasis.

The HLA DEPTH risk region in this study harbours five genes (HLA-DRB5, HLA-DRB6, AK293020, HLA-DQA1, HLA-DQB1), whereas the HLA locus collectively comprises over 130 protein coding genes that are important in regulating the immune system (22). While the HLA locus and the immune system has been known to be important in a wide range of diseases such as diabetes, rheumatoid arthritis and various autoimmune disorders (22), only recently has the role of the immune system been intensively studied for its role in cancer, and the potential manipulation of this system in treating cancer through immunotherapy (23). It is tempting to speculate whether the HLA DEPTH risk region and the five genes located within would be of value in pharmaco-genomics, or as a therapeutic target.

In conclusion, we show using two CRC GWAS data sets (CCFR and GECCO) that more risk regions can be identified using an alternative approach (DEPTH) without needing to increase sample sizes. In fact, this study reduced the number of participants compared to the original CCFR and GECCO data sets due to an extra layer of QC and data harmonisation that was applied to this study. Finally, we designed DEPTH as an exploratory tool to identify novel risk loci, with the expectation that other methods can later be used to validate these loci. For example, in this study we used penalised regression to identify the main SNPs driving risk at the HLA locus, although other methods such as bioinformatic and/or functional analyses of the SNPs at these risk loci may also inform on the biological reasons driving the risk association. We expect DEPTH will provide great utility to the GWAS field as a complementary tool to identify novel loci that are associated with complex phenotypes using existing GWAS data sets.