Developing and validating polygenic risk scores for colorectal cancer risk prediction in East Asians

Abstract

Several polygenic risk scores (PRSs) have been developed to predict the risk of colorectal cancer (CRC) in European descendants. We used genome-wide association study (GWAS) data from 22,702 cases and 212,486 controls of Asian ancestry to develop PRSs and validated them in two case-control studies (1,454 Korean and 1,736 Chinese). Eleven PRSs were derived using three approaches: GWAS-identified CRC risk SNPs, CRC risk variants identified through fine-mapping of known risk loci, and genome-wide risk prediction algorithms. Logistic regression was used to estimate odds ratios (ORs) and area under the curve (AUC). PRS_115-EAS, a PRS with 115 GWAS-reported risk variants derived from East-Asian data, validated significantly better than PRS_115-EUR derived from European descendants. In the Korea validation set, OR per SD increase of PRS_115-EAS was 1.63 (95%CI=1.46-1.82; AUC=0.63), compared with OR of 1.44 (95%CI=1.29-1.60, AUC=0.60) for PRS_115-EUR. PRS_115-EAS/EUR derived using meta-analysis results of both populations slightly improved the AUC to 0.64. Similar but weaker associations were found in the China validation set. Individuals among the highest 5% of PRS_115-EAS/EUR have a 2.52-fold elevated CRC risk compared with the medium (41-60^th) risk group and have a 12-20% risk of developing CRC by age 85. PRSs constructed using results from fine-mapping and genome-wide algorithms did not perform as well as PRS_115-EAS and PRS_115-EAS/EUR in risk prediction, possibly due to a small sample size. Our results indicate that CRC PRSs are promising in predicting CRC risk in East Asians and highlights the importance of using population-specific data to build CRC risk prediction models.

1. Introduction

Colorectal cancer (CRC) is one of the most commonly diagnosed malignancies around the world.^{1, 2} In the United States and many other developed countries, CRC incidence and mortality have steadily declined over the past few decades due to the implementation of effective population-based screening programs to detect and remove pre-cancerous lesions and early-stage cancer.^2-5 However, the incidence and mortality of this cancer continue to rise in many Asian countries where rates are lower than those of the United States. Currently, many Asian countries do not have a population-based CRC screening program, significantly hindering the prevention of this disease. Because of economic constraints and differences in CRC risk, population-based CRC screening programs currently implemented in the U.S. and European countries may not be feasible or even appropriate for Asian countries. A cost-efficient, population-specific CRC screening strategy for Asians is imminently needed.

Current guidelines for initiating CRC screening are mainly based on age and family history of CRC, while more than 80% of CRC cases occur in individuals without a positive family history.⁶ CRC has a sizable heritable fraction and is a polygenic disease.⁷ Since 2007, genome-wide association studies (GWAS), including our studies conducted in East Asians,^8-12 identified common genetic variants in more than 190 loci associated with CRC risk.^8-24 Polygenic risk scores (PRSs) constructed using CRC-associated risk variants as a measure of cumulative effect of these variants were evaluated for CRC risk prediction in several studies conducted in European-ancestry populations.^25-29 However, to date, no study has systematically developed and validated PRSs for CRC risk prediction in East Asian populations. Given the difference in genetic architectures between Asian- and European-ancestry populations, PRSs established in European-ancestry populations may not perform well in predicting CRC risk in Asians. In 2009, we established the Asia Colorectal Cancer Consortium (ACCC) to evaluate genetic susceptibility factors of CRC in Asians. In the current study, we used GWAS data from the ACCC, including 24,192 CRC cases and 214,186 controls, to develop and validate performance of PRS in CRC risk prediction.

2. Methods

2.1. Datasets

To develop CRC risk prediction models for individuals of East Asian ancestry, we used GWAS data collected from 22,702 cases and 212,486 controls shown in Table 1. These study participants were recruited from eight studies conducted in China, Korea, and Japan. Data from the Korean-National Cancer Center CRC Study (Korea validation set: a case-control study including 622 cases and 832 controls) and a case-control study nested in cohorts of the Shanghai Men’s Health Study and Shanghai Women’s Health Study (China validation set: 868 cases and 868 age- and sex-frequency matched controls) were used for model validation. Detailed information on each study was previously reported^8-12 and is described in the Supplementary Materials.

Table 1.

Sample size and selected descriptive statistics of participating studies: the Asia Colorectal Cancer Consortium (ACCC).

Participating Studies (Acronym)	Ethnicity	Sample Size		Female (%)		Age (Mean ± SD)
		Cases	Controls	Cases	Controls	Cases	Controls
Training Datasets		22,702	212,486	40.1	60.8	60.9 ± 11.4	59.7 ± 15.7
Shanghai Studies (Shanghai-1,2,3)	Chinese	3,303	4,612	50.2	85.0	61.3 ± 8.52	54.7 ± 9.60
Aichi CRC Studies (Aichi-1,2)	Japanese	625	1,396	36.6	43.6	59.7 ± 10.0	51.9 ± 15.5
Guangzhou CRC Studies (Guangzhou-1,2)	Chinese	2,479	2,227	37.5	32.8	55.5 ± 13.0	51.1 ± 13.8
Korean Cancer Prevention Study II (KCPS-II)	Korean	325	975	27.1	43.4	51.4 ± 10.6	41.3 ± 8.44
Hwasun Cancer Epidemiology Study (HCES)	Korean	6,822	5,689	36.6	72.7	63.4 ± 11.2	59.0 ± 15.3
Korea National Cancer Center Study (Korea-NCC)	Korean	1,313	1,223	38.5	39.0	58.3 ± 11.5	55.6 ± 9.26
Seoul CRC Study (Korea-Seoul)	Korean	773	619	40.8	48.0	59.1 ± 9.61	57.2 ± 9.62
BioBank Japan Study (BBJ)	Japanese	7,062	195,745	-	-	-	-
Validation Datasets		1,490	1,646	42.6	40.7	63.8 ± 11.2	63.9 ± 11.7
Korea validation set	Korean	622	832	32.3	34.1	56.4 ± 8.82	56.4 ± 9.68
China validation set	Chinese	868	868	50.0	50.8	69.1 ± 9.04	71.4 ± 9.09
Younger (Age < 70.3)		443	395	51.0	58.7	61.5 ± 5.48	63.0 ± 4.73
Older (Age ≥ 70.3)		425	473	48.9	44.2	77.0 ± 3.75	78.5 ± 4.80

2.2. Genotyping and Imputation

Details of genotyping, quality control, and imputation for the ACCC were reported previously^8-12 and are provided in the Supplementary Materials (Supplementary Table S1). We imputed genotype data using the 1000 Genomes Project Phase III data as a reference via the Michigan Imputation Server.³⁰ Only variants with minor allele frequency (MAF) > 5%, a high imputation quality (R² > 0.8), and measured in over half of our training datasets were included for further analyses.

2.3. Polygenic Risk Score Calculation

To estimate odds ratios of SNPs associated with CRC risk, we used logistic regression models adjusting for age, sex, and top 10 principal components (PCs) for genetic structure. Association analyses were performed in each of the eight ACCC datasets included in the training set, and then a meta-analysis was conducted to estimate pooled odds ratios using METAL software.³¹ PRSs were calculated as weighted sums of alleles associated with CRC risk using equation $PRS={\sum }_{k=0}^n{\widehat{\beta }}_{k}SN{P}_k$, where SNP_k was the allelic dosage and ${\widehat{\beta }}_k$ was the corresponding log-odds ratio of SNP_k associated with CRC risk derived from the meta-analysis. We developed PRSs using three different approaches as described below and summarized in Figure 1.

Figure 1.

Summary of approaches to derive polygenic risk scores (PRS) for colorectal cancer (CRC) in East Asians.

Approach 1: PRS using GWAS-identified index risk SNPs

By searching the literature, we identified common SNPs in 198 regions showing a significant association with CRC risk at P < 5.0 x 10⁻⁸ from large CRC GWASs conducted in East Asians, the European-ancestry population, or both populations.^8-24 If two SNPs are in linkage disequilibrium (LD) with R² > 0.1 in East Asians (1000 Genomes, Phase 3 v5), the SNP with lower P value was kept for PRS construction. In total, 126 SNPs with MAF > 1% were included as independent CRC risk variants. Of them, 115 risk variants were consistent in the direction of association with CRC risk in both East Asian and European descendants, and these risk variants were used to construct PRS₁₁₅. European log-odds ratios obtained from the literature were used as the weights to construct PRS_115-EUR,^8-24 while log-odds ratios derived from our ACCC training set were used to construct PRS_115-EAS. We performed a meta-analysis of log-odds ratios derived from European and East Asian populations using the fixed-effects model to estimate the pooled log-odds ratios, and used them as weights to construct PRS_115-EAS/EUR.

Approach 2: PRS using SNPs selected from fine-mapping of GWAS-identified risk loci

Approximately 78% (n=99) of these 126 independent CRC risk variants were initially identified in GWAS conducted in European descendants. To identify additional independent risk variants and risk variants more strongly associated with CRC in each of these regions in East Asians, we performed fine-mapping analyses with data from our training set using the Genome-wide Complex Trait Analysis (GCTA) method described by Yang.³² We first extracted 53,015 SNPs located at flanking 500Kbp regions of each of the 126 index SNPs and with nominal significant association at P < 0.05 in the training set. We then conducted conditional and joint analyses (COJO) in each region adjusting for the corresponding GWAS-identified lead SNP, from which we identified 156 SNPs of 114 loci showing an independent association with CRC risk at P < 1.0 × 10⁻³. Adjusted log-odds ratios of these 156 SNPs, along with log-odds ratios of the GWAS-identified risk variants for the remaining 12 loci where no SNP showed an association with CRC risk at P < 1.0 × 10⁻³ in our training data set, were used to construct the PRS (PRS₁₆₈). We also evaluated PRSs conducted using risk SNPs selected using more stringent p-values (such as 1.0 × 10⁻⁴ and 1.0 × 10⁻⁵), and results for these evaluations are shown in Supplementary Table S3.

Approach 3: PRSs based on genome-wide risk prediction algorithms

We used LDpred and PRS-CS to derive genome-wide PRSs using data from the training set. LDpred is a Bayesian approach that considers LD among SNPs^33-35 and may have a higher accuracy than PRS using only GWAS-identified risk SNPs.²⁷ We used summary statistics from our ACCC training set for model training. LD matrix was calculated using genotype data of all ACCC training sets except the BioBank Japan study (BBJ), which included 32,381 samples. Individual genotype data for BBJ were not available for the present study. According to LDpred2 recommendation, we restricted our analysis to SNPs included in the HapMap3 SNPs. After applying these criteria, 747,643 SNPs were included in this analysis. Log-odds ratios of these SNPs with CRC risk were re-estimated using LDpred2 with default settings. We applied four models from LDpred2 in this study (infinitesimal model <LDpred2-Inf>, the best model among all sparse models <LDpred2-grid-sp>, the best model among all non-sparse models <LDpred2-grid-nosp>, and the automatic model <LDpred2-Auto>).

PRS-CS is a Bayesian approach that infers posterior SNP effect sizes under continuous shrinkage (CS) priors using summary statistics and an external LD reference panel.³⁶ Summary statistics from our ACCC training set and pre-calculated LD reference panel of East Asians constructed using the UK Biobank data from PRS-CS were used for model training.

2.4. Model Performance Assessment and Risk Estimation

We evaluated the performance of PRSs derived in training sets in our validation datasets by calculating the area under the receiver operating characteristic curve (AUC). Odds ratio (OR) and 95% confidence interval (CI) per one standard deviation (SD) increase in PRS were estimated using logistic regression. Age, sex, ethnic group, study sites, and top 10 PCs were adjusted in the risk estimation model using logistic regression. We also estimated ORs for selected PRS groups (≤1^st, 2-5^th, 6-10^th, 11-20^th, 21-40^th, 61-80^th, 81-90^th, 91-95^th, 96-99^th, >99^th PRS percentile) relative to those in the medium risk group (41^th to 60^th PRS percentile).

We estimated ten-year and lifetime absolute risk of CRC by PRS categories for Chinese, Korean, and Japanese subjects using ORs from our study of the association of PRS with CRC risk and data on CRC incidence and mortality rates for China, Japan, and South Korea obtained from the GLOBOCAN database (2020)¹. All statistical analyses were conducted using R v4.1.2.

3. Results

Data from eight datasets were used for PRS training and two datasets for PRS validation. Selected characteristics of study participants are summarized by studies in Table 1. Cases and controls differed considerably by age and sex in multiple studies included in the training set and thus these two variables were adjusted for in the analysis. In the validation sets, age and sex were comparable between cases and controls.

As described in the method section, we developed PRS_115-EAS, PRS_115-EUR, and PRS_115-EAS/EUR using log-odds ratios for 115 GWAS-identified risk variants derived from GWAS conducted in East Asians, European-ancestry populations, and meta-analyses of both populations, respectively. These log-odds ratios were used as weights for constructing PRS₁₁₅ (Supplementary Table S2). PRS_115-EAS performed significantly better in discriminating CRC cases from non-cases than PRS_115-EUR (Table 2). In the Korea validation set, ORs per SD increase of PRS_115-EAS were 1.63 (95%CI = 1.46 - 1.82; AUC = 0.63), compared with OR of 1.44 (95%CI = 1.29 - 1.60, AUC = 0.60) for PRS_115-EUR. PRS_115-EAS/EUR slightly improved the AUC to 0.64 (95% CI = 0.61 - 0.67). Similar but weaker associations were found in the China validation set. Because subjects included in the China validation set were older than those included in the Korea validation set, we performed stratified analyses by age using the mean age (70.3 years) as the cut-off to evaluate if the performance of PRSs differ by age. The performance of all three PRSs was better in the younger group than the older group in the China validation set (Supplementary Table S3). For example, the PRS_115-EAS/EUR showed an AUC of 0.63 (95% CI = 0.60 - 0.67) and OR of 1.62 (1.40 – 1.88) in the younger China validation set, while its AUC in the older group is only 0.59 (0.56 - 0.63) and the OR is 1.39 (1.21 - 1.60). The difference between these two OR estimates was statistically significant (P for interaction, 0.038).

Table 2.

Associations of polygenic risk scores with colorectal cancer risk in the validation datasets.

PRS development methods	Korea validation set		China validation set
	OR (95% CI)1	AUC (95% CI)	OR (95% CI)1	AUC (95% CI)
GWAS-reported index SNPs 2
PRS115-EAS	1.63 (1.46 - 1.83)	0.63 (0.60 - 0.66)	1.51 (1.37 - 1.67)	0.61 (0.59 - 0.64)
PRS115-EUR	1.44 (1.29 - 1.60)	0.60 (0.57 - 0.63)	1.39 (1.26 - 1.53)	0.59 (0.56 - 0.61)
PRS115-EAS/EUR	1.68 (1.50 - 1.89)	0.64 (0.61 - 0.67)	1.50 (1.36 - 1.66)	0.61 (0.58 - 0.64)
SNPs selected by fine-mapping 3
PRS168	1.54 (1.38 - 1.72)	0.62 (0.59 - 0.65)	1.32 (1.20 - 1.45)	0.58 (0.55 - 0.61)
Genome-wide risk prediction algorithm 4
PRSLDpred2-Grid-Sp	1.43 (1.29 - 1.60)	0.60 (0.57 - 0.63)	1.32 (1.20 - 1.45)	0.58 (0.55 - 0.60)
PRSPRS-CS	1.34 (1.20 - 1.49)	0.58 (0.55 - 0.61)	1.35 (1.23 - 1.49)	0.58 (0.55 - 0.61)

^*PRS, polygenic risk score; OR, odds ratio; CI, confidence interval; AUC, aera under the receiver operating characteristic curve.

^1.Odds ratio and 95% confidence interval per standard deviation increase in PRS were estimated using logistic regression.

^2.All weights for European ancestry (PRS_115-EUR) were obtained from published papers. Weights for East Asian (PRS_115-EAS) were derived from the ACCC training set. Weights of PRS_115-EAS/EUR were from a meta-analysis based on weights for European ancestry and weights for East Asians.

^3.Weights for PRS were derived from the ACCC training set.

^4.Weights for PRS based on LDpred2 and PRS-CS were re-estimated using LDpred2 and PRS-CS based on the ACCC training set.

Using fine-mapping methods, we developed three PRSs based on COJO-p-values cut-off of 1.0 × 10⁻³, 1.0 × 10⁻⁴, and 1.0 × 10⁻⁵ to select SNPs for PRS construction (Supplementary Table S3). Of the three PRSs evaluated, PRS₁₆₈, based on p-value of 1.0 × 10⁻³, had the best performance, showing an AUC of 0.62 and 0.58 in the Korea and China validation sets, respectively (Table 2). Performances of other two PRSs in the validation sets were similar. None of these PRSs performed better than the PRS_115-EAS or PRS_115-EAS/EUR described above.

We developed five PRSs using the genome-wide risk prediction algorithms, based on four LDpred2 models (infinitesimal model, the best model among all sparse models, the best model among all non-sparse models, and the automatic model) and one PRS-CS model to derive weights for SNPs for PRS construction. Of these five PRSs evaluated, PRS_{LDpred2-Grid-Sp}, based on the best main model among all sparse models, demonstrated the highest discriminative ability, showing an AUC of 0.60 and 0.58 in the Korea and China validation sets, respectively (Table 2). Results based on PRS-CS (PRS_PRS-CS) showed AUCs of 0.58 in both the Korea and China validation sets. Other PRS performances in the validation sets were similar and are shown in Supplementary Table S3. Again, none of these PRSs performed better than the PRS_115-EAS or PRS_115-EAS/EUR described above.

Because the PRS_115-EAS/EUR showed the best performance in our external validation, subsequent analyses were based on this PRS. We estimated ORs according to percentile of PRS_115-EAS/EUR in Table 3. A clear dose-response association between PRS levels and CRC risk was observed in the combined validation dataset, including samples from both Korea and China (P for trend < 0.001). Because of a relatively small sample size in the validation set, we also evaluated the associations of PRS level with CRC risk in all subjects included in the ACCC (including both training and validation datasets, except BBJ data). The risk estimates were similar in these two datasets, except those at the 95^th to 99^th and >99^th percentile groups, in which the estimated ORs were higher in the combined dataset than the validation set. This difference is likely due to unstable estimates in the validation set because of a small sample size.

Table 3.

Odds ratios of colorectal cancer risk in association with PRS_115-EAS/EUR in the validation datasets¹ and ACCC datasets combined².

PRS Percentile	Validation datasets		ACCC datasets combined
	Number of Cases/Controls	OR (95% CI)	Number of Cases/Controls	OR (95% CI)
≤1	5/18	0.33 (0.11 - 0.85)	35/177	0.34 (0.23 - 0.49)
[2, 5]	21/68	0.37 (0.22 - 0.61)	183/683	0.48 (0.40 - 0.57)
[6, 10]	34/85	0.51 (0.33 - 0.78)	269/858	0.56 (0.48 - 0.65)
[11, 20]	80/170	0.58 (0.42 - 0.79)	639/1713	0.66 (0.59 - 0.73)
[21, 40]	226/339	0.81 (0.64 - 1.02)	1573/3426	0.82 (0.75 - 0.89)
[41, 60]	278/340	1.00 (Reference)	1930/3425	1.00 (Reference)
[61, 80]	332/339	1.19 (0.96 - 1.49)	2577/3426	1.34 (1.24 - 1.45)
[81, 90]	223/170	1.59 (1.23 - 2.05)	1494/1712	1.56 (1.43 - 1.71)
[91, 95]	128/85	1.93 (1.41 - 2.66)	917/857	1.88 (1.68 - 2.10)
[96, 99]	122/68	2.21 (1.58 - 3.11)	915/683	2.34 (2.09 - 2.63)
>99	34/18	2.29 (1.28 - 4.24)	332/177	3.22 (2.65 - 3.92)
P for trend		< 0.001		< 0.001
Top 5% Vs Remaining		2.21 (1.68 - 2.92)		2.40 (2.19 - 2.64)
Top 1% Vs Remaining		2.17 (1.23 - 3.95)		2.92 (2.42 - 3.53)

^1.Including 1,490 cases and 1,646 controls from both Korea and China validation sets.

^2.Including 1,7130 cases and 18,387 controls from both training and validation datasets except BBJ data. Individual genotype data for BBJ were not available for the present study.

We used incidence and mortality data from the GLOBOCAN database for the year 2020 to estimate absolute CRC risks by PRS group. Figure 2 shows the estimated 10-year absolute risk of CRC by PRS_115-EAS/EUR groups in Chinese, Japanese, and Korean subjects. The recommended age to start colorectal cancer screening in the general population is 40 years in Japan ³⁷ and 50 years in both China ³⁸ and South Korea ³⁹. At age 50, the 10-year absolute risks for CRC were 0.47% and 0.52% for average-risk individuals (at 40-60^th PRS percentile) in China and Korea, respectively (Supplementary Table S4). In Japan, the 10-year absolute risk for CRC was 0.24% for an average risk individual at age 40. However, individuals in the top 1% PRS_115-EAS/EUR group (>99^th percentile) reached this risk level by ages 38, 36, and 32 in China, Korea, and Japan, respectively, much earlier than the average risk group.

Figure 2.

Ten-year absolute risks of colorectal cancer (CRC) by PRS_115-EAS/EUR groups in China, South Korea, and Japan. The green horizontal lines show the 10-year absolute CRC risk (0.69%) for individuals at age 50 in the United States in the year 2020. The red horizontal lines show the 10-year absolute CRC risk for individuals at age 50 in China (0.47%, Figure 2A) and South Korea (0.52% Figure 2B), and 10-year absolute CRC risk (0.24%, Figure 2C) for individuals at age 40 in Japan.

The 10-year absolute CRC risk is 0.69% for average-risk individuals at age 50 in the United States. An individual with a medium PRS_115-EAS/EUR (40^th to 60^th percentile) reaches this risk level at age 54, 53, and 48 years in China, Korea, and Japan, respectively (Figure 2). However, among those in the top 1% risk group, an individual would reach this risk level at age 42, 40, and 38 years in China, Korea, and Japan, respectively.

The estimated lifetime absolute risks for East Asians by PRS_115-EAS/EUR categories for CRC are shown in Figure 3. By age 85, the absolute risk of CRC in the top 1% of PRS_115-EAS/EUR was 16.4%, 18.6%, and 26.0%, in China, Korea, and Japan, respectively. In the lowest 1% of PRS_115-EAS/EUR, by age 85, the absolute risk of CRC was 1.7%, 2.0%, and 2.7% in China, Korea, and Japan, respectively. Individuals in the top 5% of PRS_115-EAS/EUR were estimated to have 12.9%, 14.5%, and 20.4% risk of developing CRC by age 85 years in China, Korea, and Japan, respectively.

Figure 3.

Lifetime absolute risks of colorectal cancer by PRS_115-EAS/EUR groups in China, South Korea, and Japan.

4. Discussion

In this study, we built and validated PRSs to predict CRC risk for East Asians using the largest data set available to date, including 24,192 CRC cases and 214,186 controls of East Asian ancestry from the ACCC. We found that the PRSs derived using GWAS-identified CRC risk variants showed promise in predicting the risk of this common cancer, particularly when data from East Asians were used in constructing the PRS. Our study demonstrates the importance of incorporating population-specific data to build risk prediction models for CRC.

Theoretically, one would expect that PRSs built using risk variants selected from fine-mapping known CRC risk loci or using genome-wide algorithms would outperform PRS constructed using only GWAS-reported CRC risk variants. In our study, however, these PRSs showed a poorer performance than PRSs derived using GWAS-reported risk variants. The sample sizes used in fine-mapping and deriving genome-wide PRSs were relatively small, which might have affected the stability of risk estimates in this study. Therefore, future studies with a larger sample size are needed to further improve the performance of PRS in East Asians.

An interesting finding from our study is that the discriminative ability of PRS was lower in older study participants. For example, the AUC for PRS_115-EAS/EUR was 0.64 in the Korea validation set (mean age = 56.4). In the China validation set (mean age = 70.3), however, the PRS_115-EAS/EUR had a poor performance with an AUC of 0.61. When stratified by age in the China validation set, we showed that the performance of PRS_115-EAS/EUR is better in the younger group (AUC = 0.63, mean age = 62.2) than the older group (AUC = 0.60, mean age = 77.8). Our findings for a significant interaction between age and PRS were supported by recent studies conducted in European-ancestry populations.^{27, 40} It is possible that lifestyle and environmental risk factors may play a more significant role in the etiology of CRC in older than younger patients, and thus the prediction accuracy of PRS may decrease among elderly compared with younger populations.

Although this is the largest study ever conducted to date in East Asians to develop and validate CRC risk prediction using GWAS data for this cancer, the sample size is relatively small, particularly in analyses stratified by ethnic groups and sex. Because of the small sample size for the validation set, we used log-odds ratios derived from all data combined to estimate absolute risk. This could lead to potential overfitting. However, the ORs of CRC by PRS percentile estimated using data from the validation set or all ACCC data were similar, suggesting that overfitting may not be a major concern. To construct PRS_115-EUR, we obtained weights for known CRC risk variants from the literature. Some randomness in risk estimates could occur. However, given the last sample size of these previous studies, for which the weights were derived, we believe that these random errors should be small. Future studies could address this limitation by using weights from all risk variants derived from a single large study. Additionally, not all studies had data on CRC family history, thus we were not able to include family history in our analysis.

In summary, we found that PRSs derived using GWAS data of CRC showed promise in predicting the risk of this cancer in East Asians, particularly when data from East Asians are included in constructing the PRS. Our study demonstrated the need to use population-specific data to build risk prediction models in CRC. The predictive accuracy of PRS developed in our study remains moderate and could be improved in future studies with a larger sample size.

Novelty and Impact:

Several polygenic risk scores (PRS) have been developed to predict colorectal cancer (CRC) risk in European-ancestry populations. Using data from large genome-wide association studies (GWAS), we developed and validated PRS in CRC risk prediction for East Asians. PRS derived from this study showed a similar discriminatory ability in classifying cases and controls as those reported in previous studies for European-ancestry populations. Our study supports the utility of PRS in identifying high-risk individuals for CRC prevention in Asians and highlight the importance of using population-specific data to build CRC risk prediction models.

Abbreviations list:

ACCC

Asia Colorectal Cancer Consortium

AUC

Area Under the Receiver Operating Characteristic Curve

CI

Confidence Interval

CRC

Colorectal Cancer

CS

Continuous Shrinkage

GCTA

Genome-wide Complex Trait Analysis

GWAS

Genome-wide Association Studies

LD

Linkage Disequilibrium

MAF

Minor Allele Frequency

OR

Odds Ratio

PC

Principal Component

PRS

Polygenic Risk Score

Data availability statement

The data that support the findings of this study are available from the corresponding author upon reasonable request.