Inclusion of a Genetic Risk Score into a Validated Risk Prediction Model for Colorectal Cancer in Japanese Men Improves Performance Free

Incidence and mortality rates for colorectal cancer in Japan both linearly increased up to the early 1990s (1). Colorectal cancer is currently the second most common cancer diagnosis and the third leading cause of cancer-related death among Japanese, with approximately 116,342 new cases in 2009 and 48,485 disease-specific deaths in 2014 (2, 3). The very high burden of this cancer encourages research on primary and secondary prevention.

Many previous epidemiologic studies have identified a number of lifestyle and environmental risk factors for colorectal cancer (4, 5). Some of these factors are established and modifiable, and it is necessary to translate knowledge of their involvement into the development of preventive strategies. A risk prediction model is a statistical tool for estimating the absolute risk that a currently healthy individual with specific risk factors will develop a future condition, such as colorectal cancer. These models help to describe the distribution of absolute risk in a specific population and to identify individuals at high risk who might benefit from intensive surveillance, chemoprevention, or other preventive measures. Several risk prediction models for colorectal cancer have been reported (6, 7), including a model that we previously developed and validated in Japanese men, which was based on two large population-based cohorts from the Japan Public Health Center-based Prospective Study (JPHC study; ref. 7). This model, which included age, body mass index (BMI), alcohol consumption, smoking status, and daily physical activity, showed modest discriminatory accuracy for estimating colorectal cancer risk and was well calibrated in a validation cohort (7). However, further improvement of its discriminatory accuracy requires the incorporation of strong risk factors into the model. In this regard, accumulating evidence from genome-wide association studies (GWAS) for colorectal cancer has raised interest in using these genetic markers (8–24). Although SNPs identified by GWASs are generally associated with only a small increase in the risk of colorectal cancer, a combination of SNPs may confer substantial impact on risk. Indeed, several studies have developed and evaluated risk prediction models that include information on both genetic markers and nongenetic predictors (25–28). However, inconsistent findings have forestalled any decision on the value of adding genetic markers to colorectal cancer risk prediction models built with nongenetic predictors.

The aim of the current study was to evaluate whether adding genetic risk scores (GRS) using GWAS-identified SNPs to our validated model for Japanese men improves the prediction of colorectal cancer risk.

The JPHC study, which was initiated in 1990 for cohort I and in 1993 for cohort II, includes 140,420 subjects (68,722 men and 71,698 women) ages 40 to 69 years living in municipalities supervised by 11 public health centers. Approximately 80% of the eligible individuals returned a self-administered baseline questionnaire on demographic characteristics, personal and familial medical histories, anthropometric factors, physical activity, smoking and drinking habits, as well as diet. About 35% provided a blood sample at the baseline survey during 1990 to 1995. Details of the JPHC study have been described elsewhere (29). This study was approved by the Institutional Review Board of the National Cancer Center (Tokyo, Japan).

In this study, we excluded subjects in two public health centers because one center (Katsushika, Tokyo) did not collect information on cancer incidence and the other (Suita, Osaka) did not have DNA samples available (29). The 32,989 participants (11,901 men and 21,088 women) who did not report any diagnosis of cancer in the baseline questionnaire and provided a blood sample were followed until the end of 2009.

Newly diagnosed cancer cases were identified from active patient notifications from the local major hospitals in the study area and data linkage with population-based cancer registries. Death certificates were used as a supplementary information source. Colorectal cancer cases were coded according to the International Classification of Diseases for Oncology, Third Edition (C18-C20), with colon cancer as C18 and rectal cancer as C19 and C20. During follow-up, 675 cases (349 men and 326 women) of colorectal cancer were newly identified.

For each case, one control was selected, using incidence density sampling, from subjects who had no prior history of colorectal cancer at the time when the case was diagnosed. Controls were matched to cases by sex, age (within 3 years), date of blood withdrawal (within 3 months), time since last meal at blood withdrawal (within 4 hours), and study location.

We selected the same risk factors and categorized smoking status and alcohol consumption as in the validated prediction model by Ma and colleagues (also see Table 1; ref. 7), except that we did not include physical activity, because it was not available from the cohort I questionnaire.

To incorporate genetic markers into the model, we selected a total of 43 risk variants that had been identified by GWAS published up until August 2014 (8–24). In the current study, we extracted data for JPHC participants included in a GWAS of colorectal cancer in Japanese, which used the Illumina 1M-duo arrays with 1,199,064 markers (21). Genotyping and quality control procedures have been described in detail elsewhere (21). Among the 43 variants, 27 had been directly genotyped and for 9 other, we used linkage disequilibrium (LD)-based proxy variants with r² greater than 0.85 identified using SNP Annotation and Proxy Search by the Broad Institute (https://www.broadinstitute.org/mpg/snap/). Seven of 43 variants that were not available on the Illumina 1M-duo array were excluded from the current analyses.

The current analysis used data from 341 cases and 329 controls from 11,883 men in the JPHC baseline survey who had available genotyping data. Only men were included because our validated prediction model by Ma and colleagues (7) was applicable to men only.

Genotype frequencies were tested for deviation from Hardy–Weinberg equilibrium with the χ² test. The main effect of each variant was assessed using HRs and 95% confidence intervals (CI) estimated from weighted Cox proportional hazards models stratified by study areas (per-allele model) adjusted for potential confounders (age, smoking status, alcohol consumption, BMI, leisure time physical activity, past history of diabetes mellitus, and family history of cancer). An inverse probability weighted Cox regression model was used to estimate HRs and 95% CIs. We calculated the inverse probability using the methods of Samuelsen (30), which estimates the sampling proportion based on the nested case–control study design. For a GRS, we selected variants that showed a statistically marginally significant association with colorectal cancer risk (P < 0.10) in the per-allele model. If several variants located on the same gene or chromosome were selected and were in strong LD (each D' > 0.9), we selected the one variant with the lowest P value. Using these selected variants, we constructed a weighted GRS to measure the cumulative effect of multiple genetic risk variants. The number of risk alleles was weighted by the log-transformed per allele HR and summed across all selected variants to create the weighed GRS. The weighted GRS was divided into quintile categories based on control distribution.

We developed three models: a nongenetic model that included the same variables as those used in Ma's model except physical activity; a genetic model that included age and GRS; and an inclusive model that included variables used in both nongenetic and genetic models (Table 1). HRs and 95% CIs for colorectal cancer according to variables included in the three models were calculated using weighted Cox proportional hazards models. Ten-year absolute risk for colorectal cancer in men was calculated using the Breslow estimator of cumulative hazards using the estimates of weighted Cox regression analysis (31).

The prediction models were evaluated on the basis of improvement of model discrimination. To compare model discrimination, we calculated the c-statistic, the integrated discrimination improvement (IDI) and continuous net reclassification improvement (NRI; ref. 32). We also evaluated classification accuracy using quantified differences in classification by NRI. The risk category of NRI was arbitrarily defined as <1.5%, ≥1.5% and <3%, and ≥3% 10-year absolute risk of developing colorectal cancer. Because of the nested case–control sampling and matching design, we calculated all of these measures weighted using the weight of inversed sampling probability in a similar way to the weighted Cox regression model (33). We compared the c-statistic, IDI, continuous NRI, and NRI of the three models (nongenetic, genetic model, and inclusive models) using estimates primarily from averages from the 5-fold cross-validation method and secondarily from the full sampled cohort.

Calibration was assessed in the full sampled cohort by comparing the predicted and observed number of events with the Grønnesby and Borgan goodness-of-fit test (GB test; refs. 34, 35), which is similar to the Hosmer–Lemeshow test for logistic regression models. To implement the GB test, all subjects were divided into 10 groups according to the weighted deciles of predicted absolute risks in the control subjects. We added the group indicator variables into the Cox proportional hazards model with the risk factors and performed an overall test for the group indicator variables using the Wald test with robust variance. The GB test for the nested case–control study design is based on weighted martingale residuals. The 10-year observed and predicted absolute risks in each group were also estimated by the weighted Kaplan–Meier method and the weighted average of predicted absolute risks, respectively.

All reported P values are two sided, and significance level was set at P < 0.05. All statistical analyses were performed using SAS 9.3 (SAS Institute Inc.).

The association between risk variants identified by GWAS and the risk of colorectal cancer is shown in Supplementary Table S1. Among 36 variants, we observed a statistically marginally significant association for 10 variants (P < 0.10). The GRS was constructed using six variants among them (rs6983267, rs3802842, rs1035209, rs12241008, rs174537, and rs4939827) after excluding four variants that were in strong LD.

Table 1 shows HRs and 95% CIs for variables included in the nongenetic, genetic, and inclusive models. Overall, increasing age and BMI, alcohol drinking, and smoking were associated with an increased risk of colorectal cancer, although HRs for BMI and smoking were not statistically significant. We found a statistically significant association between the weighted GRS and the risk of colorectal cancer. Compared with the lowest quintile category, HR (95% CI) for the highest category was 4.15 (2.21–7.79) in the inclusive model. Similar associations were observed for both colon and rectal cancer (Supplementary Table S2).

We observed a mean c-statistic of 0.60 in the nongenetic model and slightly higher c-statistics for the genetic (0.63) and inclusive models (0.66; Table 2). We found an improvement in the inclusive model compared with the nongenetic model for the mean IDI (0.015) and mean continuous NRI (0.39). We assumed three risk categories defined by cut-off values for 1.5% and 3%. The mean NRI value (0.12) showed an improvement when the nongenetic and inclusive models were compared (Table 2). A similar pattern was observed for colon and rectal cancer, except for the mean NRI values for the inclusive model (Supplementary Table S3). In addition, results from the full sampled cohort were similar to those in Table 2 (Supplementary Table S4). The c-statistic for the inclusive model (0.68; 95% CI, 0.63–0.73) was slightly greater than that for the nongenetic model (0.62; 95% CI, 0.57–0.67). We found a statistically significant improvement in the inclusive model compared with the nongenetic model for the IDI (0.0052; 95% CI, 0.0023–0.0081), continuous NRI (0.36; 95% CI, 0.0023–0.71), and NRI (0.26; 95% CI, 0.0039–0.43).

The differences between the observed and predicted absolute risks by decile category are shown in Fig. 1. The P values for GB test were 0.14 for the nongenetic model, 0.11 for the genetic model, and 0.09 for the inclusive model. The P values by subsite (colon and rectal) were also not statistically significant for all three models in either colon or rectal cancer.

We estimated the absolute risk for developing colorectal cancer for men according to risk factor profiles and age (Table 3). The 10-year absolute risk for developing colorectal cancer varied by risk factor profile, for example, the absolute risks for a 40-year-old man in the lowest and highest risk groups were 0.16% and 3.3%, respectively, whereas the corresponding ones for a 70-year-old man were 0.53% and 11%.

This study found a statistically significant association between a weighted GRS constructed using six GWAS-identified risk variants and the risk of colorectal cancer in Japanese men in the JPHC study. This weighted GRS was a significant predictor of colorectal cancer risk. The addition of the weighted GRS to the nongenetic model predicted the risk of colorectal cancer slightly better than the nongenetic model alone based on the mean c-statistic. Moreover, the mean IDI and NRI showed improvements when compared between the nongenetic and inclusive models. Our findings provide support for including genetic risk information into colorectal cancer risk assessment models to more accurately stratify men in low- to high-risk groups.

Few studies have investigated the degree of improvement in prediction models for colorectal cancer obtained by including GRSs, and their findings have been inconsistent (25–28). Dunlop and colleagues reported very modest discriminative performance for two models, one incorporating SNP genotypes at all 10 loci alone and the second in combination with gender, age, and family history data: The average AUC for 10 iterations in cross-validation analysis was 0.57 for the former and 0.59 for the latter (25). Similarly, Hsu and colleagues compared a model that included both family history and a GRS based on 27 GWAS risk variants with a model that included family history only (26). Although AUC values were very modest, a statistically significant improvement was shown in men (from 0.51 to 0.59) but not in women (from 0.52 to 0.56). However, both studies estimated the 10-year absolute risk of colorectal cancer based on the models, which included information on common genetic variants for colorectal cancer, and suggested the practicality of identifying a high-risk group (25, 26). Regarding models that include lifestyle and environmental factors, Jung and colleagues failed to show a statistically significant improvement in AUC for a model that included both a GRS based on 7 variants selected from 23 GWAS-identified variants and traditional risk factors, such as age, sex, fasting serum glucose, exercise, and family history (AUC = 0.74), compared with a model that included traditional risk factors only (AUC = 0.73; ref. 27). Hosono and colleagues developed a risk model that included both a risk score based on 6 genetic variants selected from 23 GWAS-identified risk variants and established risk factors such as age, smoking, alcohol intake, BMI, exercise, dietary folate intake, and family history and evaluated the discriminatory accuracy of the model by the AUC method in both the derivation and replication population (28). Compared with the model that included established risk factors only, the discriminatory accuracy for the model that included both the risk score and established risk factors was significantly improved in the derivation population, from 0.7009 to 0.7167, whereas the improvement in the replication population, from 0.6391 to 0.6365, was not statistically significant.

Our validated model reported by Ma and colleagues showed modest discriminatory accuracy in a cohort for model development (c-statistic = 0.70) and in a validation cohort (c-statistic = 0.64; ref. 7). However, the mean c-statistic for a 5-fold cross-validation analysis was 0.60 for the nongenetic model in this study. This difference in c-statistics between the previously reported model and the current nongenetic model might be explained by a difference in the study populations, as the current analysis was restricted to participants who provided blood (∼35%) for the baseline survey when they underwent health checkups. We previously reported that health checkup examinees had a different socioeconomic status than nonexaminees and a more favorable lifestyle profile (36). In addition, the current model did not include daily physical activity due to a lack of information on this variable in the cohort I questionnaire.

The current genetic model showed a somewhat higher mean c-statistic (0.63) than those from the previous studies, namely 0.57 by Dunlop and colleagues and 0.60 by Hsu and colleagues (25, 26). In our current study, genetic variants were selected from the same dataset that was used to calculate the c-statistic. As a result, the c-statistic by the full sampled cohort (0.66) was likely to be inflated, although mean c-statistic in 5-fold cross-validation analysis was 0.63. Moreover, it will be important to replicate the association between the GRS and risk of colorectal cancer in an independent population. Hosono and colleagues constructed a risk score based on 6 variants selected from 23 GWAS-identified risk variants and observed a statistically significant association between the risk score group and the risk of colorectal cancer in the derivation population but not in the replication population (28). Therefore, the lack of a replication study for the GRS is a major limitation of the current study.

In the current study, the mean IDI and NRI values for comparison of the nongenetic model with the inclusive model showed improvements in model discrimination. Indeed, in our nested case–control subjects, the number of subjects in the low (<1.5%) and high (≥3.0%) risk categories were n = 162 (25.7%) and n = 84 (13.3%) by the nongenetic model versus n = 225 (35.7%) and n = 181 (28.7%) by the inclusive model, respectively. Moreover, the range of predicted values by the inclusive model (0.002–0.078) was somewhat wider than that by the non-genetic model (0.004–0.048). Although the number of subjects in the very high-risk group was relatively small [i.e., 40 subjects (6.4%) with ≥5% by the inclusive model], these findings suggest that this model or one expanded with additional risk factors could be useful for identifying subgroups of Japanese men at an elevated risk of colorectal cancer who might benefit most from screening and other prevention strategies. On this basis, start age at screening, and screening modality and intervals could be set according to the individual's absolute risk. The generalizability of this model for possible use as an absolute risk estimate for Japanese men should be further ascertained. In particular, the absolute risk provided by the model was based on a baseline risk of colorectal cancer estimated by participants in the JPHC study, which would be applicable only to populations with comparable rates. In this regard, although incidence rates of colorectal cancer from a nationwide cancer registry would be fully representative of the Japanese population and therefore desirable for estimating absolute risk, the only incidence rates for Japan can be estimated from high-quality data collected by selected local population-based cancer registries. In addition, the current study ignores the additional costs of obtaining DNA and of genotyping when comparing the models, even allowing for the fact that an improvement in risk prediction was found. Any practical application of this model will require investigating whether the improvement in risk prediction with the addition of GRS outweighs the additional costs.

In summary, this study showed that the addition of a GRS including six GWAS-identified genetic variants to our validated lifestyle-based model for Japanese men improved the prediction of colorectal cancer risk. This model might allow the stratification of Japanese men into low- and high-risk groups.

No potential conflicts of interest were disclosed.

Conception and design: M. Iwasaki, S. Tsugane

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): M. Iwasaki, N. Sawada, T. Shimazu, H. Wang, L. Le Marchand, S. Tsugane

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): M. Iwasaki, S. Tanaka-Mizuno, A. Kuchiba, T. Yamaji, A. Goto, T. Shimazu, S. Tsugane

Writing, review, and/or revision of the manuscript: M. Iwasaki, S. Tanaka-Mizuno, A. Kuchiba, T. Yamaji, A. Goto, T. Shimazu, S. Sasazuki, H. Wang, L. Le Marchand, S. Tsugane

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): H. Wang, L. Le Marchand, S. Tsugane

Study supervision: S. Tsugane

We are indebted to the Aomori, Iwate, Ibaraki, Niigata, Osaka, Kochi, Nagasaki, and Okinawa Cancer Registries for providing their incidence data.

Japan Public Health Center-based Prospective Study (JPHC Study) investigators are listed below:

Principal investigator: S. Tsugane.

Group members: S. Tsugane, N. Sawada, M. Iwasaki, S. Sasazuki, T. Yamaji, T. Shimazu, A. Goto, A. Hidaka and T. Hanaoka, National Cancer Center, Tokyo; J. Ogata, S. Baba, T. Mannami, A. Okayama and Y. Kokubo, National Cerebral and Cardiovascular Center, Osaka; K. Miyakawa, F. Saito, A. Koizumi, Y. Sano, I. Hashimoto, T. Ikuta, Y. Tanaba, H. Sato, Y. Roppongi, T. Takashima and H. Suzuki, Iwate Prefectural Ninohe Public Health Center, Iwate; Y. Miyajima, N. Suzuki, S. Nagasawa, Y. Furusugi, N. Nagai, Y. Ito, S. Komatsu and T. Minamizono, Akita Prefectural Yokote Public Health Center, Akita; H. Sanada, Y. Hatayama, F. Kobayashi, H. Uchino, Y. Shirai, T. Kondo, R. Sasaki, Y. Watanabe, Y. Miyagawa, Y. Kobayashi, M. Machida, K. Kobayashi and M. Tsukada, Nagano Prefectural Saku Public Health Center, Nagano; Y. Kishimoto, E. Takara, T. Fukuyama, M. Kinjo, M. Irei and H. Sakiyama, Okinawa Prefectural Chubu Public Health Center, Okinawa; K. Imoto, H. Yazawa, T. Seo, A. Seiko, F. Ito, F. Shoji and R. Saito, Katsushika Public Health Center, Tokyo; A. Murata, K. Minato, K. Motegi, T. Fujieda, S. Yamato and M. Doi, Ibaraki Prefectural Mito Public Health Center, Ibaraki; K. Matsui, T. Abe, M. Katagiri and M. Suzuki, Niigata Prefectural Kashiwazaki and Nagaoka Public Health Center, Niigata; M. Doi, A. Terao, Y. Ishikawa and T. Tagami, Kochi Prefectural Chuo-higashi Public Health Center, Kochi; H. Sueta, H. Doi, M. Urata, N. Okamoto, F. Ide, H. Goto, R. Fujita and Y. Sou, Nagasaki Prefectural Kamigoto Public Health Center, Nagasaki; H. Sakiyama, N. Onga, H. Takaesu, M. Uehara, T. Nakasone and M. Yamakawa, Okinawa Prefectural Miyako Public Health Center, Okinawa; F. Horii, I. Asano, H. Yamaguchi, K. Aoki, S. Maruyama, M. Ichii and M. Takano, Osaka Prefectural Suita Public Health Center, Osaka; Y. Tsubono, Tohoku University, Miyagi; K. Suzuki, Research Institute for Brain and Blood Vessels Akita, Akita; Y. Honda, K. Yamagishi, S. Sakurai and N. Tsuchiya, University of Tsukuba, Ibaraki; M. Kabuto, National Institute for Environmental Studies, Ibaraki; M. Yamaguchi, Y. Matsumura, S. Sasaki, and S. Watanabe, National Institute of Health and Nutrition, Tokyo; M. Akabane, Tokyo University of Agriculture, Tokyo; T. Kadowaki and M. Inoue, The University of Tokyo, Tokyo; M. Noda and T. Mizoue, National Center for Global Health and Medicine, Tokyo; Y. Kawaguchi, Tokyo Medical and Dental University, Tokyo; Y. Takashima and Y. Yoshida, Kyorin University, Tokyo; K. Nakamura, Niigata University, Niigata; and R. Takachi, Nara Women's University, Nara; J. Ishihara, Sagami Women's University, Kanagawa; S. Matsushima and S. Natsukawa, Saku General Hospital, Nagano; H. Shimizu, Sakihae Institute, Gifu; H. Sugimura, Hamamatsu University School of Medicine, Shizuoka; S. Tominaga, Aichi Cancer Center, Aichi; N. Hamajima, Nagoya University, Aichi; H. Iso and T. Sobue, Osaka University, Osaka; M. Iida, W. Ajiki, and A. Ioka, Osaka Medical Center for Cancer and Cardiovascular Disease, Osaka; S. Sato, Chiba Prefectural Institute of Public Health, Chiba; E. Maruyama, Kobe University, Hyogo; M. Konishi, K. Okada, and I. Saito, Ehime University, Ehime; N. Yasuda, Kochi University, Kochi; S. Kono, Kyushu University, Fukuoka; S. Akiba, Kagoshima University, Kagoshima; T. Isobe, Keio University, Tokyo; Y. Sato, Tokyo Gakugei University, Tokyo.

This study was supported by the National Cancer Center Research and Development Fund (23-A31[toku] and 26-A-2 since 2011, and 25-A-14), a Grant-in-Aid for Cancer Research from the Ministry of Health, Labor and Welfare of Japan (from 1989 to 2010), Practical Research for Innovative Cancer Control (15ck0106177h0001) from the Japan Agency for Medical Research and Development, AMED, and a Grant-in-Aid for Scientific Research (15K15956) from the Ministry of Education, Culture, Sports, Science, and Technology of Japan. The genotyping data were generated under grant CA126895 from the NCI.

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.