Evaluating the Potential of Polygenic Risk Score to Improve Colorectal Cancer Screening

This article is featured in Highlights of This Issue, p. 1247

Colorectal cancer is recognized as one of the cancers with the highest incidence and associated mortality worldwide (1). It is generally acknowledged that a vast majority of colorectal cancer cases develop from nonmalignant precursor adenomas (2). The average duration of the development of an adenoma to colorectal cancer transition is unobserved, but it is estimated to take at least 10 years (3). This long latent phase provides an excellent window of opportunity for early detection. Therefore, colorectal cancer is particularly suitable for screening. Keeping in mind the dimension of this disease, European national health systems have started population screening programs to increase early detection and improve prevention measures. Screening for colorectal cancer offers the possibility to identify the disease at an earlier stage or at a premalignant phase. For this reason, the evidence-based European Code Against Cancer recommended that men and women over 50 years of age should participate in colorectal cancer screening. This was given effect within the European Union (EU) by the 2003 Council Recommendation on cancer screening (4).

Indeed, colorectal cancer is highly preventable by detecting and removing adenomas through colonoscopy screening, but this procedure is very costly to be implemented as population screening and has an associated morbidity (5). Intermediate screens to detect occult blood in feces such as fecal immunochemical testing (FIT) are therefore often used to select patients for colonoscopy with suboptimal sensitivity (6, 7). This two-step strategy for colorectal cancer screening is the most common worldwide (8), but results in a high false positive rate due to the suboptimal specificity of the occult blood detection, necessitating unnecessary colonoscopies (9). Therefore, additional biomarkers added to the first step to the current scenario could improve colorectal cancer screening.

As for other complex diseases, colorectal cancer is caused by both genetic and environmental factors (10). Twin studies showed that around 13% to 30% of the variation in colorectal cancer susceptibility involves inherited genetic differences (11, 12). Some of the known colorectal cancer predisposition factors were already discovered in the past two decades through genome-wide association studies (13, 14). Right after their identification, the hope was raised for genetic profiling using the combination of these common, low-penetrance genetic variants to be able to identify high-risk individuals in the population that could benefit from preventive and therapeutic interventions (15). Indeed, polygenic risk scores (PRS) combining the individual, weak effects on disease risk have been developed in the past for common diseases such as colorectal cancer. Their predictive potential was limited, most likely evidencing their usefulness but their shortcomings when used alone without other clinical or environmental data (16, 17). PRS models for colorectal cancer were developed by using individual genome-wide association study (GWAS) genetic variants (from 10 to more than 100) but have recently incorporated genome-wide data to improve risk prediction by using all available SNPs genotyped in GWAS (18). Certainly, genome-wide PRS have proven to identify individuals with risk equivalent to monogenic mutations (19), which could justify its application in health care systems.

Using PRS to screen the population at medium risk for colorectal cancer is an attractive alternative to improve current results in this setting. Frampton and colleagues demonstrated that personalized screening programs for colorectal cancer, in which eligibility was based on PRS in addition to age, had the potential to greatly reduce the number of individuals screened while still detecting nearly as many cases (20). Some more recent studies have also tested the potential application of PRS on colorectal cancer screening programs, showing its value to define a personalized, risk-adapted starting ages for screening (17, 21) or personalized screening intervals after negative findings from colonoscopy (22).

In this study, almost 3,000 individuals from the Barcelona colorectal cancer screening program were enrolled. We were able to develop a PRS and evaluate its effectiveness in determining which individuals should undergo a colonoscopy after a positive FIT result considering their colorectal cancer–associated genetic risk. This study is the first investigation analyzing PRS in a two-step colorectal cancer screening program.

Our study included 2,893 subjects aged between 50 to 69 years. Participants were recruited initially in 2009 (first round) and then during 2017 to 2019 (second round) from the Barcelona colorectal cancer screening program with a positive FIT in 3 different hospitals (23). In both rounds, individuals with a FIT-positive result (≥20 μg of hemoglobin/g of feces) were advised to undergo a colonoscopy.

The histologic classification of polyps and cancer was based on World Health Organization (WHO) criteria (24) and additional evidence (25), as summarized in Table 1. It included low-risk adenomas (LRA) and high-risk adenomas (HRA). All invasive carcinomas (stages I–IV) were classified as colorectal cancer as well as carcinoma in situ (CIS) although currently this latter lesion is classified as HRA. While developing this study, the intermediate risk adenoma (IRA) was incorporated as a new category in the European Guideline for Quality Assurance in Colorectal Cancer Screening and Diagnosis (25, 26). However, before its publication IRA was included in the HRA group. Our cohort was divided in two groups (cases and controls) taking into account the outcome of the colonoscopy and their link with colorectal cancer and their different clinical surveillance after its finding. Therefore, the cases group included high-risk lesions (HRL) such as colorectal cancer, intramucosal carcinoma, HRA, IRA, and polyposis cases, whereas the control group comprised LRA and normal examination after colonoscopy (27).

The cases group included 1,221 subjects, comprising 755 men and 466 women. Controls were 1,672 subjects with no findings relevant for colorectal cancer and included 795 men and 877 women. Cases and controls included in this study were negative for a family history of hereditary or familial colorectal cancer or adenomatous polyposis (i.e., ≥2 first-degree relatives with colorectal cancer/adenomatous polyposis or 1 first-degree relative diagnosed before the age of 60). Cases with polyposis corresponded to de novo cases. Subjects with inflammatory bowel disease were also excluded.

Standard extraction procedures were used to obtain DNA from frozen peripheral blood for all samples. Mutations in genes for germline predisposition to colorectal cancer with a high-penetrance could not be excluded. However, since cases did not report a relevant family history of colorectal cancer or adenomatous polyposis, the role of germline predisposition in cases can be considered less likely and all the participants were over 50 years of age. Sex, age, and FIT values of all individuals included in the study are summarized in Table 2. This study was conducted in alignment with the Declaration of Helsinki and approved by the corresponding institutional ethics committee (Hospital Clínic of Barcelona; Barcelona, Spain; HCB/2017/0193) and written informed consent was obtained from all individuals.

We selected and genotyped 65 SNPs previously linked with colorectal cancer risk in all available DNA samples. We did not observe any inconsistency in the genotyping results between the two inclusion rounds. We selected SNPs from association results of previous GWAS, which were available before mid-2017 (www.ebi.ac.uk/gwas/, GWAS Catalog; Supplementary Table S1). We used the Biomark 96.96 Genotyping dynamic array (Fluidigm) and TaqMan assays (Thermo Fisher Scientific) to perform genotyping. Fluidigm SNP Genotyping Analysis and PLINK software were used to assess the quality of the data. We eliminated samples and SNPs with a genotyping success rate below 90% from following analyses (including rs7229639, rs3764482). We also eliminated a monomorphic SNP (rs10904849). We also tested genotyping quality by duplicating 26 samples. In this case, 100% genotype concordance was achieved. Hardy–Weinberg equilibrium (HWE) checking for deviation of the genotype frequencies in the controls from those expected was assessed by the χ² test (1df). Each SNP was in HWE in controls, with the exception of rs647161 and rs12603526 (P = 9.28e-04 and 6.42e-04, respectively). The possibility of genotyping artifacts was excluded also after manually inspecting genotype assignments for them. Therefore, both SNPs were not excluded. After data quality filtering, the final number of available samples was 2,829 (1,200 cases and 1,629 controls) and 62 SNPs. In the remaining individuals, 99.07% was the overall genotyping success rate.

We first evaluated the association between each SNP and colorectal cancer through unconditional logistic regression to observe how these SNPs behaved in our population and whether they followed the trend of previously reported studies. Models were unadjusted to capture the raw association of each SNP with colorectal cancer. ORs and 95% confidence intervals (CI) were derived from the model.

We then calculated a polygenic risk score (PRS). This score was defined as the count of risk alleles across all available SNPs. We used all 62 SNPs and not only those that were significant in our cohort, since all the SNPs selected in this study were previously shown to be associated with colorectal cancer. The total number of risk alleles was calculated for all samples and coded as 0, 1, or 2 for each SNP assuming an additive genetic effect. To allow for missing values in some SNPs, the PRS values were proportionally rescaled according to the number of non-missing SNPs.

An unweighted PRS was preferred since the published effects of each SNP were similar. We also explored the weighted model derived from the GWAS publications (used ORs in Supplementary Table S2) and the model fitted to our data, but the results were essentially the same, and the weights may be biased due to the winner's course effect. Therefore, we opted to use the unweighted score. A two-sided t test was applied to compare between cases and controls. Afterward, PRS effect was fitted in a logistic regression model to assess for genetic susceptibility comparing cases and controls. We also performed a comparison with extreme phenotypes (CRC and HRA vs. negative colonoscopy) excluding subjects with other diagnosis.

Next, we explored the combination of PRS combined with age, sex, and the FIT value to see if we could improve its predictive capacity and if it would be worthwhile to incorporate the PRS in the actual colorectal cancer screening program. To do so, we tested four different models. The first one was calculated taking into account sex and age, the second one taking into account sex, age, and PRS, the third one taking into account sex, age, and FIT value, and the fourth was a combined model taking into account all variables (sex, age, FIT value, and PRS). FIT value was used as a qualitative variable cutting it in three categories. All models were developed using a general linear model.

The predictive accuracy of the different models was assessed with the area under the ROC curve (AUROC). We also compared AUROC values in the different models using the Delong test. To account for potential overfitting that could overestimate the effect of the different models, a 10-fold cross-validation was used to estimate AUROC. The cross-validated AUROC and Delong P were calculated with all samples combined. In each step of the cross-validation the model was estimated with 90% of the samples and predictions applied to the remaining 10%. This step was repeated 10 times. The AUROC and Delong P were calculated with the combined predictions of the 10 steps. We then selected four cut-offs (P10, P20, P80, and P90). Subsequently, we calculated the positive and negative predictive value of our “combined model” by comparing the observed case/control status and the predicted probability estimated by our “combined model”.

All analyses were performed using PLINK v1.09 and R statistical software (version 3.4.1, R Foundation for Statistical Computing). All statistical tests were two-sided, and P values < 0.05 were considered statistically significant. Bonferroni correction was used for multiple testing adjustment (P_adjusted = 0.05/62, SNPs = 8.06e-04).

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

The cases group included 1,221 subjects, comprising 755 men and 466 women. Controls were 1,672 subjects with no relevant findings related to colorectal cancer and included 795 men and 877 women. Afterward, a total of 1,200 cases and 1,629 controls were successfully genotyped for 62 SNPs previously linked with genetic susceptibility to colorectal cancer. Firstly, the frequency of colorectal cancer risk alleles between cases and controls was compared and those significantly enriched in the cases cohort were detected (Supplementary Table S2). 13 colorectal cancer SNPs showed statistically significant associations with colorectal cancer (rs10936599, rs704017, rs1035209, rs174537, rs1535, rs174550, rs3217810, rs4939827, rs2241714, rs1800469, rs961253, rs6061231, and rs4925386) and these genetic associations were similar to those previously reported for colorectal cancer genetic susceptibility. Despite being not significant, most remaining SNPs (47/62), showed ORs in the same directions as those previously described in the literature for colorectal cancer susceptibility.

We then calculated a PRS and compared the number of colorectal cancer risk alleles between cases and controls. Figure 1 shows the distribution of risk by allele number for the 62 genotyped SNPs, both for cases and controls. Risk alleles followed a normal distribution in both cases and controls. It was apparent a shift towards a higher number of risk alleles in cases coherent with a cumulative impact of colorectal cancer risk alleles. The mean number of risk alleles in control individuals was 56.47 compared with 57.54 in cases and there was a statistically significant difference in the mean number of risk alleles between cases and controls (difference: −1.07; two-sided t test P = 1.6e-07).

Then, we calculated PRS for cases and controls taking into account the cumulative number of colorectal cancer risk alleles. We took into account 56, the median number of risk alleles in controls, as reference. We grouped cases and controls when subjects carried ≤46 risk alleles and ≥70 alleles, since they corresponded to a small number of subjects. We observed that the risk score was higher in the cases group compared with the control cohort (per allele OR = 1.04; 95% CI, 1.02–1.06; P < 0.0001). A 2-fold increase in colorectal cancer risk for subjects in the highest decile of risk alleles (≥65), compared with those in the first decile (≤54; OR = 2.22, 95% CI 1.59–3.12; P < 0.0001) was also detected. As presented in Fig. 1, a linear increase in risk per allele was apparent, indicating the independent additive contribution of each allele to predispose for a HRL. We also checked that there was no association between PRS and age (OR = 0.99; 95% CI, 0.92–1.07; P = 0.86) and sex (OR = 0.99; 95% CI, 0.91–1.07; P = 0.75). Results shown here correspond to the unweighted model. We also explored the weighted model derived from the GWAS publications and the model fitted to our data, but the results were essentially the same, and the weights may be biased due to the winner's course effect (Supplementary Fig. S1). Finally, a comparison of extreme phenotypes (colorectal cancer and HRA as cases vs. negative colonoscopy as controls) was also explored (Supplementary Fig. S2). When doing so, results did not change relevantly but showed some improvement. For instance, we detected that the risk score was higher in the cases group compared with the control cohort (OR = 1.05; 95% CI, 1.01–1.08; P = 0.015). We also observed that there was a 3-fold increase in colorectal cancer risk was detected for subjects in the highest decile of risk alleles (≥64), compared with those in the first decile (≤51; OR = 3.15; 95% CI, 1.06–9.41, P = 0.04).

Predictive modeling to discriminate between cases and controls was performed considering PRS along with age and sex, since although age and sex are not associated with PRS, they modify the risk of developing colorectal cancer. FIT value for each individual was also included in the model.

The results achieved in the PRS-based predictive model (AUROC: 0.614; 95% CI, 0.593–0.635) were superior to those obtained using only age and sex (AUROC: 0.597; 95% CI, 0.576–0.618) but lower than those obtained using age, sex, and FIT value (AUROC: 0.626; 95% CI, 0.605–0.647). Moreover, the results were statistically significant when compared with the baseline model (sex-age–based model). However, the model resulting from the combination of sex, age, FIT value, and PRS reached the highest accuracy for identifying patients with a HRL (AUROC: 0.639; 95% CI, 0.619–0.660) when compared with participants with nonrelevant findings for colorectal cancer risk at colonoscopy, illustrating the potential usefulness of incorporating PRS into the current colorectal cancer screening scenario. Results are shown in Table 3 and Fig. 2.

A 10-fold cross-validation was also performed. All 10-fold cross-validated AUROC were very similar than the direct estimate of the model not supporting overfitting (Supplementary Table S3).

We then calculated the sensitivity, specificity, predictive positive and predictive negative values for four cut-offs (P10, P20, P80, P90) for the sex-age-FIT– and PRS-based model. As shown in Table 4, our positive predictive value (PPV) ranged from 44.25% to 61.75% and our negative predictive value (NPV) ranged from 74.11% to 59.75%. Therefore, when the cut-off point is low, the capacity to predict true positives is lower whereas the ability to detect true negatives is higher. On the other hand, as the cut-off point increases, PPV becomes higher and the NPV is reduced. For instance, comparing with the current scenario where every FIT-positive individual undergoes a colonoscopy, the P10 cut-off shows the impact in results by reducing 10% the number of colonoscopies when applying the sex-age-FIT and PRS-based model. By doing so, there will be still a sensitivity of approximately 94% (94% of individuals who have any relevant colorectal cancer lesion will test positive) and an improvement in specificity of approximately 13% (13% of additional individuals without any relevant colorectal cancer lesion will test negative and will not undergo a colonoscopy). Below this cut-off, the negative likelihood ratio (false negative/true negative) would be 0.48 and 209 unnecessary colonoscopies will not be performed (true negatives) but 73 HRL would not be detected (false negatives) only two of them being colorectal cancer. These results from our model show that number of colonoscopies could be significantly reduced by applying our model incorporating PRS and still maintain acceptable screening results.

As previously mentioned, we also explored a comparison of extreme phenotypes (colorectal cancer and HRA as cases vs. negative colonoscopy as controls). We calculated AUROC values (Supplementary Fig. S3) and discriminative capacity including also a 10-fold cross-validation (Supplementary Tables S4 and S5). Again, results did not change relevantly but showed some improvement, an outcome likely related to considering phenotypically extreme individuals. This outcome could be expected since SNPs used in PRS were detected by comparing colorectal cancer cases and controls in previous GWAS and not HRA, IRA, or other HRLs. However, since a smaller set of samples was used, results were not as significant.

Estimation of colorectal cancer risk can determine who should be screened. Being screened can result in prevention, by detecting and removing precancerous lesions and colorectal cancer, providing an earlier effective treatment for this disease. Screening includes FIT and colonoscopy, and ideally they should be tailored based on personal risk. FIT is inexpensive, safe, but less sensitive, whereas colonoscopy is expensive, invasive and more risky. Colorectal cancer has a heritable fraction being in most cases polygenic with thousands of genetic variants contributing to its development. Therefore, utilizing colorectal cancer variants to predict risk holds promise for risk stratification for primary and secondary prevention. Our study has explored the potential utility of PRS to improve colorectal cancer screening through genotyping 62 SNPs previously associated to colorectal cancer in 2,829 individuals from a colorectal cancer screening program.

We first evaluated the association between each SNP and colorectal cancer. Our results showed that most of the SNPs showed ORs in the same directions as previously described in the literature for colorectal cancer susceptibility, indicating that most of the SNPs used in our study behaved as expected in our population, even those SNPs described in Asian populations.

We then compared the median number of risk alleles between cases and controls to evaluate if there was a high number of risk alleles in the case group. The obtained results suggest that there is a significant increase in the number of risk alleles in the case group compared with controls. Moreover, by comparing the individuals in the highest decile of risk alleles with those in the first decile, we observed that there was almost a 2-fold increase in colorectal cancer risk (OR = 2.22; 95% CI, 1.59–3.12; P < 0.0001). Thus, the higher the number of risk alleles, the higher the risk of having a HRL. Based on this, it seems feasible to identify a subgroup of participants from the Barcelona colorectal cancer screening program whose risk of having precancerous lesions or colorectal cancer is explained by a high PRS, as to warrant the application of specific surveillance measures (regular colonoscopies) to those individuals exceeding a defined value or PRS cut-off.

Regarding our predictive model for colorectal cancer screening, the directly estimated results of our models evidenced an improvement of the discrimination compared to the current predictive model used in the colorectal cancer screening program and based on FIT alone. A 10-fold cross-validation was also performed. All 10-fold cross-validated AUROC were very similar than the direct estimate of the model not supporting overfitting. However, we believe that PRS should not be considered the best biomarker to discriminate between cases and controls in a two-step colorectal cancer screening program although it shows real potential to be used along with sex, age, FIT value, and other additional biomarkers. It could be also hypothesized that a PRS-only model could be used in a different scenario to identify those individuals with a high number of risk alleles who would benefit from starting the screening program before the age of 50.

Most studies previously developed using PRS have used individuals from case-control or cohort studies not recruited specifically in a colorectal cancer screening scenario (16–18, 20, 28, 29). Likewise, there are few studies based on individuals from colorectal cancer screening programs where colonoscopy is used (21, 22, 30, 31). Accordingly, these previous studies focused mainly on improving the definition of risk-adapted screening ages. On the contrary, the individuals included in this study were part of the Barcelona colorectal cancer screening program where FIT was used as intermediate test to select who should undergo a colonoscopy when positive. Thus, we aimed at improving the discriminating power of the intermediate test used to better direct individuals to colonoscopy. Therefore, our study corresponds to the first investigation analyzing PRS in a two-step colorectal cancer screening program.

The number of genetic variants included to calculate PRS has increased as more colorectal cancer risk components were unfolded by GWAS and data meta-analyses over the years. Previous studies used the first 10 variants (16), 21 variants (29), 27 variants (28), 37 variants (20), 48 variants (21), 63 variants (17), and the final currently known 140 variants (18). Also, besides using the colorectal cancer risk linked to individual GWAS variants to estimate PRS, other recent approaches have escalated variant selection to use a selection of SNPs by linkage disequilibrium (10,000 variants), or even a genome-wide approach (18). Our study used 65 colorectal cancer risk variants since when genotyping started in 2017, it corresponded to the maximum number of variants known at that time. Even though our study did not used the maximum number of known colorectal cancer susceptibility genetic variants, a more modest sample size, or other additional approaches, it reached a similar AUROC values when compared with previous studies, exemplifying again the limited capacity of the colorectal cancer genetic susceptibility to be used in risk prediction. In fact, it should be expected from complex diseases such as colorectal cancer that risk prediction could not rely only on genetic susceptibility but also in other causes contributing to disease such as environmental factors or microbiome, and those should be taken into account in the risk prediction model.

Another matter of concern in PRS models is their transferability to all ancestral populations so their application in disease prediction in the clinical setting is robust. Most PRS have been developed and optimized using genotyping data from individuals of European ancestry and may be less accurate when risk prediction is applied to other populations (32). To circumvent this matter to some extent, we included three genetic variants (rs12080929, rs11987193, rs3987) associated with colorectal cancer risk and identified in previous GWAS performed in the Spanish population (33, 34).

This study has some limitations. It should be commented that our cohort sample size or the number of analyzed genetic variants to calculate PRS may probably be not large enough to reach stronger conclusions. Moving into affordable genome-wide genotyping in the near future could be possible and useful for the risk prediction of colorectal cancer and other diseases. It is also feasible that including additional information in the PRS model such as gut microbiota data could enhance risk prediction. It is a plausible hypothesis that certain colonic microbes or alterations of the typical resident colonic flora may create a microenvironment that is more favorable to tumor development (35). In that direction, microbiome analysis is being performed by using DNA extracted from FIT samples from this same cohort and hopefully it would add an additional layer of information to the conclusions reached in the present study. Also, additional information if available regarding body mass index, metabolic syndrome, smoking, anti-inflammatory drugs, and antibiotic use could surely enrich our prediction model.

In summary, our study corresponds to the first investigation analyzing PRS in a two-step colorectal cancer screening program. PRS could help improve the current results in colorectal cancer screening programs. However, its capacity is limited as proven by this study and should be complemented by additional biomarkers such as microbiome or environmental factors.

B. Bellosillo reports personal fees from Biocartis, Novartis, Qiagen, Merck-Serono, BMS, Pfizer, Amgen, Janssen; grants and personal fees from Thermo Fisher Scientific, AstraZeneca; and personal fees from Eli Lilly and Company outside the submitted work. J.M. Borràs reports grants from Department of Health during the conduct of the study. V. Moreno reports grants from Agency for Management of University and Research Grants (AGAUR) of the Catalan Government, Instituto de Salud Carlos III, cofunded by FEDER funds; and grants from Spanish Association Against Cancer (AECC) Scientific Foundation during the conduct of the study. No disclosures were reported by the other authors.

C. Arnau-Collell: Conceptualization, data curation, software, formal analysis, validation, investigation, visualization, methodology, writing–original draft. A. Díez-Villanueva: Data curation, software, formal analysis, validation, investigation, visualization, methodology, writing–review and editing. B. Bellosillo: Resources, data curation, validation, investigation, visualization, methodology, writing–review and editing. J.M. Augé: Resources, data curation, formal analysis, validation, investigation, visualization, methodology, writing–review and editing. J.Muñoz: Data curation, validation, investigation, visualization, methodology, writing–review and editing. E.Guinó: Resources, data curation, software, formal analysis, validation, investigation, visualization, methodology, writing–review and editing. L. Moreira: Resources, data curation, formal analysis, validation, investigation, visualization, methodology, writing–review and editing. A. Serradesanferm: Resources, data curation, validation, investigation, visualization, writing–review and editing. À. Pozo: Resources, data curation, validation, investigation, visualization, writing–review and editing. I.Torà-Rocamora: Resources, data curation, validation, investigation, writing–review and editing. L. Bonjoch: Resources, validation, investigation, visualization, methodology, writing–review and editing. G.Ibañez-Sanz: Resources, data curation, software, validation, investigation, visualization, methodology, writing–review and editing. M. Obon-Santacana: Resources, data curation, software, validation, investigation, visualization, methodology, writing–review and editing. F. Moratalla-Navarro: Resources, data curation, software, validation, investigation, visualization, methodology, writing–review and editing. R. Sanz-Pamplona: Resources, data curation, software, validation, investigation, visualization, methodology, writing–review and editing. C. MárquezMárquez: Resources, data curation, investigation, visualization, writing–review and editing. R. Rueda Miret: Resources, data curation, investigation, visualization, writing–review and editing. R. Pérez Berbegal: Resources, data curation, investigation, visualization, writing–review and editing. G. Piquer Velasco: Resources, data curation, investigation, visualization, writing–review and editing. C. HernándezRodríguez: Resources, data curation, investigation, visualization, writing–review and editing. J. Grau: Resources, data curation, supervision, investigation, project administration, writing–review and editing. A. Castells: Resources, formal analysis, supervision, funding acquisition, investigation, project administration, writing–review and editing. J.M. Borràs: Resources, data curation, supervision, funding acquisition, investigation, project administration, writing–review and editing. X. Bessa: Resources, data curation, supervision, funding acquisition, investigation, project administration, writing–review and editing. V. Moreno: Conceptualization, resources, data curation, software, formal analysis, supervision, funding acquisition, validation, investigation, visualization, methodology, writing–original draft, project administration. S. Castellví-Bel: Conceptualization, resources, formal analysis, supervision, funding acquisition, validation, investigation, visualization, methodology, writing–original draft, project administration, writing–review and editing. CRIPEV Consortium: Resources, data curation, funding acquisition, investigation, project administration, writing–review and editing.

This study was funded by the Strategic Plan for Health Research and Innovation - PERIS program 2016–2020 (SLT002/16/00398, Generalitat de Catalunya; to all authors), and complemented with grants from Fondo de Investigación Sanitaria/FEDER (PI14/00613 to V. Moreno, PI17/00878 to S. Castellví-Bel, PI17/00092 to V. Moreno, PI20/00113 to S. Castellví-Bel), Fundació La Marató de TV3 (2019–202008 to S. Castellví-Bel), Fundación Científica de la Asociación Española contra el Cáncer (GCB13131592CAST to A. Castells, GCTRA18022MORE to V. Moreno, PRYGN211085CAST to S. Castellví-Bel), CERCA Program (Generalitat de Catalunya CERCA-2648; to all authors) and Agència de Gestió d'Ajuts Universitaris i de Recerca (Generalitat de Catalunya, GRPRE 2017SGR21 to S. Castellví-Bel, GRC 2017SGR653 to A. Castells; GRC 2017SGR735 to J.M. Borràs, GRC 2017SGR80 to X. Bessa, GRC 2017SGR723 to V. Moreno). Centro de Investigación Biomédica en Red de Enfermedades Hepáticas y Digestivas (CIBEREHD) and Consortium for Biomedical Research in Epidemiology and Public Health (CIBERESP) are funded by the Instituto de Salud Carlos III (to all authors). C. Aranau-Collell and J. Muñoz were supported by a contract from CIBEREHD (CB06/04/0016). A. Díez-Villanueva and L. Bonjoch were supported by PERIS contract (SLT017/20/000042) and a Juan de la Cierva postdoctoral contract (FJCI-2017–32593), respectively. This article is based upon work from European Cooperation in Science and Technology (COST) Action CA17118 to all authors, supported by COST (www.cost.eu). We are also sincerely grateful to the participants, and Biobanks of Hospital Clínic–IDIBAPS, IDIBELL, and Hospital del Mar (Barcelona, Spain). The work was carried out (in part) at the Esther Koplowitz Centre.

We acknowledge the contribution of the CRIPREV consortium, which permitted to develop this study. The members are listed as follows:

Fundació Institut d´Investigació Biomèdica de Bellvitge (IDIBELL; Barcelona, Spain) – ICO: J.M. Borràs, E. Guinó, G. Ibañez-Sanz, M. Obon-Santacana, F. Moratalla-Navarro, A. Diez-Villanueva, R. Sanz-Pamplona, and V. Moreno.

Fundació Clínic per a la Recerca Biomèdica (IRS-IDIBAPS; Barcelona, Spain): C. Arnau-Collell, J. Muñoz, J.M. Augé, L. Bonjoch, A. Serradesanferm, À. Pozo, L. Moreira, Marcos Díaz-Gay, Sebastià Franch-Expósito, Cristina Herrera-Pariente, Yasmin Soares de Lima, Lorena Moreno, Teresa Ocaña, Sabela Carballal, Ariadna Sánchez, Francesc Balaguer, J. Grau, A. Castells, S. Castellví-Bel, Elena Asensio, Sara Lahoz, Carolina Parra, Clàudia Galofré, Iván Archilla, Miriam Cuatrecasas, and Jordi Camps.

Institut Hospital del Mar d'Investigacions Mèdiques (IMIM; Barcelona, Spain): Joan Gibert, Raquel Longaron, Clara Montagut, X. Bessa, B. Bellosillo, C. Márquez Márquez, R. Rueda Miret, R. Pérez Berbegal, G. Piquer Velaso, Joan Carles Balboa, Ana Cristina Alvarez Urturi, Ines Ana Ibañez Zafon, Sandra Cordero Cerrudo, Miriam Parrilla Carrasco, and Bouchra Alouali Moussakhkhar.

Institut de Recerca Biomèdica de Barcelona (IRB; Barcelona, Spain): Toni Gabaldón, Ester Saus, and Olfat Khannous-Lleiffe.

Fundació Institut d'Investigació en Ciències de la Salut Germans Trias i Pujol (Barcelona, Spain): Sergio Alonso, Beatriz González, Maria Navarro-Jiménez, Andreu Alibés, Mar Muñoz, Berta Martin, and Miguel A. Peinado.

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.