Do Polygenic Risk Scores Add to Clinical Data in Predicting Pancreatic Cancer? A Scoping Review

The majority of pancreatic cancer cases are discovered at an advanced, unresectable stage, with poor 5-year survival [Surveillance Epidemiology and End Results (SEER; RRID:SCR_006902)]. If we could identify pancreatic cancer earlier when it remains resectable, we could improve survival fourfold (1). Because universal pancreatic cancer screening is not feasible (2), cost-effective methods to identify high-risk individuals to undergo targeted screening and surveillance would significantly improve early detection.

Susceptibility to pancreatic cancer is likely determined by many routinely available clinical factors and some additional genetic factors (3–7). Routine clinical predictors of pancreatic cancer include alcohol use, diabetes, cigarette smoking, and obesity (3–5, 7–12). Genetic factors have been most commonly evaluated as polygenic risk scores (PRS; ref. 13), which sum an individual's alleles associated with pancreatic cancer risk across the genome based on genome-wide association studies (GWAS; refs. 14–23).

While PRS have been developed for pancreatic cancer (13, 24), their additional clinical utility and discrimination beyond prediction models based on routine clinical factors is unknown. We aimed (i) to identify all published and preprint manuscripts that evaluate a pancreatic cancer PRS and pancreatic cancer risk and (ii) to evaluate the independent prognostic discrimination of pancreatic cancer PRS beyond clinical risk.

This scoping review was conducted and reported in accordance with the PRISMA Extension for Scoping Reviews (25) and Levac and colleagues’ recommendations for scoping review methodology (26) with the PRISMA Checklist (Supplementary Table S1). The study protocol is available on Open Science Framework (https://osf.io/97hxw).

A medical librarian (A.A. Grimshaw) conducted an exhaustive search of the literature in Cochrane Library, Google Scholar, Ovid Embase, Ovid MEDLINE, PGS Catalog, PubMed, Scopus, and Web of Science Core Collection databases to find relevant articles published from each database inception to March 10, 2023. Search terms (Supplementary Table S2) were a combination of keywords and controlled vocabulary for PRS and pancreatic cancer and not limited by language, publication type, or year. A second medical librarian peer reviewed the search by using the Peer Review of Electronic Search Strategies. Using Citationchaser (27), we searched the reference lists of included studies and retrieved articles that cited the included studies with the goal of finding additional relevant studies not retrieved by the initial database search.

We imported citations from all databases into Endnote 20, removing duplicates using the Yale Reference Deduplicator (28). The deduplicated results were imported into the Covidence systematic review management program for screening (29). Two independent screeners performed a title and abstract review. Three separate individuals (L. Wang, C. Mezzacappa, and N.R. Larki) performed full text review and resolved disagreements by consensus. Criteria for inclusion in the study included original manuscripts and preprint manuscripts with the inclusion of a pancreatic cancer PRS and testing of the association between the PRS with either pancreatic cancer or specifically pancreatic ductal adenocarcinoma.

We extracted information regarding author, journal, year of publication, study design, source/country of origin of the study, aim/purpose, funding source, description of the source population, population size, ancestry, definition of pancreatic cancer cases and controls (if applicable), follow-up time, genetic variants included in the PRS, methods for analysis, model inputs (use of age/sex, inclusion of clinical risk factors), and evaluation of the discriminatory ability of models with and without PRS. Sudlow and colleagues (30) provided details on ascertainment of ancestry in the UK Biobank. We evaluated whether studies included confidence intervals (CI) surrounding the ΔAUC or P value for the ΔAUC to determine if there was a statistically significant improvement in pancreatic cancer prediction after inclusion of the PRS. In addition, following TRIPOD guidelines for risk prediction models, we also reported on the clinical usefulness of the PRS addition using measures such as net reclassification improvement (NRI), which sums the reclassification of individuals with and without pancreatic cancer into high- and low-risk categories as defined by a prespecified probability cutoff.

The authors confirm that the data supporting the findings of the study are available within the article and supplementary materials.

The search resulted in 828 records (Fig. 1); after duplicates were removed, a total of 489 citations were identified from searches of electronic databases. On the basis of the title and abstract, 450 citations were excluded, with 39 full text articles retrieved and assessed for eligibility. Of these, 20 citations were excluded for the following reasons: duplicate citation, duplicate study data (the exact same study data used between two studies), conference abstract, pancreatic cancer PRS association with pancreatic cancer was not evaluated, no original data, no pancreatic cancer PRS, and thesis paper (Supplementary Table S3). Two additional articles were located from citation chasing, as described previously in the methods. The remaining 21 studies were considered eligible for this review (19, 21, 22, 24, 31–47).

Studies were published between 2012 and 2022, and most included populations from the United Kingdom (n = 10) and United States (n = 10), with fewer than three studies from Brazil, China, and non-UK European countries (Germany, Italy, Poland, Greece, Czech Republic). Determination of ancestry varied among studies and included genetically derived ancestry, proxies of self-reported race/ethnicity, country of origin, or a combination of the first two methods. On the basis of the study's definition of ancestry, thirteen studies focused exclusively on European ancestries, and six studies included individuals of non-European ancestries. Most studies employed data from the UK Biobank (n = 9) or pancreatic cancer consortia (n = 6). The remainder were institutional biobanks or hospital-based studies. Studies used a case–control study design (n = 16) or a cohort design (n = 5). Less than half the included cases supplied directly from cancer consortia or included pathologically defined pancreatic cancer (n = 10). Of note, 8 studies had as the primary aim developing a risk prediction model for pancreatic cancer, 10 studies incorporated the use of PRS for evaluating multiple cancers, 2 studies were primarily GWAS for pancreatic cancer, and 1 primarily evaluated PRS in the context of methods for improving PRS performance. Of note, 14 studies used effect sizes from the original GWAS or meta-analysis, which can result in overestimation of the effect sizes of newly discovered genome-wide significant SNPs (48). Two studies used effect sizes from a replication GWAS, and it was unclear in 5 studies where effect sizes were obtained. Additional details on the studies’ first author, journal, year, study design, population source and description, aim/purpose, sample size, determination of ancestry, case and control definitions, variants used for the PRS, choice of analysis, and funding source are in Table 1 and Supplementary Table S4.

We also compared the controls used in case–control and cohort studies. The median age of onset of pancreatic cancer is 70 years old (SEER; RRID:SCR_006902), and the known risk factors (49) of pancreatic cancer include diabetes (3, 5), cigarette smoking (9, 10), alcohol use (11, 12), pancreatitis (50–52), pancreatic cysts (53), and increased body mass index (BMI; refs. 54, 55). Many of the control groups were younger and healthier than those typically at risk for pancreatic cancer or drew from studies without relevant clinical exposures for pancreatic diseases. By providing skewed exposure levels of nongenetic risk factors in the controls, individuals with high genetic liability (e.g., pancreatic cancer PRS) would not have had the opportunity to develop pancreatic cancer, limiting the interpretability of the pancreatic cancer PRS on risk for pancreatic cancer. Specifically, the average age of UK Biobank individuals was 56.5 years old, and other studies drew controls from healthy blood donors, patients undergoing surgery, or healthy community members (46). One study specifically excluded individuals with clinically relevant risk factors for pancreatic cancer including pancreatitis and pancreatic cysts (24). In other cases, studies drew cases and controls from separate populations, such as different hospitals in various regions of the nation with varying levels of care complexity (31), or hospitals versus community centers (46).

While nearly all studies included age and sex in their models (n = 18), only seven studies adjusted for routine clinical risk factors (Table 2). These factors included smoking (22, 24, 37, 40, 41), family history of any cancer or specifically pancreatic cancer (22, 37, 42), and BMI or waist circumference (37, 41, 42), diabetes (24, 42), alcohol use (40, 41), and Townsend index (33). Of these, 5 that included clinical risk factors with a PRS also evaluated model discrimination and calibration, mostly using AUC. Only 3 studies (refs. 37, 41, 42; 14.3%) evaluated the change in discrimination with the addition of PRS to clinical factors versus clinical factors alone and reported improvements in discrimination [(i) AUC: 0.715 to 0.745 (ΔAUC: 0.030), no 95% CI or P value provided; (ii) AUC 0.791 to 0.830 (ΔAUC: 0.039) P value = 0.0002; (iii) analysis in two separate datasets: UK Biobank, AUC from 0.694 to 0.711 (ΔAUC: 0.017) and Michigan Genomics Initiative (MGI): AUC from 0.609 to 0.619 (ΔAUC: 0.01); no 95% CI or P value provided for either model].

Kachuri and colleagues (37) developed a cohort study among individuals with self-reported European ancestry in the UK Biobank (413,753 total individuals, including 493 pancreatic cancer cases) and evaluated the additive predictive value of a PRS for 16 separate cancers, including pancreatic cancer. This study used 22 SNPs from the combination of previously reported GWAS (15, 18–20, 23, 49, 56) that were identified in populations of European ancestry. Follow-up time included the date of enrollment to the date of cancer diagnosis (based on ICD codes), death, or end of follow-up (January 1, 2015), whichever came first. Clinical factors included BMI, smoking status, cigarette pack-years, family history of cancer (prostate, breast, lung or colon/rectum). Using Cox proportional hazards regression, the authors found that the addition of PRS to the model with clinical factors improved the AUC from 0.715 to 0.745 (ΔAUC: 0.030), no 95% CI or P value given. To assess clinical usefulness, the authors evaluated NRI, which we previously defined in Materials and Methods. They found the total percentile NRI index for the PRS addition to be 0.228 [95% CI, 0.174–0.284; NRI_event 0.228 (95% CI, 0.173–0.284); NRI_{no event} 0.001(−0.002 to 0.003)], however the set threshold for high versus low risk of pancreatic cancer was not reported.

Sharma and colleagues (42) performed a case–control study among individuals of European ancestry in the UK Biobank to test the ability of PRS to discriminate between individuals with and without pancreatic cancer (1,042 cases and 10,420 controls). This study used 5 separate PRS of varying SNP sizes (5–49 SNPs), derived exclusively from individuals of Japanese ancestry (22), individuals of European ancestry (24, 36, 57), or a combination of both. Incident cases were based on ICD codes and self-reported cases of pancreatic cancer, and controls were age and sex matched participants without any history of pancreatic cancer. Clinical risk factors included age at diabetes diagnosis, diabetes onset, waist circumference, and family history of pancreatic cancer. Using the Cox proportional hazards regression, the AUC improved from 0.791 to 0.830 (ΔAUC: 0.039; P value = 0.0002). The authors did not include further measures for evaluating the clinical utility of the model.

Salvatore and colleagues (41) used the data from the UK Biobank and MGI to evaluate the combination of phenotype risk scores, PRS and clinical risk models for pancreatic cancer and assess their discriminatory ability and calibration. This study used 18 independent, previously reported SNPs from 5 previous studies (19–21, 23, 58). In this case/control study, incident cases (1,088 individuals) were based on diagnostic ICD codes and controls (430,570 individuals) were based on age and sex-matched individuals without pancreatic cancer. Clinical risk factors included BMI, alcohol, and smoking. Separate analyses for each dataset through logistic regression showed ΔAUC improvement (UK Biobank, AUC from 0.694 to 0.711, ΔAUC: 0.017; MGI, best AUC improved from 0.609 to 0.619, ΔAUC: 0.01; no 95% CI or P value provided). The authors did not further evaluate the clinical utility of the model.

In this scoping review, we comprehensively identified published and preprint studies evaluating a pancreatic cancer PRS associated with pancreatic cancer published between 2010 to 2023 and the prognostic discrimination of PRS beyond established clinical risk factors. We found some suggestion of independent improvement in discrimination, based on very limited research. We have identified the current limitations and highlighted the need for additional studies on the independent prognostic discrimination of pancreatic cancer PRS in at-risk populations to help ensure the clinical utility of future studies.

Only one third of studies (33.3%) included routine clinical factors beyond age and sex and less than one tenth (9.5%) reported a significant improvement with PRS to routine clinical factors. Currently, only three pancreatic cancer–specific PRSs have demonstrated improved discrimination beyond existing clinical factors. In addition, existing studies are prone to confounding from shared location or relatedness of individuals, stand-ins for environmental or shared lifestyle risk factors (e.g., diet, exercise) respectively (13). Without accounting for clinical risk factors, models likely inflate the clinical utility of the PRS in real-world settings.

Compared with the actual at-risk pancreatic cancer population, many of the control groups or study cohort populations were either younger/healthier or drew from populations excluding relevant exposures of pancreatic diseases, such as pancreatitis and pancreatic cysts. By unnaturally skewing the exposures of nongenetic risk factors among cases versus controls or study cohort versus at-risk population, these studies artificially created an environment where individuals with high genetic liability (e.g., pancreatic cancer PRS) did not have the opportunity to develop pancreatic cancer (e.g., younger, healthier, or without risk factors), limiting the interpretability of the PRS risk and biasing the model including pancreatic cancer PRS towards the null. Other studies drew cases and controls from separate source populations, skewing the comparison of these groups. In addition, most of the studies only included individuals of European ancestry, which could exacerbate disparities given the limited predictive capability in non-European ancestry individuals (59).

Three studies reported an improvement in the AUC after the addition of PRS to clinical factors, and one study reported a significant P value for the ΔAUC, while the other two did not report testing for significance. Because AUC can also be clinically difficult to interpret, TRIPOD guidelines also suggest testing for clinical usefulness, and one study showed an improvement in the NRI with the addition of PRS to clinical factors. However, the value for NRI is heavily dependent on the set threshold for high vs. low risk of pancreatic cancer, which was not reported by the authors.

Our study had several strengths. First, this is the first comprehensive scoping review of PRS for pancreatic cancer association. We incorporated both published and unpublished literature and reported on a wide range of study designs and methodologies. In addition, to ascertain the clinical utility of these models, we reported on the added discriminatory ability of PRS to models containing clinical risk factors to understand the improvement in prediction for PRS beyond established risk factors.

Our scoping review has some limitations. Given the limited subject matter in this area, we included manuscripts with a broad scope of aims provided they also addressed our study aim. However, evaluating the discriminatory ability of pancreatic cancer PRS may have only been a secondary aim of the original manuscript.

This is the first scoping review to comprehensively identify all studies that evaluate a pancreatic cancer PRS associated with pancreatic cancer and to understand the independent contribution of PRS beyond clinical risk factors alone. However, out of 21 studies, only seven studies accounted for well-established clinical risk factors, and three studies evaluated the marginal prognostic discrimination of PRS in addition to clinical factors and all three showed an improvement in model discrimination. Only one study addressed the clinical usefulness of the PRS addition. Neither the ability for PRS to improve clinical risk prediction for pancreatic cancer nor its effect on clinical decision making has been adequately addressed. Additional investigations are needed to evaluate the prognostic discrimination and clinical utility of models with PRS independent of established clinical risk factors in diverse ancestral populations representative of those at risk for pancreatic cancer.

No disclosures were reported.

L. Wang: Conceptualization, data curation, formal analysis, methodology, writing–original draft, writing–review and editing. A.A. Grimshaw: Data curation, writing–review and editing. C. Mezzacappa: Data curation, writing–review and editing. N.R. Larki: Data curation, writing–review and editing. Y.X. Yang: Conceptualization, methodology, supervision, writing–review and editing. A.C. Justice: Conceptualization, methodology, supervision, writing–review and editing.

This paper was supported by a VA CDA-2 grant from the VA Research and Development office (Grant: 1IK2BX005891–01) awarded to L. Wang. In addition, Catherine Mezzacappa was supported by the NIH T32 grant: NIH T32DK007356.