Prediction of clinically significant prostate cancer using polygenic risk models in Asians

The incidence of prostate cancer (PCa) is constantly on the rise, currently second in cancer diagnosis and sixth in cancer mortality in men worldwide as of 2020 [1]. This trend is more apparent in Asia where PCa has traditionally had low prevalence, with literature pointing to the gradual transition to a westernized high-fat intake diet and obesity increase as well as broader implementation of nationwide PCa screening as notable factors [2]. As such, genetic origins of PCa pathogenesis have been of keen clinical interest, and with the approval of PARP (poly [adenosine diphosphate-ribose] polymerase) inhibitors as novel therapy for actionable genetic mutations including BRCA 1/2, the National Comprehensive Cancer Network (NCCN) now recommends germline testing for PCa not only in men with a positive family history but also in high- to very-high risk regional or metastatic PCa. Alterations in homologous DNA repair and mismatch repair genes are strongly linked to 3- to 6-fold increase in overall risk [3].

However, not all PCa can be explained with nominal genes of high penetrance, but rather can be attributed to the cumulative effect of small variant alleles that confer varying levels of individual risk to disease [3, 4]. With the pivotal study that identified 5 singular single nucleotide polymorphisms (SNPs) having strong association with metastatic PCa [5], the search for PCa susceptible variants have culminated in over 100 SNPs found in large-scale genome-wide association studies (GWASs), attributing to 60% of heritability [6]. Associations from germline mutations provide valuable information in predicting lifetime trajectory, allowing individualized medical care at an early stage of screening and diagnosis. Polygenic risk score (PRS) takes advantage of common polymorphisms to assess a person’s genetic risk of disease, and when used with other clinical variables, has added value in anticipating tumor aggressiveness and escaping unnecessary biopsy [7, 8].

Current literature on PRS in PCa are scarce compared to the vast number of reported SNPs with variable levels of performance, limiting application in clinical practice. Most large scale models are also constructed in males of mainly Caucasian and European ancestry, resulting in lower predictive power when replicated in Asian cohorts likely due to non-shared genomic loci [4]. As such, ethnic-specific PRS is required to account for genetic variations. In this study, we developed and validated a PRS model predicting clinically significant PCa (csPCa) in a Korean population.

The mean age of PCa cases and controls were 67.8 and 64.7 years, respectively. The mean BMI of cases and controls was 24.6 kg/m² and 24.1 kg/m², respectively. The mean prostate-specific antigen (PSA) level of PCa cases were 35.1 ng/mL with standard deviation of 209.4 ng/mL. Most of PCa cases had GS 7 (n=2,105), followed by GS 9 or higher (n=497) and GS 8 (n=411) (Table 1).

Table 1
Characteristics of the study population

• Click for larger image
• Click for full table
• Download as Excel file

We identified 83 SNPs that were previously reported to be PCa-associated in one or more studies (Table 2). rs72725879 located within PRNCR1 was most significantly associated with PCa risk (OR, 0.597; 95% confidence interval [CI], 0.555–0.641; p=4.3×10^-45). Other 38 SNPs also showed negative associations between minor allele and PCa risk. Of the 83 SNPs, 23 SNPs were located within 8q24.21 region, spanning PRNCR1, CASC8, PCAT2, PCAT1, LOC105375751, and CCAT2, 8 SNPs were located within HNF1B (17q12) and 5 SNPs were located within 11q13.3. After LD clumping, 31 SNPs located within 26 distinct genomic loci had been extracted as candidate SNPs for PRS construction (Table 3).

Table 2
SNPs associated with development of PCa

• Click for larger image
• Click for full table
• Download as Excel file

Table 3
Candidate SNPs for PRS construction

• Click for larger image
• Click for full table
• Download as Excel file

We calculated individual PRS in the test set (311 cases and 822 controls) and tested the performance in predicting PCa risk (Fig. 1). PRS models composed of top 29 SNPs, top 26 SNPs and top 28 SNPs showed comparably superior performance compared to other models (Table 4). The AUC of the prediction model of PRS composed of 29 SNPs was estimated to be 0.700 (95% CI, 0.667–0.734), with a sensitivity and specificity of 0.672 and 0.662, respectively (Fig. 2A).

Fig. 1
Distribution of PRS in cases and controls. Mean PRS in controls: pink dashed line, Mean PRS in cases: blue dashed line, and optimal PRS cutoff: black solid line. PRS, polygenic risk score.

• Click for larger image
• Download as PowerPoint slide

Fig. 2
(A) ROC curve of PRS models in prediction of PCa risk. (B) Distribution of PRS in cases and controls with two cutoff values to categorize PRS risk groups into moderate, high and very-high risk group. (C) Risk of development of PCa by PRS risk group. AUC, area under the receiver operating characteristics curve; OR, odds ratio; ROC, receiver operating characteristic; PRS, polygenic risk score; PCa, prostate cancer.

• Click for larger image
• Download as PowerPoint slide

Table 4
Predictive performance of PRS models for development of PCa

• Click for larger image
• Click for full table
• Download as Excel file

From the best performing PRS model composed of top 29 SNPs, we chose two different cutoff values, i) the optimal PRS value at the maximum Youden’s index (-0.05) and ii) 95th percentile PRS cutoff (0.02) (Fig. 2B). We defined those with PRS below the optimal cutoff, those with PRS above 95th percentile cutoff and those with PRS between the two cutoff values as moderate, very-high and high-risk group, respectively. High PRS and very-high PRS group showed 3.80-fold and 5.93-fold increased risk, respectively, for PCa, compared to the moderate PRS group (the reference group) (Fig. 2C). The very-high PRS group had a 4.92-fold increased risk compared to the remaining population.

Based on the OR estimates of each SNP in PCa specific to a Korean population, we developed and evaluated PRS in Korean males. Of the 83 SNPs identified with significant association, 31 SNPs were selected for PRS construction. The best performing PRS composed of 29 SNPs reported an AUC of 0.700, with the top 5th percentile having more than a 4-fold increased risk of Grade Group (GG) ≥2 PCa. To note, minor alleles of many SNPs included in this study showed protective effects on PCa risk, which highlights that common risk alleles may have a cumulative impact in the development of PCa.

With trends of increasing prevalence worldwide, familial PCa and heritability is thoroughly researched for potential identification of actionable mutations. Hereditary PCa is caused by inherited genes of high penetrance, including homologous DNA pair genes (e.g., BRCA1/2, ATM, CHEK2, PALB2) and mismatch repair genes (e.g., MLH1, MSH2, MSH6, PMS2), the latter related to Lynch syndrome. While such nominal mutations are easily identified and linked to PCa risk, not all familial PCa can be explained by singular mutations, but rather is accountable by the sum of multiple common variants with small effect size. PRS is a novel and only biomarker that uses the concept of polygenic risk as a single combinatory value estimate of the individual’s genetic propensity of a complex disease to potentially stratify high- and low-risk patients in order to apply different methods of screening and prevention.

While the exact number required to achieve the full potential of PRS is unknown, summary statistics based on sufficient data is required to detect rare but significant variants shared between phenotypes and approximate the populational burden of disease. PCa is reported to have variants with relatively larger effect size and hence has been suggested to require a smaller sample to assess heritability compared to other, more polygenic cancers. While falling short of over 60,000 cases required to explain 80% of heritability as suggested by Zhang et al. (2020) [14], one of the largest cohorts of over 2,700 males in biopsy and RP-proven GS ≥7 PCa was utilized, compared to previous Korean [15] and Japanese [16] PRSs developed from 1,001 and 689 cases, respectively.

Selection of appropriate SNPs for inclusion critically influences PRS performance, due to obvious uncertainties inherent in estimating effect size from a potentially biased population. As such, statistical techniques must be applied prior to PRS development, a process necessary to weed out insignificant variants and to avoid overfitting or under-fitting of risk alleles. p-value thresholds are often implemented as a cut-off, and while p<1×10^-5 is most commonly used, it is often an arbitrary value. A strict, parsimonious threshold risks the loss of otherwise powerful variants confounded by nearby SNPs resulting in false negatives, whereas an overly flexible threshold risks inclusion less predictive alleles as well as a resulting SNP combination too large to be feasible in actual practice. Hence, to optimize SNP selection, a relative lenient p-value threshold of p<10^-3 was applied to filter a relatively large set of potential candidates, followed by LD clumping to extract lead SNPs and quality control via imputation for missing genotypes as well as SNP combination comparisons, a similar approach described in our previous report [15].

To additionally bolster performance, our study selectively used previously reported, well-established PCa-susceptible SNPs, many found in the 8q24.21 region. 8q24 is commonly considered a “desert” region due to scarcity of genes, but has been the source of numerous PCa susceptible variants, contributing to 8% to 9% of two-fold increase in familial risk [17]. A proto-oncogene Myc is closest, and while polymorphisms in this region have been associated with regulatory enhancers and androgen receptor (AR) response, mechanisms of action are not yet well understood. Nonetheless, strong associations have been made in both European and East Asian populations [18, 19], with rs7837688 and rs1016343 more common in Korean males [15], a finding comparable in this study as 4th and 12th in statistical significance. Eleven candidate SNPs were associated with PRNCR1 in the 8q24 region, whose upregulation is implicated in aggressive PCa via AR-mediated transcriptional programs [20]. Thirteen were located in HNF1B (17q12) and 11q13.3, both well-documented as susceptible loci in males of European and Asian descent [21, 22].

The ultimate goal of this research was to evaluate whether PRS can be utilized in actual clinical practice to avoid unnecessary intervention in low-risk PCa. Therefore, we selected patients with biopsy or pathology-proven csPCa with ≥GG2, excluding males under the age of 60. While generally preferred in the Korean population over active surveillance (AS) or watchful waiting, early intervention and surgery in low-grade PCa risks overtreatment and economic burden. As such, imaging modalities and biomarkers have been introduced to focus biopsy and treatment in csPCa, which generally includes any of histopathology ISUP grade ≥2 (≥GG2) and/or tumor volume ≥0.5 cc [23]. Our study, selecting only males of intermediate- to high-risk PCa category, showed performance comparable to previous literature. Seibert et al. (2018) [7], defining aggressive PCa as any cancer not eligible for surveillance based on the NCCN guidelines (i.e., any ≥GG2, stage T3–4, PSA ≥10 ng/mL, and any nodal or distant metastasis), developed a PRS from 54 SNPs to predict early age at onset of aggressive PCa (z=11.2, p<10^-16), with males in the 98th percentile having a 2.9-fold increased risk. Stockholm3 combined a PRS generated from 232 SNPs with 6 plasma proteins and other clinical variables to predict ≥GG2 PCa, outperforming PSA alone with a resulting AUC of 0.74 (95% CI, 0.72–0.75) [8]. Takata et al. (2019) [24] published results from a large Japanese cohort with a development set of 5,088 PCa cases from Biobank Japan and 10,682 controls from multiple institutions, but 1,806 (35.5%) patients had missing GS, and 1,293 (25.4%) had GS ≤6. Our effect size estimates calculated from csPCa better represent genetic polymorphisms that require intervention in actual practice.

Generalizability is key to the performance of a PRS, and population bias from a limited cohort poses a unique challenge in genomic prediction. Previously reported models are generated from cohorts of largely European ancestry and hence often fail to have the same predictive power in Asian males, primarily because unshared genomic loci, as well as variable population-specific effect size. A meta-analysis of multiancestry GWAS data found East Asians to have a 0.73-fold lower risk score compared to their European counterparts, highlighting the probability of potent germline variations among ethnic populations [25]. As such, accurate prediction of PCa risk requires tailoring to each race and ethnicity. Also, with increasing prevalence, even a modest performance is able to substantially stratify absolute populational risk, allowing application into clinical practice of early detection and prevention. AUC of 0.700 in our study is one of the highest discriminatory performances of PRS developed in an Asian cohort to date [4], compared to a previous report by our team conferring an AUC of 0.605 for any PCa utilizing 5 SNPs [15] and those developed in Japanese and Chinese populations with AUCs of 0.60 and 0.659, respectively [16, 26]. The 4.92-fold increased risk of csPCa in the 95th percentile of our study is comparable to results validated in western counterparts with a 4 to 5-fold increased risk in the 99th percentile [18, 27].

Despite successful replication of significant PCa susceptible polymorphisms and outperforming some previous models [15, 16, 28, 29], our study is not without limitations. First, although we utilized data from 2,702 males with csPCa, it is still a relatively small sample size to accurately represent true populational variation. Second, because both development and validation cohorts consisted of Koreans, future evaluation in other race and ethnicities are required to fully assess predictive performance and generalizability. Due to the inherent selection bias in design, not all significant SNPs may be adequately represented in our data. In addition, while we limited the study population to csPCa, applying our model to males with any PCa at any age may further PRS utility, as both patients and clinicians can have additive information to guide screening schedules and opt for early intervention in AS-eligible males with high genetic risk.

Nonetheless, this study provides valuable evidence for prediction of csPCa based on the cumulative effects of small genetic variants specific to Asian populations. PRS calculated from individual GWAS data is the hallmark of modern precision medicine with the possibility of predicting a patient’s lifetime trajectory for disease, offering the chance for early screening and personalized treatment, as in the case of coronary heart disease [30] where genetics is making great strides in real-world practice. Future acquisition of more sample summary statistics will further potentiate PRS performance in PCa and produce more accurate stratification of risk groups.

We successfully developed and validated PRS models for csPCa risk developed in a large sample of 2,702 ≥GS7 PCa cases and 7,485 healthy controls. A model utilizing the top 29 SNPs conferred the best predictive performance with an AUC 0.700, with over 4-fold risk predicted in males in above the 95th percentile of PRS. Future prospective, large-scale studies are required to further validate our results.