PSA is a protein secreted specifically from the prostate gland, and has been widely accepted as a serum biomarker for PC. However, other medical conditions, such as benign prostatic hypertrophy and inflammation can cause serum PSA elevation 
. Hence, the diagnostic specificity of PSA is quite low, especially at boarder-line levels of PSA, or ‘gray-zone’. Patients suspected to have PC by PSA screening usually undergo prostate needle biopsy, which is an invasive procedure that accompany complications, some of which are severe. In addition, recent randomized controlled trials have shown no or little benefit of PSA screening in extending cancer-specific survival 
. Economic burden of prostate needle biopsies, followed by overdiagnosis and overtreatment for PC, is another serious issue since it is estimated that each year, more than one million patients undergo prostate needle biopsies in the US, a procedure which costs $500–1,000 for each 
. Therefore, there is a world-wide controversy over PSA screening, and additional biomarkers which can better identify the patients that need prostate needle biopsies are definitely required.
The risk of PC is 2.5 times higher in the patients with a positive family history of PC in their first degree relatives 
. However, only a small proportion of patients have known positive family history of PC in Asians including the Japanese 
, and collection of detailed family history is often problematic. In our GWAS of the Japanese which includes 4,584 PC, only 6.6% of the patients had positive family history of PC, and data was missing in 15.5% of the patients. Contrarily, risk prediction models based on easily accessible genetic information can be applied to the general population. Furthermore, unlike other biomarkers which show some degree of fluctuation that could affect their reproducibility, genetic risk scores are stable in each individual. So far, few studies have studied the reproducibility of genetic risk prediction models using independent sets of samples. Although reproducibility of genetic risk prediction models could be affected by sample collection bias, using independent sets of samples we have shown that our model is highly reproducible in a single ethnic group. In fact, our model showed similar reproducibility when it was applied to a sample set of Japanese from the Multiethnic Cohort in Hawaii and California (MEC
). These data warrant similar predictive performance of our model in other Japanese population or East Asian population as well.
Genetic risk prediction of PC was first reported using only five common susceptibility variants 
. The model was established by simply counting the number of risk alleles. Subsequently, models incorporated increasing number of variants, and logistic regression models were adapted to account for the effect size of each variant 
. In the present study, we have also created a model based on step-wise logistic regression analysis. When compared to the model incorporating 9 of the 16 variants that remained significant in a step-wise model, the predictive performance of the model including all the 16 variants was modestly but statistically significantly superior (Figure S4
). Over forty PC susceptibility variants have been reported so far, and still more remains to be identified. Inclusion of novel variants newly identified to be associated with PC is expected to further improve the predictive performance of the model, although careful selection of the variants based on association studies in each ancestry group is necessary. Of note, even when we have included only 16 PC susceptibility variants identified to be associated with PC in GWAS of Japanese, one variant showed negative association with PC in the samples used to create the risk prediction model, suggesting that a variant with very mild effect would show either positive or negative association in different sample sets even in a single ancestry group, which may lower the predictive performance of the model.
There is still a large debate over the clinical utility of genetic risk prediction models. The overall predictive performances of genetic risk prediction models as assessed by ROC analysis are usually modest, since the distribution of the ORs between the case and controls largely overlap. However, it has been implicated in breast cancer that genetic risk prediction models could be clinically useful among a subset of high risk patients 
. In case of PC, patients can be risk-stratified using PSA, and genetic risk prediction models can be a useful compensatory marker at gray-zone PSA, where patients have relatively high risk of PC, and the diagnostic ability of PSA is the lowest. Furthermore, PCs are generally slow growing, and even if patients with PC are false negatively classified as low risk by a genetic prediction model, they can still be followed with serial PSA measurements, and can have prostate biopsy with increasing PSA before reaching advanced stages except in rare cases of very aggressive tumor. Identification of aggressive PCs is another important issue in PC diagnosis. Most of the PC susceptibility variants identified by GWAS have fallen short of discriminating aggressive from non-aggressive PCs 
, and there was no significant difference in the distribution of ORs between the aggressive and non-aggressive PCs in our genetic risk prediction model as well (Figure S5
). Additional biomarkers that can discriminate aggressive and indolent PCs should be explored.
The circumstance where biomarkers are applied is also important in assessing their clinical values. When the prevalence of a disease is very low, even a marker with very high sensitivity and specificity shows low positive predictive value. Since the overall predictive value of genetic risk prediction models are low, it is important to utilize the model in a situation where disease prevalence or pre-test probability is relatively high. We have demonstrated that the predictive performance of the genetic risk prediction model is not affected by serum PSA level, and that the model can be utilized in the patients with gray-zone PSA, where there is usually 20–25% probability of a positive prostate needle biopsy. Recently, a few common variants have been reported to be associated with serum PSA level 
, and two of them, rs10993994 at 10q11
and rs7501939 at 17q12
were included in our genetic risk prediction model. It is presumed that the small effect size of these variants on PSA level did not significantly affect overall predictive performance of our risk prediction model at gray-zone PSA. Some other studies that have evaluated a model incorporating PSA and genetics have argued that there was only minor increase in AUC when genetic information was added 
. However, PSA fluctuation is problematic in PSA gray-zone, and it is possible that the true effect of combining PSA with genetic risk information is confounded by these variances in PSA.
We have shown that while the genetic risk model may not be helpful clinically in all the patients with gray-zone PSA, it may largely influence decision making in a portion of patients. In our clinical simulation, 24.2% of the patients had OR<0.5, and these patients had 10.7% chance of being positive after prostate needle biopsy. Considering the complications of prostate needle biopsies, these patients might chose serial PSA follow-up rather than immediate prostate needle biopsy. On the other hand, 9.7% of the patients with OR>2, who have more than 42.4% chance of being positive for prostate cancer, may choose to undergo immediate prostate needle biopsy. Although the genetic risk prediction model should further be evaluated prospectively in clinics, our data suggests that it can be an additional biomarker that can risk-stratify individuals at gray-zone PSA in Japanese, leading to personalized medicine.