Polygenic Risk Scores Stratify Keratinocyte Cancer Risk among Solid Organ Transplant Recipients with Chronic Immunosuppression in a High Ultraviolet Radiation Environment

and a,b Solid organ transplant recipients (SOTRs) have elevated risks for basal cell carcinoma (BCC) and squamous cell carcinoma (SCC), especially in high UVR environments. We assessed whether polygenic risk scores can improve the prediction of BCC and SCC risks and multiplicity over and above the traditional risk factors in SOTRs in a high UV setting. We built polygenic risk scores for BCC (n ¼ 594,881) and SCC (n ¼ 581,431) using UK Biobank and 23andMe datasets, validated them in the Australian QSkin Sun and Health Study cohort (n > 6,300), and applied them in SOTRs in the skin tumor in allograft recipients cohort from Queensland, Australia, a high UV environment. About half of the SOTRs with a high genetic risk developed BCC (absolute risk ¼ 45.45%, 95% conﬁdence interval ¼ 33.14 e 58.19%) and SCC (absolute risk ¼ 44.12%, 95% conﬁdence interval ¼ 32.08 e 56.68%). For both cancers, SOTRs in the top quintile were at > 3-fold increased risk relative to those in the bottom quintile. The respective polygenic risk scores improved risk predictions by 2% for BCC (area under the curve ¼ 0.77 vs. 0.75, P ¼ 0.0691) and SCC (area under the curve ¼ 0.84 vs. 0.82, P ¼ 0.0260), over and above the established risk factors, and 19.03% (for BCC) and 18.10% (for SCC) of the SOTRs were reclassiﬁed in a high/ medium/low risk scenario. The polygenic risk scores also added predictive accuracy for tumor multiplicity (BCC R 2 ¼ 0.21 vs. 0.19, P ¼ 3.2 (cid:2) 10 e 3 ; SCC R 2 ¼ 0.30 vs. 0.27 , P ¼ 4.6 (cid:2) 10 e 4 ).

To date, KC cancer prevention has relied on the assessment of traditional risk factors, but new approaches harnessing genetic information through polygenic risk scores (PRSs) have shown recently to have good potential for improving risk stratification. We have previously reported that in a low UV setting such as in the United Kingdom, (i) PRSs derived from the general population enable effective risk stratification for KC cancers among SOTRs; (ii) transplant recipients with a high genetic risk (PRS) have 3.3fold and 2.1-fold increased risk per 1 SD increase in BCC or SCC PRSs, respectively; and (iii) the PRS improves BCC predictions over and above the traditional risk factors with a 3% increase in the prediction accuracy (area under the curve [AUC]) (Seviiri et al., 2021). Other studies have also shown that PRSs generated from the nontransplant general population can predict the risk of BCC and SCC among SOTRs in low UV settings (Stapleton et al., 2020(Stapleton et al., , 2019. However, given that high UV exposure and chronic immunosuppression are strong risk factors for BCC and SCC, it remains to be determined whether the findings mentioned earlier apply to SOTRs with chronic immunosuppression in a high UV setting. Secondly, in high UV settings such as in Australia, where many in the population have pale skin, KC cancer incidence rates and tumor multiplicity are extremely high (Pandeya et al., 2017;Way et al., 2020). It is hence of interest to know whether a PRS can predict not only the risk but also tumor burden (multiplicity).
Therefore, this study aims to assess whether PRSs generated from the general population can improve BCC and SCC risk prediction over and above the traditional risk factors in SOTRs in a high UV index environment and whether it can predict multiplicity of KC cancer.

Performance of the PRSs prediction models in the independent QSkin Sun and Health Study validation cohort
The F2 model with a linkage disequilibrium (LD) radius of 5,000 kilobase (kb) and a fraction of causal SNPs of 0.01 was the best predictive model for BCC risk in the QSkin Sun and Health Study (Qskin) with a Nagelkerke's variance (R 2 ) of 33.7 % (Figure 1a). For SCC, the best predictive model was F3 with an LD radius of 5,000 kb and a casual fraction of SNPs of 0.001 in QSkin with Nagelkerke's R 2 of 35.5% (Figure 1b).

Baseline characteristics in the skin tumors in allograft recipients cohort
The analysis for BCC and SCC was restricted to 331 and 337 participants, respectively, who had complete data on all important variables. At baseline, participants had an average (SD) duration of immunosuppression of 9.61 (8.50) years, they reported a mean (SD) age of 44.4 (14.2) at the first transplantation, and the majority (217, 65.6%) were male. Further baseline characteristics   Figure 1. The performance of the BCC and SCC PRS prediction models in the QSkin validation cohort. (a) The performance of BCC PRS prediction models in the validation cohort. The x-axis represents the prediction models with different fractions of causal SNPs. Fi represents the infinitesimal model, whereas F0, F1, F2, F3, F4, and F5 represent fractions of causal SNPs of 1, 0.1, 0.01, 0.001, 0.0001, and 0.00002, respectively. The red and cyan colors represent the prediction models at the LD radius of 2,000 kb and 5,000 kb, respectively. The y-axis represents Nagelkerke's variance (R 2 ) (%) for each of the prediction models. The black dashed line highlights the best predictive model (with the highest Nagelkerke's R 2 ). (b) The performance of SCC PRS prediction models in the validation cohort. The x-axis represents the prediction models with different fractions of causal SNPs. Fi represents the infinitesimal model, whereas F0, F1, F2, F3, F4, and F5 represent fractions of causal SNPs of 1, 0.1, 0.01, 0.001,0.0001, and 0.00002, respectively. The red and cyan colors represent the prediction models at the LD radius of 2,000 kb and 5,000 kb, respectively. The y-axis represents Nagelkerke's R 2 (%) for each of the prediction models. The respective PRSs were associated with the risks of BCC (OR per SD ¼ 1.52, 95% CI ¼ 1.15e2.00, P ¼ 3.0 Â 10 e3 ) and SCC (OR per SD ¼ 1.69, 95% CI ¼1.25e2.28, P ¼ 7.2 Â 10 e4 ) after adjusting for the established risk factors and the first 10 principal components (PCs) (Figure 2).
PRSs and risk stratification for BCC and SCC among SOTRs in the skin tumors in allograft recipients cohort About half of the people with a high genetic risk (in the respective top PRS quintiles) developed BCC (AR ¼ 45.45%, 95% CI ¼ 33.1458.19%) and SCC (AR ¼ 44.12%, 95% CI ¼ 32.08e56.68%) during follow-up ( Figure 3a). Despite having a low genetic risk (bottom quintile), SOTRs in the skin tumors in allograft recipients (STARs) cohort had an AR for BCC 2.6 times higher than that in the QSkin validation cohort of 40,438 nontransplant recipients in the same high UV setting after about the same period of follow-up (AR in STAR¼ 25.37%, 95% CI ¼ 15.53e37.49% vs. AR in QSkin ¼ 9.57%, 95% CI ¼ 9.28e9.86%). Similarly, SOTRs in STAR in the bottom quintile of the PRS had an AR for SCC that was five times higher than that in the QSkin cohort after a similar duration of follow-up (AR in STAR ¼ 20.59%, 95% CI ¼ 11.74e32.12% vs. AR in QSkin ¼ 4.16%, 3.97e4.36% for QSkin) (Figure 3a).
Compared with the SOTRs with a low genetic risk (bottom quintile), SOTRs with a high genetic risk (top quintile) had a 3.5-fold increased risk of developing BCC (OR ¼ 3.66, 95% CI ¼ 1.54e8.72, P ¼ 3.3 Â 10 e3 ), whereas those with a moderate genetic risk (the middle 60%) had a 2.0-fold increased risk (OR ¼ 1.95, 95% CI ¼ 0.94e4.04, P ¼ 0.0716), after adjusting for the established risk factors and the first 10 PCs (Figure 3b). Similarly, SOTRs in the top quintile had a 3.2-fold increased risk (OR ¼ 3.21, 95% CI ¼ 1.27e8.17, P ¼ 0.0135) of developing SCC when compared with SOTRs in the bottom quintile ( Figure 3b). the R 2 to 0.21. The analysis of variance between the two models indicated that adding the PRS significantly improved the model fit (P-value for ANOVA test ¼ 3.2 Â 10 e3 ). Adding the SCC PRS improved the prediction of SCC multiplicity over and above the established risk factors (established risk factors þ PRS model R 2 ¼ 0.30 vs. established risk factors model R 2 ¼ 0.27, P-value for ANOVA test ¼ 4.6 Â 10 e4 ).

PRS models for UK Biobank D 23andMe versus UK Biobank only for BCC and SCC
Comparing the PRS derived from the UK Biobank (UKB) þ 23andMe versus that derived from only UKB, we found that the AUC was very slightly higher for SCC for the UKB þ 23andMe scenario but that the reverse was true for BCC (Supplementary Table S2). However, the results were so similar that we cannot conclude that the larger training dataset (UKB þ 23andMe) performed better than the smaller one (UKB only). In general, the larger sample size available from UKB þ 23andMe should enable better prediction than that available from one source alone, although in practice, the performance was similar, perhaps owing to the 23andMe phenotype being based on self-report.

DISCUSSION
This study evaluated whether a PRS generated from the general population can be used to stratify the risk of BCC and SCC among SOTRs with chronic immunosuppression in a high UV environment. We found that transplant recipients with a high genetic risk, that is, those in the top quintile, have a high risk of BCC and SCC, with about half of them developing BCC and SCC and having a 3-fold increased risk of BCC and SCC relative to those in the bottom quintile. Despite the strong environmental effect of UVR, the PRS improved both the BCC and SCC risk predictions over and above the established risk factors by 2% and 19.03%, respectively, and 18.10% of the SOTRs had their risk category changed for BCC and SCC. It further showed that the PRS can improve the prediction of BCC and SCC multiplicity over and above the established risk factors.
Our results are consistent with previous findings that SOTRs with a high genetic risk (e.g., those in the top PRS quintile) have a substantially higher risk of developing BCC or SCC than their counterparts with a low genetic risk (e.g., those in the bottom quintile) and that the PRS improves BCC and SCC risk prediction over and above the established clinical and skin cancer risk factors (Seviiri et al., 2021;Stapleton et al., 2019). This study differs from other previous studies in a number of ways. First, we used Ldpred (Vilhjálmsson et al., 2015), a method that considers a large number of genetic markers at the PRS generation stage, in contrast to the LD clump method used in the previous studies (Roberts et al., 2020;Seviiri et al., 2021;Stapleton et al., 2020Stapleton et al., , 2019 this problem by using the Ldpred method and training the models in an independent cohort using both different LD radius blocks (r 2 ¼ 2,000 kb and r 2 ¼ 5,000 kb) and the fractions of causal variants. Second, it has assessed the performance of a PRS in a high UV environment, where environmental factors greatly increase background BCC or SCC incidence. Previous studies have evaluated the PRS in environments with typically lower UV such as the United Kingdom and United States (Roberts et al., 2020;Seviiri et al., 2021;Stapleton et al., 2020Stapleton et al., , 2019. About half of the patients in the top quintile developed BCC and SCC within the relatively short 3-year follow-up period. In contrast, as we reported previously, only about 23% of SOTRs in the top quintile in the United Kingdom (a low UV setting) had developed BCC and SCC by late middle age (Seviiri et al., 2021). Third, as opposed to the follow-up of patients immediately after receiving their organ transplant, this study assessed SOTRs with a mean (SD) duration of 9.61 (8.50) years after transplantation and thus with chronic immunosuppression, another key risk factor for KC cancer in SOTRs.
Despite the chronic immunosuppression and other established risk factors, the PRSs were able to stratify the risks of both BCC and SCC. Therefore, a PRS can be of clinical importance at any stage of follow-up after transplantation.

Clinical utility
This study has shown that SOTRs with chronic immunosuppression in high UV settings can benefit from the PRS for BCC and SCC risk stratification and prediction (risk and multiplicity). Those at high, medium, and low genetic risk have markedly different ARs, with the risk stratification benefits continuing in the long term (10 years after transplantation). Indeed, the 19.03% (for BCC) and 18.10% (for SCC) of individuals whose risk category changed after adding the PRS may have their treatment options changed. For example, the 9.67% (for BCC) and 8.90% (for SCC) of SOTRs who are reassigned to a higher risk group may consequently have more intense KC cancer preventive interventions than their counterparts in the previously assigned group. The reverse may be applied to the 9.37% (for BCC) and 9.20% (for SCC) who move to a lower risk group. We show that a PRS that can (in combination with established risk factors) identify SOTRs at a very high AR of developing KC cancer in a high UV environment. These individuals may benefit from enhanced review and screening for KC cancer for purposes of early detection and prevention of KC cancer. In the Australian setting, all SOTRs are at nonnegligible KC cancer risk and are frequently placed on waiting lists for specialist dermatology care; further studies are merited to assess how effective a PRS-based approach would be in directing finite resources to those at highest risk. Internationally, in both the high UV setting considered in this study and in a lower UV setting considered previously (Seviiri et al., 2021), a PRS-based approach offers good stratification of risk in SOTRs, and future studies should assess country-specific economic factors underlying when the practical benefits of implementing PRS-based screening may be realized. PRSs improve the BCC and SCC risk and multiplicity predictions over and above the established risk factors among SOTRs with chronic immunosuppression in a high UV environment. The incorporation of PRSs into the clinical guidelines for KC cancer prevention, including screening, risk stratification, and prediction, may contribute to the reduction of the burden of these cancers among SOTRs with chronic immunosuppression in a high UV setting.

METHODS AND MATERIALS
Discovery cohorts for the PRS derivation: The UKB cohort and 23andMe We used the UKB and 23andMe cohorts to derive the discovery GWAS summary statistics for BCC and SCC. Detailed descriptions on recruitment, genotyping, quality control, and imputation procedures and processes for the UKB and 23andMe cohorts have been published elsewhere (Bycroft et al., 2018;Chahal et al., 2016b;Sudlow et al., 2015). 23andMe participants provided written informed consent and participated in the research online, under a protocol approved by the external Association for the Accreditation of Human Research Protection Programs-accredited Institutional Review Board, Ethical & Independent Review Services. Participants were included in the analysis on the basis of the consent status as checked at the time data analyses were initiated. For the UKB, the study was approved by the United Kingdom's National North West Multi-Centre Research Ethics Committee, and all participants provided written informed consent. In the UKB, we selected 307,684 nontransplant recipients (20,791 cases and 286,893 controls) for BCC and 294,294 (7,402 cases and 286,892 controls) nontransplant recipients for SCC of European ancestry. Cases were based on International Classification of Diseases 10 or International Classification of Diseases 9 codes, together with histologic International Classification of Diseases for Oncology, 3rd edition codes for BCC and SCC, excluding the respective in-situ BCC and SCC cases. Controls had no history of any cancer. The 23andMe cohort included 287,197 participants (12,945 BCC cases and 274,252 controls) for BCC and 287,137 (6,579 SCC cases and 280,558 controls) for SCC of European ancestry with self-reported data but with >90% sensitivity and >98% specificity for BCC and SCC with a low misclassification rate <10% (Chahal et al., 2016b). The UKB cases were ascertained through linkage of participant data with records at the national cancer registries.

Validation cohort for the PRSs: The QSkin prospective cohort
QSkin is a prospective population-based cohort of adult participants (n~43,000) residing in Queensland, a high UV index setting in Australia (annual average noon clear sky UV index ¼ 10). Participants were aged 40e60 years (mean age ¼ 56 years) and were randomly recruited from the population from 2011 with both clinically validated and self-reported data on skin cancers (Olsen et al., 2012). Details of the phenotype distributions have been published before (Olsen et al., 2012). In 2017, over 17,000 participants were genotyped using the Illumina GSA arrays and imputed to the Haplotype Reference Consortium panel (Loh et al., 2016

Prospective test cohort: the STAR cohort
The STAR study is a cohort of over 600 kidney, liver, and lung OTRs recruited respectively through the Princess Alexandra Hospital (Brisbane, Australia) and Prince Charles Hospital (Brisbane, Australia), the central referral hospitals for OTRs in the state of Queensland. Full details of the cohort and collection of phenotype data have been published previously (Hartman et al., 2018;Iannacone et al., 2015;Plasmeijer et al., 2019). Briefly, baseline recruitment was between 2012 and 2014, with subsequent annual follow-ups until the middle of 2016. At baseline, the key variables recorded were sex; date and type of transplantation (kidney, liver, and lung); duration of immunosuppression (time since transplantation); age at transplantation; skin reaction to the sun (only tans, burns then tans, and always burns); lifetime painful sunburns; sun exposure (during both weekdays and over the weekend); skin color (medium, olive, and fair); red hair color; and type of immunosuppressive medication, including calcineurin inhibitors (cyclosporine, tacrolimus), antimetabolites (azathioprine, mycophenolate), mammalian target of rapamycin inhibitors (sirolimus, everolimus), and corticosteroids (prednisone, used only in addition to other medication). Because all participants had been taking at least one immunosuppressive medication since the time of transplantation, the immunosuppressive medication variable was further reclassified as monotherapy (one immunosuppressive medication), double therapy (two medications), and triple therapy (at least three).
Dermatologists conducted skin examinations for every participant at study baseline and annual follow-up clinics, and all clinically diagnosed BCC/SCC lesions were referred for histologic confirmation. Between annual clinics, patients received quarterly phone calls to ascertain skin cancer treatments, and treating physicians confirmed all histologically diagnosed incident cancers. In addition, regular reviews of pathology laboratories' databases ensured the documentation of all newly diagnosed skin cancers. The study was approved by the human research ethics committees at Queensland Institute of Medical Research (QIMR) Berghofer Medical Research Institute and at Metro South Hospital and Health Service, Brisbane, Australia. All participants provided written informed consent.

Genotyping, quality control, and imputation of genetic data
In 2019, we extracted DNA for 375 adult participants, comprising 252 kidney, 30 liver, and 93 lung OTRs. DNA samples were genotyped using an Illumina GSA chip. We performed standard GWAS quality control procedures on the genotyped data. Individuals were excluded if they failed the sex and heterozygosity check, they were closely related (pihat >0.2), they had high genotype missingness (>3%), or they had divergent ancestry from CEU (European ancestry) (>6 SDs) of the HapMap phase 3 ( Figure 5). We further computed the first 10 PCs using selected autosomal SNPs to account for any subtle population stratification effects within the European ancestry group in the subsequent analyses. We also excluded SNPs with a call rate <95%, minor allele frequency <1%, and Hardy-Weinberg equilibrium P < 1 Â 10 e6 . Next, we imputed the genetic data to the Haplotype Reference Consortium reference panel (version r1.1 2016, European population) (Loh et al., 2016) using the Michigan Imputation Server (Das et al., 2016) and Eagle, version 2.4 (Loh et al., 2016) for phasing and Minimac4 for imputation. We retained the SNPs with an imputation score >0.3.

Skin cancer GWAS.
We used a training GWAS as conducted in our previous work (Seviiri et al., 2021). Briefly, using UKB data, we performed two case-control GWAS for BCC (20,791 BCC cases and 286,893 controls) and SCC (7,402 cases and 286,892 controls) that excluded SOTRs. The GWAS was conducted using a scalable and accurate implementation of a generalized mixed model (Zhou et al., 2018). Next, we obtained the 23andMe GWAS data for BCC (12,945 cases and 274,252 controls) (Chahal et al., 2016b) and SCC (6,579 cases and 280,558 controls) (Chahal et al., 2016a), and using a fixed-effects inverse variance weighted model, we performed a meta-analysis between the UKB GWAS and 23andMe GWAS for both BCC (total ¼ 33,736 cases and 561,145 controls) and SCC (total ¼ 13,981 cases and 567,450 controls) using METAL (Willer et al., 2010). We restricted the analysis to nonambiguous, autosomal, biallelic SNPs with a minor allele frequency >1%. In the initial GWAS, sex, age, and population stratification using PCs were adjusted for in both the UKB and 23andMe analyses.
Next, we identified the SNPs that were present in both our validation (QSkin) and target (STAR) cohorts. This resulted in 6,559,345 and 6,559,527 SNPs for BCC and SCC discovery GWAS metaanalysis, respectively.
Generation of the PRS models.
Our overall approach was to generate multiple PRS models and later select the one that performed best in an independent validation cohort (QSkin). We generated 14 PRS models (for each trait) in a systematic way using the LDpred method (Vilhjálmsson et al., 2015). LDpred is a Bayesian method that uses all SNPs in the GWAS and weights every SNP by the posterior mean of its conditional effect and LD information from the reference panel. First, using an LD reference panel of 2,000 unrelated individuals of European ancestry from UKB and the GWAS meta-analysis summary statistics (generated as discussed earlier) for BCC and SCC, we generated LDpred-adjusted effect estimates (log ORs) for BCC and SCC separately using different parameters. We first used an LD radius of 2,000 kb with varying fractions of causal SNPs, that is, Fi (infinitesimal model), F0 (1), F1 (0.1), F2 (0.01), F3 (0.001), F4 (0.0001), and F5 (0.00002). Then, we generated similar models using an LD radius of 5,000 kb but maintaining the fractions of causal SNPs mentioned earlier. Therefore, in total, we generated 14 PRS models for BCC and SCC that we applied to our validation data set to select the best predictive model.
Validation of the PRSs in the QSkin cohort.
Next, using the LDpred-adjusted effect sizes (log ORs) for the 14 models mentioned earlier as SNP weights and the imputed allelic dosages for the genotypes in QSkin, we generated individual PRSs using PLINK 1.9 (Chang et al., 2015). To select the best predictive model for each trait, we compared the model fit between a model with a BCC or SCC~PRS þ age þ sex þ10 PCs and a null model. We selected the best performing model using Nagelkerke's R 2 (Nagelkerke, 1991) computed using the predictABEL R package (Kundu et al., 2011). Model performance for both BCC and SCC in QSkin is presented in Figure 2a and Figure 2b, respectively. This process of selecting the single best model based on the QSkin cohort ensures that we do not induce bias from overfitting when applying our derived PRS to the STAR cohort.
Application of the best predictive PRS model in the STAR cohort.
Using the LDpred-adjusted effect estimates for the best predictive models for BCC and SCC, we generated the individual PRSs in the STAR cohort using imputed allelic dosages and PLINK 1.9. PRSs were normalized to express OR per SD increase in the PRS for the associations. We assessed the association between the PRS and the BCC/SCC risk by performing both simple and multiple logistic regression analyses (i.e., model 1, BCC or SCC~PRS þ 10 PCs; model 2, BCC or SCC~PRS þ 10 PCs þ age at transplantation þ sex; model 3, BCC or SCC~PRS þ major factors; and model 4, BCC or SCC~PRS þ all known factors). In model 3, major factors included age at transplantation, sex, type of transplantation, immunosuppressive medication (monotherapy, double therapy, and triple therapy), duration of immunosuppression, and skin color. All known factors in model 4 included established risk factors: age at transplantation, sex, type of organ transplantation, immunosuppressive medication, history of BCC or SCC, duration of immunosuppression, skin color, sun exposure, lifetime painful sunburns, and skin reaction to the sun. Model 4 was used as the final model.
Next, we divided the PRSs into quintiles. To evaluate whether the PRSs stratify the risk of BCC and SCC in SOTRs, we computed the ORs and 95% CIs for the BCC and SCC risk for participants with high genetic risk (those in the top quintile) and moderate risk (those in the middle 60%) compared with those of individuals in the bottom quintile (adjusting for the established skin cancer risk factors mentioned earlier and 10 PCs). Although the selection of these strata was arbitrary, they have been widely used in similar previous studies (Inouye et al., 2018;Torkamani et al., 2018). Next, we evaluated the ARs for BCC and SCC in the three strata mentioned earlier by computing the proportions of the participants who had developed BCC and SCC within the 3 years of follow-up. We further compared the ARs for BCC and SCC for the SOTRs in the STAR cohort with those in our independent QSkin validation cohort (in the same UV environment) after a similar follow-up period.
Next, we evaluated whether the PRSs improve the BCC and SCC risk predictions over and above the established risk factors by comparing the AUC for the prediction models with and without the PRS, that is, AUC for established risk factors þ 10 PCs versus AUC for established risk factors þ 10 PCs þ PRS. The AUC and 95% CI were computed using the PROC package (Robin et al., 2011) in R (R Foundation for Statistical Computing, Vienna, Austria). In addition, we calculated the NRI when the PRS is added to traditional risk factor models for both BCC and SCC using the predictABEL package (Kundu et al., 2011). We evaluated the NRI in two scenarios: (i) tertiles of risk (high, medium, low) and (ii) two categories (top 20% vs. bottom 80%).
Next, we evaluated whether the PRSs improve the predictions of multiple incident BCCs and/or SCCs per person during the study period over and above the established risk factors by comparing the R 2 explained for linear models with and without the PRS using the ANOVA test using R.

Sensitivity analyses.
We further explored whether the results were materially influenced by the 23andMe data by comparing the key findings for the original UKB þ 23andMe PRS versus the UKBonly PRS models. First, as described earlier, we used the UKB GWAS for BCC and SCC to develop and validate the UKB-only PRS models (for BCC and SCC). We compared the associations with BCC or SCC, prediction of BCC or SCC risk and multiplicity, AR, and the percentage of people reclassified when we used the original UKB þ 23andMe PRS versus the UKB-only PRS models (Supplementary  Table S2).

Data availability statement
The underlying data used to develop the polygenic risk scores are available from the UK Biobank with an approved UK Biobank application. The specific UK Biobank data fields (https://biobank. ndph.ox.ac.uk/showcase/search.cgi) used for the analysis of basal cell carcinoma and squamous cell carcinoma were fields 40006 and 40013 for International Classification of Diseases codes and field 40011 for International Classification of Diseases for Oncology 3 codes. For basal cell carcinoma, we analyzed the International Classification of Diseases for Oncology 3 codes 8090, 8091, 8092, 8093, 8094, 8097, and 8098 and codes 8070, 8071, 8072, 8073, 8074, 8075, 8076, and 8078 for squamous cell carcinoma. 23andMe Research Company allows applications to access previously published datasets (https://research.23andme.com/datasetaccess/), and we accessed and used specific GWAS summary statistics for basal cell carcinoma (Chahal et al., 2016b) and squamous cell carcinoma (Chahal et al., 2016a). The UK Biobank-only polygenic risk scores for both basal cell carcinoma and squamous cell carcinoma can be accessed at the polygenic risk score catalog (https://www.pgscatalog.org/) on publication. Data for validation and application of the polygenic risk score can be accessed through application to the QSkin Sun and Health Study principal investigator David Whiteman (David.Whiteman@qimrberghofer.edu.au) and the skin tumors in allograft recipients cohort principal investigator Adele Green (Adele.Green@qimrberghofer.edu.au).  Sun exposure: mild, 5 hours during weekdays and weekends; moderate, 5þ hours during either the weekdays or weekends; excessive, 5þ hours during both the weekdays and weekends.