This study has limitations First

This study has limitations. First, the four studies were designed independently and involved different methods for recording toxicity. Endpoints needed harmonizing and the approach might be sub-optimal, but serves as a useful guide for future radiogenomics studies involving multiple cohorts. Although endpoints were dichotomized to minimize differences in toxicity grades across the cohorts, the simulation experiment showed minimal loss of statistical power. Second, there was the heterogeneity in radiotherapy protocols. For example, Gene-PARE included individuals who received brachytherapy with EBRT, which increases urinary toxicity (Sanda et al., 2008). Similarly, the proportion of individuals who received androgen deprivation differed across the studies. This heterogeneity was adjusted for using multivariable analysis and by a meta-analysis approach. Importantly, none of the top SNPs showed evidence of heterogeneity in effect size across studies (Tables 5–8). The first SNPs identified are those that rise above such noise in the data, and the significance of our work is that we can find variants despite imperfect datasets. We attempted to control for differences in radiation dose by adjusting for total biologically effective dose in each GWAS. However, this is a surrogate for the doses received by specific normal tissues. Future cohorts with detailed dosimetry data to the different gsk-3 at risk will improve on our ability to identify SNPs. The heterogeneity in cohorts should be embraced, as any predictive test must use SNPs that are associated with toxicity across treatment centers and protocols. Last, in order to minimize confounding by population structure, the present analysis was restricted to individuals of European ancestry. It will be important for future studies to focus on other ancestral groups, both for replication of the loci identified here and for discovery of additional loci. While this study was successful in identifying radiosensitivity SNPs via a meta-analysis of GWAS, it is modestly sized compared with GWAS of other diseases and traits, and it should be viewed as the starting point for expanded radiogenomics studies. Given that our study was not powered to detect SNPs with odds ratio<2, many true positive SNPs will have been missed. Also, though the SNPs identified here reach statistical significance and show consistency across multiple independent studies, there is still a possibility that they will fail to replicate in other cohorts. This is a challenge in GWAS (McCarthy et al., 2008), but meta-analysis and replication studies have proved successful in validating hundreds of risk SNPs for a wide variety of diseases and phenotypes. Future large GWAS meta-analyses will identify additional SNPs, and this work is underway. The Radiogenomics Consortium is a participant in the OncoArray Network, which is a large collaborative effort to gain new insight into the genetic architecture and mechanisms underlying several cancers as well as outcomes related to the their treatment. Approximately 5000 additional DNA samples from individuals treated with radiotherapy for prostate cancer are being genotyped using the customized OncoArray, and meta-analysis of this greatly expanded set of GWAS data using the methods developed in this paper will uncover additional risk SNPs, provide a platform for further validation of the SNPs identified here, and serve as the basis for future post-GWAS analyses on these confirmed loci. The last decade saw a rapid expansion of knowledge of the genetics of disease susceptibility in the general population. Numerous large collaborative GWAS showed that polygenic risk profiles can be built based on multiple SNPs each conferring small effects but together a significant proportion of susceptibility to common diseases. With the rapid decline in costs for genetic testing, there is growing acceptance that risk prediction models incorporating genetic and environmental factors will be important in future healthcare provision (Chatterjee et al., 2016). In 2012 there were an estimated 32.6million people alive five years after being diagnosed with cancer (http://www.cdc.gov/cancer/international/statistics.htm), and many will be living with the consequences of treatment. Research developing models predicting susceptibility to long-term radiation effects is important and the findings reported here shows that heterogeneous radiotherapy cohorts can be combined to identify common genetic variants associated with toxicity. The work provides a basis for larger collaborative efforts to identify enough variants for a future test involving polygenic risk profiling.