Blood lipids and prostate cancer: a Mendelian randomization analysis

1125 ORIGINAL RESEARCH Blood lipids and prostate cancer: a Mendelian randomization analysis Caroline J. Bull1,2,3,4, Carolina Bonilla1,2,3, Jeff M. P. Holly3,4, Claire M. Perks3,4, Neil Davies1,2,3, Philip Haycock1,2,3, Oriana Hoi Yun Yu5, J. Brent Richards5,6, Rosalind Eeles7,8, Doug Easton9, Zsofia KoteJarai7, Ali Amin Al Olama9, Sara Benlloch9, Kenneth Muir3,10,11, Graham G. Giles12,13, Robert J. MacInnis12,13, Fredrik Wiklund14, Henrik Gronberg14, Christopher A. Haiman15, Johanna Schleutker16,17, Børge G. Nordestgaard18, Ruth C. Travis19, David Neal20,21, Nora Pashayan9,20,21,22, Kay-Tee Khaw23, Janet L. Stanford24,25, William J. Blot26, Stephen Thibodeau27, Christiane Maier28,29, Adam S. Kibel30,31, Cezary Cybulski32, Lisa Cannon-Albright33, Hermann Brenner34,35,36, Jong Park37, Radka Kaneva38, Jyotsna Batra39, Manuel R. Teixeira40,41, Agnieszka Micheal42, Hardev Pandha42, George Davey Smith1,2, Sarah J. Lewis1,2,3,a & Richard M. Martin1,3,43,a The PRACTICAL consortiumb 1School of Social and Community Medicine, University of Bristol, Bristol, United Kingdom 2MRC/University of Bristol Integrative Epidemiology Unit, University of Bristol, Bristol, United Kingdom 3Integrative Cancer Epidemiology Programme, University of Bristol, Bristol, United Kingdom 4IGFs and Metabolic Endocrinology Group, School of Clinical Sciences North Bristol, University of Bristol, Bristol BS10 5NB, United Kingdom 5Department of Medicine, Division of Endocrinology, Epidemiology Biostatistics and Occupational Health McGill University, Montreal, Quebec, Canada 6Department of Twin Research, King’s College London, London SE1 7EH, United Kingdom 7The Institute of Cancer Research, London SM2 5NG, United Kingdom 8The Royal Marsden NHS Foundation Trust, London SW3 6JJ, United Kingdom 9Centre for Cancer Genetic Epidemiology, Department of Public Health and Primary Care, University of Cambridge, Strangeways Laboratory, Worts Causeway, Cambridge, United Kingdom 10Warwick Medical School, University of Warwick, Coventry CV4 7AL, United Kingdom 11Institute of Population Health, University of Manchester, Manchester M13 9PL, United Kingdom 12Cancer Epidemiology Centre, The Cancer Council Victoria, 615 St Kilda Road, Melbourne, Victoria, Australia 13Centre for Epidemiology and Biostatistics, Melbourne School of Population and Global Health, The University of Melbourne, Victoria, Australia 14Department of Medical Epidemiology and Biostatistics, Karolinska Institute, Stockholm, Sweden 15Department of Preventive Medicine, Keck School of Medicine, University of Southern California/Norris Comprehensive Cancer Center, Los Angeles, California 16Department of Medical Biochemistry and Genetics, University of Turku and Tyks Microbiology and Genetics, Department of Medical Genetics, Turku University Hospital, Turku, Finland 17Institute of Biomedical Technology/BioMediTech, University of Tampere and FimLab Laboratories, Tampere, Finland 18Department of Clinical Biochemistry, Herlev and Gentofte Hospital, Copenhagen University Hospital, Herlev Ringvej 75, Herlev DK-2730, Denmark 19Cancer Epidemiology, Nuffield Department of Population Health, University of Oxford, Oxford, United Kingdom 20Surgical Oncology (Uro-Oncology: S4), University of Cambridge, Addenbrooke’s Hospital, Box 279, Hills Road, Cambridge, United Kingdom 21Cancer Research UK Cambridge Research Institute, Li Ka Shing Centre, Cambridge, United Kingdom 22Department of Applied Health Research, University College London, 1-19 Torrington Place, London WC1E 7HB, United Kingdom 23Cambridge Institute of Public Health, University of Cambridge, Forvie Site, Robinson Way, Cambridge, CB2 0SR, United Kingdom 24Division of Public Health Sciences, Fred Hutchinson Cancer Research Center, Seattle, Washington 25Department of Epidemiology, School of Public Health, University of Washington, Seattle, Washington 26International Epidemiology Institute, 1455 Research Blvd., Suite 550, Rockville 20850, Maryland 27Mayo Clinic, Rochester, Minnesota 28Department of Urology, University Hospital, Ulm, Germany 29Institute of Human Genetics University Hospital, Ulm, Germany 30Brigham and Women’s Hospital/Dana-Farber Cancer Institute, 45 Francis StreetASB II-3, Boston, Massachusetts 02115 31Washington University, St. Louis, Missouri 32International Hereditary Cancer Center, Department of Genetics and Pathology, Pomeranian Medical University, Szczecin, Poland 33Division of Genetic Epidemiology, Department of Medicine, University of Utah School of Medicine, Salt Lake City, Utah 34Division of Clinical Epidemiology and Aging Research, German Cancer Research Center (DKFZ), Heidelberg, Germany 35Division of Preventive Oncology, German Cancer Research Center (DKFZ) and National Center for Tumor Diseases (NCT), Heidelberg, Germany 36German Cancer Consortium (DKTK), German Cancer Research Center (DKFZ), Heidelberg, Germany 37Division of Cancer Prevention and Control, H. Lee Moffitt Cancer Center, Magnolia Dr., Tampa, Florida 12902 38Molecular Medicine Center and Department of Medical Chemistry and Biochemistry, Medical University Sofia, 2 Zdrave St, Sofia, 1431, Bulgaria 39Australian Prostate Cancer Research Centre-Qld, Institute of Health and Biomedical Innovation and School of Biomedical Sciences, Queensland University of Technology, Brisbane, Australia 40Department of Genetics, Portuguese Oncology Institute, Porto, Portugal 41Biomedical Sciences Institute (ICBAS), Porto University, Porto, Portugal 42The University of Surrey, Surrey GU2 7XH, United Kingdom 43National Institute for Health Research, Bristol Nutrition Biomedical Research Unit, Bristol, United Kingdom © 2016 The Authors. Cancer Medicine published by John Wiley & Sons Ltd. This is an open access article under the terms of the Creative Commons Attribution License, which permits use, distribution and reproduction in any medium, provided the original work is properly cited. Cancer Medicine Open Access 1126 © 2016 The Authors. Cancer Medicine published by John Wiley & Sons Ltd. C. J. Bull et al. Blood Lipids and Prostate Cancer Introduction Prostate cancer is the most prevalent male cancer in Europe and a major cause of cancerrelated deaths [1]. Lifestyle factors and related intermediate phenotypes have been associated with prostate cancer development and progression in epidemiological studies, including a positive association between circulating cholesterol levels and prostate cancer [2–4]. However, conclusions are conflicting and it is not clear whether these findings reflect causality or Keywords Cholesterol, Mendelian randomization, prostate cancer, statins Correspondence Richard M. Martin, School of Social and Community Medicine, University of Bristol, Bristol, United Kingdom. Tel: 01179287321; Fax: 0117 928 7236; E-mail: [email protected] Funding Information C. J. B. is funded by the Wellcome Trust 4-year studentship WT083431MA. The Integrative Cancer Epidemiology Programme is supported by Cancer Research UK programme grant C18281/A19169. The MRC IEU is supported by the Medical Research Council and the University of Bristol (MC_UU_12013/1-9). The NIHR Bristol Nutrition Biomedical Research Unit is funded by the National Institute for Health Research (NIHR) and is a partnership between University Hospitals Bristol NHS Foundation Trust and the University of Bristol. The CRUK study and PRACTICAL consortium is supported by the Canadian Institutes of Health Research, European Commission’s Seventh Framework Programme grant agreement no. 223175 (HEALTH-F2-2009-223175), Cancer Research UK Grants C5047/A7357, C1287/A10118, C5047/A3354, C5047/A10692, and C16913/ A6135. The National Institutes of Health (NIH) Cancer Post-Cancer GWAS initiative grant no. 1 U19 CA 148537-01 (the GAME-ON initiative) and NIHR support to the Biomedical Research Centre and The Institute of Cancer Research and Royal Marsden NHS Foundation Trust. The funders had no role in the design and conduct of the study; collection, management, analysis, and interpretation of the data; and preparation, review, or approval of the manuscript. Received: 13 November 2015; Revised: 3 February 2016; Accepted: 8 February 2016 Cancer Medicine 2016; 5(6):1125–1136 doi: 10.1002/cam4.695 aThese authors contributed equally to the work. bAdditional cohort members are listed in the supporting information. Abstract Genetic risk scores were used as unconfounded instruments for specific lipid traits (Mendelian randomization) to assess whether circulating lipids causally influence prostate cancer risk. Data from 22,249 prostate cancer cases and 22,133 controls from 22 studies within the international PRACTICAL consortium were analyzed. Allele scores based on single nucleotide polymorphisms (SNPs) previously reported to be uniquely associated with each of lowdensity lipoprotein (LDL), highdensity lipoprotein (HDL), and triglyceride (TG) levels, were first validated in an independent dataset, and then entered into logistic regression models to estimate the presence (and direction) of any causal effect of each lipid trait on prostate cancer risk. There was weak evidence for an association between the LDL genetic score and cancer grade: the odds ratio (OR) per genetically instrumented standard deviation (SD) in LDL, comparing high- (≥7 Gleason score) versus lowgrade (<7 Gleason score) cancers was 1.50 (95% CI: 0.92, 2.46; P = 0.11). A genetically instrumented SD increase in TGs was weakly associated with stage: the OR for advanced versus localized cancer per unit increase in genetic risk score was 1.68 (95% CI: 0.95, 3.00; P = 0.08). The rs12916T variant in 3hydroxy3methylglutarylCoA reductase (HMGCR) was inversely associated with prostate cancer (OR: 0.97; 95% CI: 0.94, 1.00; P = 0.03). In conclusion, circulating lipids, instrumented by our genetic risk scores, did not appear to alter prostate cancer risk. We found weak evidence that higher LDL and TG levels increase aggressive prostate cancer risk, and that a variant in HMGCR (that mimics the LDL lowering effect of statin drugs) reduces risk. However, inferences are limited by sample size and evidence of pleiotropy. 1127 © 2016 The Authors. Cancer Medicine published by John Wiley & Sons Ltd. Blood Lipids and Prostate Cancer C. J. Bull et al. are the product of confounding by common causes of both cholesterol levels and prostate cancer (e.g., aspects of diet), bias, or reverse causality (the cancer causing altered cholesterol metabolism) [5]. As serum cholesterol levels can be modified by lifestyle changes [6] and statin therapy [7], clarifying the causality of this association could inform the development of prevention interventions for prostate cancer. Statins lower cholesterol levels by inhibiting HMGCoA reductase, the ratelimiting enzyme in cholesterol synthesis. A metaanalysis of 27 observational studies concluded that statin therapy reduced prostate cancer by 7% (RR 0.93; 95% CI: 0.87, 0.99; P = 0.03) [8], but a separate metaanalysis of four randomized trials showed minimal evidence of any association (RR 1.08; 95% CI: 0.91, 1.30; P = 0.38) [9]. Both analyses revealed considerable heterogeneity between the included studies (I2 values >70%) and, as prostate cancer was assessed as a secondary outcome in the trials, misclassification of outcome could have biased the results. As yet, evidence regarding statin therapy for prostate cancer is controversial [10]. Mendelian randomization uses genetic variants robustly associated with traits of interest (in this case, circulating lowdensity lipoprotein [LDL], highdensity lipoprotein [HDL], and triglyceride [TG] levels) as instrumental variables to make inferences about whether associations between exposures and disease are likely to be causal [11]. The principle of Mendelian randomization is that analysis of groups defined by common genetic variants is analogous to that of an intentiontotreat analysis in a randomized controlled trial, based on Mendel’s laws of segregation and independent assortment. Using genetic variants as “instrumental variables” to proxy modifiable exposures should be unconfounded by environmental factors, represent lifelong exposure, and not be subject to reverse causality with respect to the phenotype proxied by the genotype. A genetic score using several single nucleotide polymorphisms (SNPs) in combination can be constructed to represent the additive effect of multiple gene variants to explain more of the variance in the risk factor of interest, avoiding weak instrument bias, and increasing power [12]. Genetic scores using multiple gene variants for lipid traits have been used previously to investigate associations between blood lipids and vascular disease [13–15]. If the instrumental variable assumptions hold, as shown in Figure 1, then a test of the association between the instrument and the outcome is a test of the presence of a causal effect of the intermediate on the outcome [16]. Here we test for the possible presence of a causal effect of lipid fractions (LDL, HDL, TG) on prostate cancer using genetic variants for these traits combined in genetic risk scores using the principle of Mendelian randomization [11]. As far as we know, this study is the first to employ Mendelian randomization to investigate the association between circulating lipids and prostate cancer risk. Methods Genetic risk scores as instruments for circulating lipid fractions were developed using SNPs previously identified by genomewide association studies (GWAS). These scores were then applied to 22,249 prostate cancer cases and 22,133 controls within the international PRACTICAL consortium [17] for whom genetic data were available. Study populations PRACTICAL consortium (prostate cancer association group to investigate cancerassociated alterations in the genome) We investigated the association between genetic risk scores for lipid traits and prostate cancer risk in an individual participant metaanalysis of men in 22 studies of the international PRACTICAL consortium (Table 1). Fifteen of these studies were based in Europe, five in North America, and two in Australia. Genotypic information was available for 44,382 participants (22,249 prostate cancer cases and 22,133 controls). Subjects were predominantly of European descent (mean across studies = 99%). All studies met the appropriate ethical criteria for each country in accordance with the principles embodied in the Declaration of Helsinki. Further details are available from Figure 1. Mendelian randomization. Using genetic variants as instrumental variables to establish whether an exposure is causally related to cancer. An instrumental variable (genetic variation) [Z] acts as a proxy for environmental exposure [X], postulated to influence cancer [Y]. Z is independent of measured or unmeasured confounders [U]. Z only influences Y if X →Y is causal. Genetic risk scores for LDL, HDL and TG. Comprised of genetic variants [Z] Circulating LDL, HDL and TG [X] Prostate cancer [Y] Measured and Unmeasured confounders [U] 1128 © 2016 The Authors. Cancer Medicine published by John Wiley & Sons Ltd. C. J. Bull et al. Blood Lipids and Prostate Cancer Table 1. Summary of 22 PRACTICAL case–control studies, n = 44,382 men. Study Acronym Country n Mean % Controls Cases Age at diagnosis (years) PSA at diagnosis (ng/mL) European ethnicity Family history of prostate cancer High grade1 Advanced cancer1 Cancer of the Prostate in Sweden CAPS Sweden 664 1153 66.7 79.6 100 11.9 50.0 30.3 Copenhagen Prostate Cancer Study 1 CPCS1 Denmark 2771 848 59.5 48.0 100 8.2 71.2 — Copenhagen Prostate Cancer Study 2 CPCS2 Denmark 1009 265 58.0 36.0 100 14.7 52.2 — European Prospective Investigation Into Cancer and Nutrition (BPC3) EPIC Europe 1079 722 61.7 — 100 — 27.9 4.0 Epidemiological investigations of the chances of preventing, recognizing early and optimally treating chronic diseases in an elderly population ESTHER Germany 318 313 65.4 58.7 100 6.4 48.0 27.6 Fred Hutchinson Cancer Research Centre FHCRC USA 729 761 59.7 16.1 91.4 16.2 41.7 20.2 Portuguese Oncology Institute, Porto IPOPorto Portugal 66 183 54.0 8.3 100 20.0 84.2 13.1 Mayo Clinic MAYO USA 488 767 65.3 15.5 100 23.3 55.3 45.5 Melbourne Collaborative Cohort Study MCCS Australia 1170 1698 56.0 19.4 99.8 29.6 53.1 14.5 Multiethnic Cohort Study (BPC3) MEC USA 829 819 70.0 — 100 10.7 — 12.5 The Moffitt Group MOFFITT USA 100 414 64.5 6.5 89.4 20.1 43.1 3.5 Prostate Cancer study Medical University Sofia PCMUS Bulgaria 140 151 68.2 23.1 100 4.1 59.6 46.7 The Poland Group Poland Poland 359 438 65.5 40.2 100 6.9 32.8 37.1 Prostate Project Foundation–Postgraduate Medical School, Surrey PPFUNIS UK 188 245 67.9 19.3 99.6 25.2 45.5 28.2 Prostate testing for cancer and Treatment ProtecT UK 1474 1563 61.2 6.3 99.7 7.0 29.9 11.4 Retrospective Queensland Study (QLD) and the Prostate Cancer Supportive Care and Patient Outcomes Project (ProsCan) QLD Australia 87 186 64.3 6.7 100 25.9 83.1 — Study of Epidemiology and Risk factors in Cancer Heredity SEARCH UK 1244 1371 58.9 53.2 99.9 16.2 56.8 17.8 Stockholm 1 STHMI Sweden 2224 2006 66.7 — 100 17.1 45.5 14.4 Finnish Genetic Predisposition to Prostate Cancer Study TAMPERE Finland 2413 2754 68.2 69.1 100 — 43.7 21.4 U.K. Genetic Prostate Cancer Study and The Prostate Cancer Research Foundation Study UKGPCS UK 4182 4549 61.1 46.9 100 24.7 50.5 36.0 Institut fuer Humangengetik Ulm ULM Germany 354 603 62.4 19.1 100 33.4 51.3 40.5 UTAH Study UTAH USA 245 440 64.0 — 100 33.1 — 17.2 Missing data were excluded from each analysis. —, no data available. 1Advanced cancer (T3 to T4 or SEER staging regional or distant), high grade (Gleason ≥7), low grade (Gleason ≤6). 1129 © 2016 The Authors. Cancer Medicine published by John Wiley & Sons Ltd. Blood Lipids and Prostate Cancer C. J. Bull et al. the consortium website (http://www.practical.ccge.medschl. cam.ac.uk). Data concerning cancer grade and stage were collected by each study. Cancers were categorized as low grade (Gleason score ≤6) or high grade (Gleason score ≥7), and localized (T1 or T2 on TNM staging, or if not available, “localized” on SEER staging) or advanced (T3 or T4 on TNM staging, or if not available, “regional” or “distant” on SEER staging). Data were not available for grade in two (MEC/UTAH), or stage in three (CPCS1/CPCS2/QLD), of the PRACTICAL studies, respectively. GWAS identification of lipid SNPs SNPs associated with the lipid traits under investigation were identified by a review of the current published literature in which papers concerning lipid SNPs were identified by entering appropriate search terms into the web of science database (“Lipid,” “SNP,” “GWAS”). We selected SNPs that were exclusively associated, at genomewide significance, with each lipid trait of interest. The majority of SNPs were taken from two landmark lipid GWAS [18, 19], the latter being the largest genetic association study of blood lipid levels to be conducted to date (188,577 individuals). For completeness, a number of SNPs from other publications were also included in the analysis [14, 20]. The EPICNorfolk cohort was used by the referenced GWASs to identify SNPs associated with lipid traits; we therefore excluded this study from our analysis to reduce the chance of underestimating the causal effect of the lipids on prostate cancer. Genotyping data PRACTICAL samples were genotyped using an Illumina Custom Infinium genotyping array (iCOGS) consisting of 211,155 SNPs designed for the Collaborative Oncological Gene–Environment Study (COGS) (details of which may be found on their website http://www. cogseu.org). The array was specifically designed for the evaluation of genetic variants in breast, ovarian, and prostate cancer. As genotypic information was not available for all SNPs in the genetic risk scores, we also used SNPs that had been imputed using IMPUTE2 software [21]. As a sensitivity analysis, allele frequencies and ORs for prostate cancer outcomes using imputed and genotyped data were compared by cross tabulation (Table S1). As results were similar, imputed data were used in all subsequent analyses. All SNPs with an indication of poor imputation quality were removed (r2 hat<0.30), as were those with a minor allele frequency of <1%, a call rate of <95%, or those that violated the Hardy–Weinberg equilibrium (P < 0.05). Gene variants used to create the genetic risk scores Genetic risk scores were used as instruments to proxy exposure to circulating blood lipids in a Mendelian randomization framework [22]. This analysis assumes that the genetic risk scores used in the analysis influence prostate cancer risk only via their ability to alter the specific lipid trait that they are acting as proxies for (e.g., LDL, Fig. 1). If other biochemical processes or traits are associated with the genetic risk scores (i.e., horizontal pleiotropy is present, where pathways from lipidassociated SNPs to disease are involved that are independent of lipids, see Box 1) and also directly alter prostate cancer risk, this violates a major assumption of Mendelian randomization by introducing genetic confounding. Therefore, only SNPs exclusively associated at genomewide significance (P < 5 × 10−8) with one lipid trait (either LDL or HDL or TGs, but not more than 1) were chosen for each genetic risk scores to make them as specific to one trait as possible (as in Holmes et al. [15]) (Table S2), minimizing the possibility of pleiotropic effects. Of the 118 SNPs identified as associated with only one lipid trait, 62 were either genotyped or imputed in the PRACTICAL consortium and considered eligible for inclusion in the genetic risk scores based on current biological knowledge. SNPs selected for the genetic risk scores were tested for evidence of linkage disequilibrium (LD) using the SNAP pairwise online tool (http:// www.broadinstitute.org/mpg/snap/ldsearchpw.php), an open access resource which uses pairwise LD data based on phased genotype data from the International HapMap project. A threshold r2 value of ≥0.85 was used to indicate LD; where SNPs were in LD, the SNP with the largest effect on the lipid trait was selected for the genetic risk scores. SNP genotypes were coded as 0, 1, or 2 depending on exposure to the risk allele. Dosage values (ranging from 0 to 2) were generated for imputed SNPs. LDL, HDL, and TG scores were composed of 11, 36, and 15 SNPs, respectively. Genetic risk scores were created by summing the number of “risk” alleles that each of the 22,249 prostate cancer cases and 22,133 controls were exposed to, such that the greater the number of “risk” alleles a man had, the higher the score. “Risk” alleles were those that were positively related to serum LDL or TG, or negatively related to serum HDL. Published effect sizes (the effect Box 1. Pleiotropy Vertical pleiotropy: A genetic locus is linked to a cascade of events. This is not generally a problem for Mendelian randomization studies. Horizontal pleiotropy: A genetic locus is related to multiple phenotypes. This violates Mendelian randomization assumptions. 1130 © 2016 The Authors. Cancer Medicine published by John Wiley & Sons Ltd. C. J. Bull et al. Blood Lipids and Prostate Cancer of the risk allele on the trait in SD) were applied to each SNP and summated to give a “weighted” genetic risk score for each trait [23], for each individual man in each study, so that when estimating the effect of a unit increase in the genetic risk score directly translated to the effect of a SD change in the trait upon the outcome. In a subsidiary analysis we examined the association of a SNP in 3hydroxy3methylglutarylCoA reductase (HMGCR) (rs12916) in relation to prostate cancer outcomes. rs12916T has been used previously to mimic statin intervention in order to estimate the causal association of statin use with type 2 diabetes and adiposity measures [24]. ORs for prostate cancer outcomes were reported per rs12916T allele. Validation of genetic risk scores and investigation of potential pleiotropy in the ALSPAC cohort The Avon Longitudinal Study of Parents and Children (ALSPAC) is a birth cohort established to investigate environmental and genetic factors in health and development [25]. ALSPAC data were used to validate the genetic risk scores as instruments for circulating lipid exposure using linear regression. Ten of the 11 SNPs in the LDL score could be included in score validation analyses using ALSPAC data (due to poor imputation quality in ALSPAC, rs1801689 could not be included in the LDL score validation). Thirtyfive of the 36 HDL genetic risk score SNPs could be validated in ALSPAC (rs1084651 was poorly imputed) and 14 of the 15 TG genetic risk score SNPs were available for validation in the ALSPAC cohort (rs11649653 was poorly imputed). Statistical analysis The genetic risk scores were entered into logistic regression models to estimate the effect per genetically instrumented SD increase in the lipid trait on prostate cancer outcomes. Outcomes investigated were: all prostate cancer (case vs. control status), grade (high [≥7] versus low [≤6] Gleason score), and stage (advanced versus localized TNM or SEER staging). We conducted the analyses within each of the individual studies that make up the PRACTICAL consortium and then combined the results into a summary odds ratio (OR) for each outcome per unit increase in the genetic risk scores by fixed effect metaanalysis using the “metan” command in Stata v.13 [26]. To test that the instruments (genetic risk scores) were not associated with confounders, we investigated whether each genetic risk score was associated with available covariables that could be potential confounding factors (age, diagnostic prostate specific antigen (PSA) levels, and family history of prostate cancer). To account for potential confounding by population stratification, we adjusted for the top eight principle components (variables concerning the population’s genetic architecture). All analyses were performed in Stata v.13 (Stata Corp LP, 2013, College Station, TX). Results The genetic risk scores were validated in ALSPAC participants at the age of 7 years (Table 2). The genetic risk score for LDL explained 0.3% of the variability in circulating LDL and not only was most strongly associated with LDL (linear regression coefficient; 0.56 mmol/L LDL; F stat: 14.11; P = 2 × 10−4), but was also weakly associated with HDL (linear regression coefficient; 0.19 mmol/L HDL; F stat: 5.32; P = 0.02). The HDL genetic risk score was the strongest instrument, explaining 0.9% of variability in circulating HDL and was associated exclusively with HDL (linear regression coefficient; −0.25 mmol/L HDL; F stat: 37.60; P = 9.5 × 10−10). The TG genetic risk score was most strongly associated with TG and explained 0.2% Table 2. Weighted genetic risk score validation in ALSPAC (N = 4081). Genetic risk score Change in trait levels (mmol/L) per unit score195% CI P value r2 (%) F LDL (10 SNPs) LDL 0.56 0.27, 0.86 2 × 10−4 0.34 14.11 HDL 0.19 0.03, 0.35 0.021 0.13 5.32 lnTG −0.18 −0.41, 0.06 0.141 0.05 2.17 HDL (35 SNPs) HDL −0.25 −0.32, −0.17 9.50 × 10−10 0.91 37.60 LDL −0.12 −0.25, 0.04 0.16 0.05 1.97 lnTG 0.05 −0.06, 0.17 0.36 0.02 0.36 TG (14 SNPs) lnTG 0.35 0.13, 0.57 0.002 0.24 9.76 LDL 0.12 −0.16, 0.39 0.40 0.02 0.70 HDL −0.19 −0.34, −0.04 0.01 0.15 6.28 1TG levels have been natural log transformed. 1131 © 2016 The Authors. Cancer Medicine published by John Wiley & Sons Ltd. Blood Lipids and Prostate Cancer C. J. Bull et al. of the variability in circulating TG (linear regression coefficient; 0.35 mmol/L lnTG; F stat: 9.76; P = 2 × 10−3); however, it was also associated with HDL (0.15% variability explained, linear regression coefficient; −0.19 mmol/L HDL; F stat: 6.28; P = 0.01). Associations of the genetic risk scores with each lipid trait remained unaltered following stratification of participants by sex (data available on request). Individual data from 22,249 case and 22,133 control men in 22 PRACTICAL studies were included in our analysis (Table 1). The percentage of highgrade cancers reported varied between studies (27.9–84.2%), as did the proportion of advanced stage cancers (3.5–46.7%). There was little evidence to support an association between genetic risk scores and PSA at recruitment. Family history of prostate cancer was weakly, and imprecisely, associated with the LDL (OR: 0.50, 95% CI: 0.21, 1.19; P = 0.12) and HDL (OR: 1.62, 95% CI: 1.02, 2.58; P = 0.04) genetic risk scores. Four of the top eight principal components were associated with at least one of the genetic risk scores. Age at interview was not associated with the genetic risk scores, with the exception of the TG score (linear regression coefficient: −2.76, 95% CI: −5.52, −0.0003; P = 0.05) (Table 3). The pooled ORs for overall prostate cancer risk by lipid trait, estimated in instrumental variable analysis using the genetic risk score, were 1.24 (95% CI: 0.90, 1.69; P = 0.18; I2 = 4.4%), 0.99 (95% CI: 0.84, 1.17; P = 0.90; I2 = 0%), and 1.09 (95% CI: 0.80, 1.50; P = 0.57; I2 = 14.4%) per genetically instrumented SD increase in LDL, HDL, and TG, respectively (Fig. 2). The ORs for prostate cancer outcomes stratified by grade and stage are summarized in Table 4. There was weak evidence to suggest an effect for a genetically instrumented SD increase in LDL between highversus lowgrade cancer cases (OR: 1.50; 95% CI: 0.92, 2.46; P = 0.11). There was little evidence to suggest an association between the HDL and TG genetic risk scores and prostate cancer grade: When highand lowgrade cancers were compared, a genetically instrumented SD decrease in HDL gave an OR of 1.03 (95% CI: 0.79, 1.34; P = 0.82), and a genetically instrumented SD increase in TG gave an OR of 0.93 (95% CI: 0.57, 1.52; P = 0.77). For cancer stage, a genetically instrumented SD increase in LDL gave an OR for advanced versus localized cancers of 0.91 (95% CI: 0.51, 1.64; P = 0.77). A genetically instrumented SD decrease in HDL gave an OR for advanced versus localized cancers of 1.02 (95% CI: 0.74, 1.39; P = 0.92). The OR for advanced versus localized prostate cancer per SD genetically instrumented increase in TG was 1.68 (95% CI: 0.95, 3.00; P = 0.08). EPICNorfolk was excluded from the analysis, as the cohort was included in the referenced GWAS discovery panels. As a sensitivity analysis, we included EPICNorfolk in the metaanalysis. The results for EPICNorfolk were in agreement with the combined estimates for the odds of prostate cancer by the genetic risk scores (Fig. S1). As some associations were present between the genetic risk scores and family history or age at diagnosis (Table 3), we performed a sensitivity analysis adjusting for these variables. The results were unaltered following adjustment for age with the TG score and for family history with the LDL and HDL scores (data available on request). The rs12916T variant in HMGCR was weakly associated with a decreased risk of prostate cancer overall (OR: 0.97; 95% CI: 0.94, 1.00; P = 0.03; I2 = 0%) (Fig. 3). Table 3. Association genetic risk scores with potential confounding variables in 22,133 PRACTICAL control men. Variable nLDL HDL TG Change in variable per unit increase genetic risk score1 (95% CI), P value Principle component 1 22,133 −0.20 (−0.45, 0.04), 0.10 −0.08 (−0.29, 0.14), 0.47 −0.17 (−0.51, 0.16), 0.30 Principle component 2 22,133 0.80 (0.43, 1.17), 2 × 10−4 −0.06 (−0.19, 0.07), 0.35 0.70 (0.24, 1.15), 4 × 10−3 Principle component 3 22,133 −0.54 (−0.81, −0.27), 4 × 10−4 0.01 (−0.10, 0.12), 0.82 −0.36 (−0.60, −0.13), 4 × 10−3 Principle component 4 22,133 0.28 (−0.19, 0.74), 0.23 0.06 (−0.15, 0.28), 0.54 −0.49 (−0.82, −0.16), 0.01 Principle component 5 22,133 0.35 (−0.18, 0.87), 0.18 −0.10 (−0.29, 0.09), 0.29 0.21 (−0.20, 0.63), 0.30 Principle component 6 22,133 −0.56 (−1.00, −0.15), 0.01 0.11 (−0.06, 0.28), 0.19 0.19 (−0.19, 0.57), 0.32 Principle component 7 22,133 −0.08 (−0.39, 0.23), 0.61 0.23 (−0.06, 0.53), 0.11 0.03 (−0.39, 0.44), 0.90 Principle component 8 22,133 0.28 (−0.19, 0.75), 0.22 0.03 (−0.18, 0.25), 0.75 0.09 (−0.23, 0.41), 0.58 PSA (ng/mL) 5012 −0.08 (−1.79, 1.63), 0.86 −0.16 (−0.56, 0.23), 0.22 0.17 (−0.42, 0.76), 0.35 Age (years) 18,962 −0.68 (−3.80, 2.43), 0.65 −0.33 (−2.24, 1.58), 0.73 −2.76 (−5.52. −0.0003), 0.05 OR family history per unit increase in genetic risk score (95% CI), P value Family history210,955 0.50 (0.21, 1.19), 0.12 1.62 (1.02, 2.58), 0.04 0.60 (0.25, 1.42), 0.24 Studies with more than 20% missing data were excluded from each analysis. Linear regression models take clustering by substudy into account. LDL, lowdensity lipoprotein; HDL, highdensity lipoprotein; TG, triglyceride; OR, odds ratio. 1Higher LDL/TG scores reflect increasing circulating LDL/TG, higher HDL scores reflect lower circulating HDL. 2Family history of prostate cancer (in father or brother): compares Yes versus No (logistic regression). 1132 © 2016 The Authors. Cancer Medicine published by John Wiley & Sons Ltd. C. J. Bull et al. Blood Lipids and Prostate Cancer There was little evidence of an association with prostate cancer stage (OR, advanced vs. localized: 0.97; 95% CI: 0.92, 1.02; P = 0.26) or grade (OR, high vs. low: 1.03; 95% CI: 0.98, 1.07; P = 0.21) (Table 5). Discussion We did not find evidence to suggest an association between circulating LDL, HDL, or TG (proxied by genetic risk scores) and overall prostate cancer. Although power to investigate advanced/highgrade cancer is limited (as evidenced by wide confidence intervals), our results may indicate a potential role for LDL and TG in prostate cancer progression. These findings are clinically important, as they highlight a distinction between indolent disease and more aggressive cancers. Variation in HMGCR has been exploited previously to demonstrate the implications of statin treatment on type 2 diabetes and bodyweight [24]. Our results from an analysis using a variant in HMGCR (rs12916T) to proxy statin intervention suggest that statins may hold potential in prostate cancer prevention, but our inference is tentative and requires further investigation in larger sample sizes. Associations observed between the LDL and HDL genetic risk scores and family history findings may be an artifact of multiple testing. However, these findings warrant further investigation with larger numbers to increase the precision of the point estimates. Metaanalyses of observational studies and randomized controlled trials present heterogeneous findings for associations of cholesterol with prostate cancer, making it difficult to conclude whether cholesterol plays a role in prostate cancer. The most recent and comprehensive assessment of observational data did not show any association between HDL and LDL and prostate cancer, but could not draw definitive conclusions on highgrade prostate cancer due to limited data [5]. Our Mendelian randomization approach has several advantages over conventional observational epidemiology: it eliminates the problem of reverse causality, as prostate cancer status cannot alter Table 4. Caseonly analysis: weighted genetic risk scores and prostate cancer stage and grade (PRACTICAL consortium). Outcome Localized/low grade (n) Advanced/high grade (n) OR195% CI P value LDL score Advanced versus localized 13,707 4301 0.91 0.51, 1.64 0.77 High grade versus low grade 9237 8515 1.50 0.92, 2.46 0.11 HDL score Advanced versus localized 13,707 4301 1.02 0.74, 1.39 0.92 High grade versus low grade 9237 8515 1.03 0.79, 1.34 0.82 TG score Advanced versus localized 13,707 4301 1.68 0.95, 3.00 0.08 High grade versus low grade 9237 8515 0.93 0.57, 1.52 0.77 Advanced (T3 to T4 or SEER staging regional or distant), localized (T1 to T2 or SEER staging localized), high grade (Gleason ≥7), low grade (Gleason ≤6). LDL, lowdensity lipoprotein; HDL, highdensity lipoprotein; TG, triglyceride; OR, odds ratio. 1Per unit increase in genetic risk score (SD trait), adjusted for top eight principal components (higher LDL/TG scores reflect increasing circulating LDL/ TG, higher HDL scores reflect lower circulating HDL). Figure 2. Meta-analysis OR prostate cancer per unit increase in genetic risk scores (SD trait). (A) Low-density lipoprotein (LDL): OR 1.24 (95% CI: 0.90, 1.69), P = 0.18. (B) High-density lipoprotein (HDL): OR 0.99 (95% CI: 0.84, 1.17), P = 0.90. (C) Triglyceride (TG): 1.09 (95% CI: 0.80, 1.50), P = 0.57. Cases: 22,249; controls: 22,133. Adjusted for top eight principle components. ABC 1133 © 2016 The Authors. Cancer Medicine published by John Wiley & Sons Ltd. Blood Lipids and Prostate Cancer C. J. Bull et al. one’s germline genetic makeup; genetic risk scores represent an individual’s exposure to lipid traits over their lifetime, reducing biological and technical sources of measurement error that arise from oneoff blood sampling at one point in a person’s life; and confounding by behavioral, lifestyle, and other related intermediate traits should be minimized as individuals are effectively randomly allocated to a low or high level of exposure based on their genotype, randomly generated at conception (Mendel’s second law of independent assortment) [27]. However, Mendelian randomization is susceptible to genetic confounding if the SNPs used as instruments for the trait of interest have effects on other phenotypes besides the specific lipid of interest, and it is these other phenotypes which lead to prostate cancer (horizontal pleiotropy) [16]. There is also potential for confounding due to population stratification; however, as we have adjusted for principle components in our regression models, this should be minimized. We cannot be sure that the LDL score is exclusively associated with LDL as validation of the score in ALSPAC revealed possible pleiotropy with circulating HDL. However, the HDL score which is strongly associated with HDL, and not LDL or TG in ALSPAC, was not associated with prostate cancer; therefore, the weak association noted between the LDL score and highgrade prostate cancer is likely an LDL, not an HDL effect. As F statistics for the genetic risk scores relate to ALSPAC and not PRACTICAL (the dataset for the outcome), conventional thresholds, such as F > 10 are not relevant for this study. The dataset for the outcome is far larger than the dataset in which we tested the genetic risk score–trait association; therefore, it may be that the genetic risk scores are in fact stronger instruments for circulating lipid traits in PRACTICAL than we estimated. Figure 3. Cases versus controls. Metaanalysis OR prostate cancer per rs12916T allele. OR 0.97 (95% CI: 0.94, 1.00), P = 0.03. Cases: 22,733; controls: 23,050. Adjusted for top eight principle components. Table 5. Stratified analysis by cancer stage and grade. OR per rs12916T allele. Adjusted for top eight principle components. Outcome Localized/low grade (n) Advanced/high grade (n) OR 95% CI P value rs12916T allele Advanced versus localized 13,707 4301 0.97 0.92, 1.02 0.26 High grade versus low grade 10,038 8543 1.03 0.98, 1.07 0.21 Advanced (T3 to T4 or SEER staging regional or distant), localized (T1 to T2 or SEER staging localized), high grade (Gleason ≥7), low grade (Gleason ≤6). OR, odds ratio.