Circulating Metabolic Biomarkers of Screen-Detected Prostate Cancer in the ProtecT Study

Prostate cancer is the most frequently diagnosed malignancy among men worldwide (1). Despite huge geographical variation in incidence and mortality (suggestive of environmental causes), the only established risk factors are age, race, and family history (2), while adiposity is associated with advanced prostate cancer (3). Metabolic dysregulation is a hallmark of carcinogenesis (4), and various circulating metabolites have been associated with both prostate cancer presence and aggressiveness (2, 5–15). However, there are limitations with the existing evidence-base. First, the metabolites detected as being associated with prostate cancer have varied between studies, with no clear pattern of associations emerging. Second, most previous studies have been small, having fewer than 100 cases (Supplementary Table S1 in the Supplementary Data; ref. 5) and hence liable to the play of chance, yielding both false-positive and false-negative findings. Third, the nature of any observed associations must be interpreted cautiously, because epidemiologic studies are highly susceptible to various biases that preclude causal inference (16, 17). For instance, metabolites may be associated with other factors that affect prostate cancer development (confounding), or the presence of prostate cancer may alter metabolites (reverse causation). Mendelian randomization (MR) is a method for appraising causality that uses genetic instrumental variables to proxy for exposures that can be otherwise confounded or subject to reverse causation. Such an approach may be used to distinguish causally relevant intervention targets from biomarkers that are noncausally associated with prostate cancer. The latter may nevertheless be of potential value in risk prediction (e.g., if the biomarker acts as a strong proxy for other factors leading to disease) or disease detection (e.g., if early disease leads to alterations in the circulating metabolome).

We undertook a case–control study within the Prostate Testing for Cancer and Treatment (ProtecT) trial (ISRCTN20141297; refs. 18–20). During recruitment to ProtecT, 228,966 men ages 50–69 years at 337 general practices in nine UK centers (Birmingham, Bristol, Cambridge, Cardiff, Edinburgh, Leeds, Leicester, Newcastle, Sheffield) were invited between 2001 and 2009 to attend a clinic for consideration of a prostate-specific antigen (PSA) test. Overall, 100,444 men attended the clinic where a PSA test was offered to 82,429 men deemed eligible to enroll into the ProtecT treatment trial, after a fully informed, 30-minute discussion with a nurse. Men with PSA levels ≥3 ng/mL were offered a 10-core diagnostic biopsy. Tumors were histologically confirmed, assigned a Gleason score by uropathologists, and clinically staged using the TNM classification (21). Men with a PSA <3 ng/mL or a raised PSA (≥3 ng/mL) but a negative biopsy without development of prostate cancer during the follow-up protocol were eligible as controls. Controls were randomly selected from the same five-year age-band (age at PSA test) and GP/family practice, as cases (22).

In this analysis, in concert with the guidelines of the National Institute for Health and Care Excellence (NICE; ref. 23), men with stage T3 or T4 tumors (“localized advanced”), Gleason score ≥8, or with a PSA level at diagnosis >20 ng/mL, were classified as “high-risk.” Men with stage T1 or T2 tumors, Gleason score <8, or with a PSA level at diagnosis ≤20 ng/mL were classified as “low-risk.” Participants in the present analysis consisted of those with self-reported (white) European ancestry.

All men provided written informed consent prior to inclusion into ProtecT. The Trent Multicentre Research Ethics Committee (MREC) approved ProtecT (MREC/01/4/025) and the linked ProMPT study, which collected biological material (MREC/01/4/061), including serum used for the current study of metabolites.

Two hundred and twenty-seven quantified metabolic traits (henceforth “metabolites”) were obtained per sample of serum using a proton nuclear magnetic resonance (NMR) spectroscopy-based metabolomics platform (Nightingale Health, Helsinki, Finland). Details of the methodology have been described elsewhere (24). Briefly, 100-μL serum was mixed with sodium phosphate buffer and transferred to NMR tubes using an eight-channel, Varispan Janus liquid handling robot (PerkinElmer). Two 1D NMR spectra were acquired using a 500 MHz Bruker Avance III HD spectrometer and analyzed bioinformatically for absolute quantification of lipoprotein subclasses, their particle concentrations and composition, lipoprotein particle size, apolipoprotein A-I and B, multiple cholesterol and triglyceride measures, albumin, various fatty acids, as well as numerous low-molecular-weight metabolites covering amino acids (including branched-chained and aromatic), glycolysis-related measures, and ketone bodies. The method has been widely used in epidemiologic research and recently reviewed in refs. 24, 25.

All analyses were performed using R (version 3.4.1). A total of 2,291 men with screen-detected prostate cancer (348 high-risk; 1,939 low-risk) and 2,661 controls had NMR metabolites measured in ProtecT. The distribution of baseline characteristics in cases versus controls was compared using Wilcoxon rank-sum tests for continuous variables and a χ² statistic for categorical variables. Multiple imputation using the “mice” R package, and based on a subset of 78 metabolites chosen at random (given imputation constraints), was used to impute family history of prostate cancer, unknown for 11% of participants. Family history and age were selected as covariates in multivariable models of prostate cancer risk, as those factors are strongly associated with prostate cancer and are potential confounders of the exposure–outcome relationship. We also adjusted for the primary care center where patients were registered. Metabolite trait concentrations/ratios were log-transformed and then scaled to SD scores to allow direct comparison of the magnitude of the effect of traits with different units on prostate cancer. A dictionary of metabolic traits with units before standardization is available in Supplementary Table S2.

Multivariable logistic regression was performed to compare the odds of total prostate cancer (versus controls) per log-transformed, then SD-scaled metabolite concentration, such that each metabolite has a standard deviation of one. As a sensitivity analysis, we also examined the odds of prostate cancer by high- and low-risk case status and performed tests of the differences between ORs [took the absolute difference between the ORs (δ); calculated the standard error (SE) for δ using the SEs from each comparison set, such that |$SE_1^2$| and |$SE_2^2$| refer, respectively, to the SEs of the first comparison and second comparison sets, |$\sqrt {SE_1^2 + SE_2^2} $|⁠; calculated z scores, |$\frac{{\rm{\delta }}}{{SE( {\rm{\delta }} )}}$|⁠; and calculated p-values for the z scores) for the following comparisons: differences in odds ratios for high-risk results and total results, the low-risk results and total results, and high-risk and low-risk results. In addition (also as a sensitivity analysis) we examined the correlation between the metabolites and PSA, given that our population of participants was screen-(PSA) detected.

To account for multiple testing and the correlation between the metabolic measures, principal component analysis was carried out on z-scored metabolic trait data (26). We calculated that the first 37 principal components explained >99% of the variance in the data and set our statistical threshold to P < 0.05/37 ( = 0.0014), equivalent to P < 0.05 after adjusting for multiple testing.

To assess causality, we used MR, a causal analysis method which exploits the random assortment of alleles in an instrumental variable (IV) framework to address confounding and reverse causation that preclude causal inference in observational studies (27, 28). Germline genetic variants associated with each metabolite of interest can serve as proxies (IVs) for those metabolites in models examining the causal effects of metabolic traits on prostate cancer, if a number of assumptions are met: (i) the IVs (genetic variants) are robustly associated with metabolites; (ii) the IVs are independent of confounders of the metabolites and prostate cancer; and iii) the IVs are not pleiotropically associated with the prostate cancer; i.e. they are associated with prostate cancer only through the metabolites they are instrumenting and not associated with prostate cancer through other exposures (29).

From the literature, we know there are strong associations between single-nucleotide polymorphisms (SNPs) and metabolite levels (30–33); therefore, these SNPs can serve as instruments in Mendelian randomization analyses (34–36). For instance, the median proportion of variance explained for metabolite associations in Kettunen and colleagues (2016) was 5% and ranged from 0.2% for acetoacetate to 12.5% for glycine (33). To implement MR, we identified independent (those not in linkage disequilibrium; r² < 0.01) SNPs that were robustly associated at genome-wide signficance (i.e. P < 5 × 10⁻⁸) with metabolites in the Kettunen and colleagues (2016) genome-wide association study (GWAS) of 123 circulating metabolites in 24,925 participants from 14 European cohorts (33). These SNPs were chosen as IVs for our metabolites. We could not instrument 113 of the 227 NMR-quantified metabolic traits; 65 of these traits were ratio measures not included in the GWAS and 48 were other types of traits that had no genetic proxy.

To leverage power from large samples, we performed two-sample MR (27, 37, 38), whereby we obtained summary data on the effects of the SNPs that acted as genetic instruments for each metabolite on prostate cancer from a separate data source, the Prostate Cancer Association Group to Investigate Cancer Associated Alterations in the Genome (PRACTICAL) consortium. This consortium involved 52 prostate cancer case–control studies on which genome-wide association studies (GWAS) had been conducted (39–44). The current MR analysis was based on 44,825 prostate cancer cases and 27,904 controls from within 48 of the 52 PRACTICAL cohorts of men with European ancestry.

To implement MR, we undertook the following process for each of the SNPs identified as a proxy for each metabolite: we extracted the effect and noneffect allele, and the log OR and standard error per copy of the effect allele from the PRACTICAL GWAS. We used these data to construct our genetic instruments for our circulating metabolic traits and calculated the log OR for prostate cancer per SD increase in metabolic measure levels using the two-sample MR Maximum likelihood estimator (ref. 45; Supplementary Table S3 contains the characteristics of the genetic variants associated with metabolites that were selected as instruments.) In computing the effect estimates based on MR, the effect estimates for each SNP-prostate cancer association were meta-analyzed.

Two-sample MR analyses were performed in R using the TwoSampleMR package (46).

Table 1 compares the distribution of selected characteristics in cases versus controls. The median age (63 years) and body mass index (BMI, 27 kg/m²) were the same, but more cases than controls had a family history of prostate cancer (8% vs. 5%; P < 0.001).

Thirty-five metabolites were associated with odds of prostate cancer at P < 0.0014 (Table 2; Fig. 1; Supplementary Table S4). The following increased the odds of prostate cancer: (i) lipids and lipoprotein particle concentration, specifically, total lipids (TL) in small high-density lipoprotein (HDL) and concentration of small HDL particles; (ii) total cholesterol (TC) and TC compositional ratios, namely, TC in HDL3, TC in medium low-density lipoprotein (LDL), small HDL, and small LDL; and TC:TL ratios in medium LDL, small HDL, small LDL, and small very low-density lipoprotein (VLDL); (iii) cholesterol esters (CE) and CE compositional ratios, specifically: CE in medium LDL, small HDL, and small LDL; CE-to-TL ratios in medium LDL, CE:TL ratios in small HDL and small LDL; (iv) free cholesterol (FC) and a FC compositional ratios, namely, FC in IDL, large LDL, and medium HDL, and FC-to-TL ratio in medium HDL; (v) phospholipids (PL) and PL compositional ratios, including: PL in intermediate-density lipoprotein (IDL) and very small VLDL, and PL:TL ratios in medium VLDL and very small VLDL; (vi) and the protein albumin; (vii) the ratio of omega-6 fatty acids (FA) to total FA.

The following decreased the odds of total prostate cancer: VLDL particle size, PL-to-TL ratios in medium LDL; triglycerides (TG)-to-total lipid ratios in small and medium VLDL; amino acids (AA), specifically, the branched chain AA, isoleucine, leucine, and valine, and the aromatic AA tyrosine; and saturated FA-to-total FA.

In the sensitivity analysis of the effect of metabolic traits on high-risk prostate cancer versus controls, albumin was associated with high-risk case status [OR 1.12; 95% confidence interval (CI) 1.08–1.36; P < 0.0014]; 138 (61%) had ORs reversed from those in the combined (total, case–control) analysis; and, although 53 ORs were statistically different from those in the total analysis (P value threshold < 0.05), none of the differences survived multiple comparisons (P < 0.05/227; 0.0002; Supplementary Tables S5 and S6). Consistent with these results, in the comparison of high-versus low-risk ORs, 78 metabolites had ORs that were statistically different at the < 0.05 threshold and two at the multiple-testing threshold (P <0.0002: TC:TL in small HDL and PL:TL in small HDL); 63% of metabolites had directionally reversed ORs. Notably, among the 35 top metabolites in the total analysis, eight were included in the set of those with statistically different ORs in the high- versus low-risk comparison (P value for multiple testing set to 0.05/35 = 0.0014; Supplementary Tables S7 and S8). Conversely, the sensitivity analysis for the effect of metabolites on low-risk prostate cancer (vs. controls) revealed patterns of association that mirrored the magnitude and direction of effects observed for the total analysis versus controls; only four (0.02%) metabolites in the low-risk analysis had ORs directionally reversed from those in the total analysis; and none of the ORs were statistically different from those in the total analysis (P < 0.05) (Supplementary Tables S9 and S10).

None of the metabolite–PSA correlations exceeded |0.06| (Supplementary Table S11).

Fourteen of the top 35 metabolites observationally associated with total prostate cancer were analyzable using MR. Of the 14 metabolites that were instrumental, none appear causal for prostate cancer risk (Fig. 2; Supplementary Tables S12 and S13).

We identified 35 potential biomarkers for prostate cancer. The majority of these were cholesterols, followed by glycerides and phospholipids. Steroid hormones, including androgens that drive prostate cancer, are derived from cholesterol (47), and high levels of cholesterol are required by rapidly proliferating cells (48). Hence, it is possible that our findings point to the underlying relationship between prostate cancer and androgens. Moreover, the observed effects appear to be driven solely by the low-risk cases, which were more abundant in our screen-detected cohort. The weak correlations between the metabolites and PSA suggest that our findings are not a byproduct of screening.

Fifteen of the top noninstrumented metabolites were ratios, which means that we were able to test the causal effects for the majority (70%) of our top metabolites that were not ratios (14/20).

A few recent studies have explored the relationship between serum metabolites and prostate cancer using metabolites detected from chromatography–mass spectrometry (2, 14, 15). In a pilot study, Mondul and colleagues (2014) compared 420 metabolic compounds in fasting serum collected prospectively from 74 clinically detected prostate cancer cases and 74 matched controls within the Alpha-Tocopherol, Beta-Carotene Cancer Prevention Study (ATBC) cohort. In their study, circulating 1-stearoylglycerol (1-SG) was inversely associated with prostate cancer (OR 0.34; 95% CI, 0.20–0.58; ref. 15). In this study, we did not quantify 1-SG. In their replication study, also within the ATBC cohort, Mondul and colleagues (2015) analyzed fasting serum collected prospectively for 626 metabolic compounds in 200 clinically detected cases and 200 matched controls (14). Notably, there was no overlap between the findings of this study and those of Mondul and colleagues (2015; 14).

Similarly, Huang and colleagues (2016) undertook an investigation of prostate cancer within the Prostate, Lung, Colorectal, and Ovarian Cancer Screening Trial (PLCO), for which they prospectively examined 695 known serum metabolites in 380 screen-detected cases and 380 matched controls. Their set of top metabolites differed from both from ours and the set observed in the ATBC (clinically detected) studies (2).

This study differed from both the two ATBC and the PLCO metabolome studies—studies perhaps the most comparable with ours—in that, instead of using chromatography–mass spectrometry to detect agnostically any measurable serum metabolites, we used a quantitative high-throughput NMR metabolomics platform with a prechosen set of metabolites that cover metabolic pathways for lipoprotein lipids and subclasses, fatty acids, amino acids, and glycolysis precursors. As such, some of the deviation between our findings and theirs are explained by this—we examined different sets of metabolites. Another difference is that the ATBC and PLCO studies were prospective and the observational portion of this study is cross-sectional.

We observed a family history of prostate cancer in 8% of cases, likely reflecting that they were identified in a screening versus a clinical setting.

Our analysis uses MR to interrogate whether some of our top findings (those with genetic instruments) appear causal. It is the first study of circulating metabolic traits and prostate cancer to do so. Moreover, it is the largest (more than 4× larger than the largest previous study; ref. 2) examination of the role of circulating metabolites in prostate cancer, and it yielded novel and promising associations with metabolic traits that may be useful clinically as biomarkers to better distinguish presence of disease and disease severity.

Our study has a few limitations. As the blood samples were collected at diagnosis for cases, we were unable to determine the direction of causality in our observational analyses. Likewise, there is potential, due to the way we selected our controls [men with PSA <3 ng/mL or a raised PSA (≥3 ng/mL) and negative biopsy] for there to be some misclassification of case status. Use of MR, at least for the instrumentable metabolites, allowed us, nonetheless, to appraise causality for a subset of our top findings, and we had at least 80% power to detect effect estimates within the range of those observed in our observational analysis for most metabolites. Another limitation is that there was a lack of specificity for many of the available genetic instruments, potentially biasing our causal analysis toward the null. Given this, while our MR found no evidence for causality, future MR analyses containing a larger number of specific genetic instruments for the metabolites are needed to strengthen causal assessment of the role of the metabolites we have detected as marking the presence of prostate cancer.

We identified 35 circulating metabolites associated with prostate cancer presence, but found no evidence of causality for those 14 testable with MR. Thus, the 14 metabolites tested with MR are unlikely to be mechanistically important in prostate cancer risk. We cannot speculate about the causality for those not tested with MR.

P. Wurtz is an employee of and has ownership interest in Nightingale Health Ltd. R.A. Eeles reports receiving speakers bureau honoraria from Janssen. N. Usmani reports receiving commercial research funding from Best Medical and is a consultant/advisory board member for Bayer, Amgen, and Astellas. P. Townsend reports receiving speakers bureau honoraria from MedLabs; is named as an inventor in patents unrelated to this manuscript; is a consultant/advisory board member for DeepMed, NED, and is a scientific advisor for Aptamer Group; and receives shares/license fees from Karus Therapeutics Ltd., Pentagon Biotechnology Ltd, and PRECignature Ltd. No potential conflicts of interest were disclosed by the other authors.

Conception and design: C. Adams, G.D. Smith, C.L. Relton, B.E. Henderson, F.R. Schumacher, A. Wolk, A.S. Kibel, M. Kogevinas, M.G. Dominguez, R.M. Martin

Development of methodology: C. Adams, D.L. Santos Ferreira, A. Razack, R.M. Martin

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): P. Wurtz, J.L. Donovan, F.C. Hamdy, D.E. Neal, R.A. Eeles, C.A. Haiman, Z. Kote-Jarai, F.R. Schumacher, K. Muir, S.I. Berndt, F. Wiklund, S.J. Chanock, S.M. Gapstur, C.M. Tangen, J. Batra, J.A. Clements, N. Pashayan, J. Schleutker, D. Albanes, A. Wolk, C.M.L. West, L.A. Mucci, G. Cancel-Tassin, S. Koutros, K.D. Sørensen, L. Maehle, R. Hamilton, S.A. Ingles, B.S Rosenstein, Y.-J. Lu, A. Vega, M. Kogevinas, J.Y. Park, J.L. Stanford, C. Cybulski, B.G. Nordestgaard, H. Brenner, E.M. John, M.R. Teixeira, S.L. Neuhausen, K. DeRuyck, A. Razack, L.F. Newcomb, D. Lessel, R.P. Kaneva, N. Usmani, F. Claessens, P. Townsend, M.G. Dominguez, M.J. Roobol, F. Menegaux, K.-T. Khaw, L.A. Cannon-Albright, H. Pandha, S.N. Thibodeau, R.M. Martin

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): C. Adams, R.C. Richmond, W. Spiller, V.Y. Tan, J. Zheng, P. Wurtz, A.A. Al Olama, D.V. Conti, S.J. Chanock, B.S Rosenstein, A.S. Kibel

Writing, review, and/or revision of the manuscript: C. Adams, R.C. Richmond, D.L. Santos Ferreira, V.Y. Tan, J.L. Donovan, D.E. Neal, J.A. Lane, G.D. Smith, C.L. Relton, R.A. Eeles, Z. Kote-Jarai, F.R. Schumacher, A.A. Al Olama, S. Benlloch, K. Muir, S.I. Berndt, F. Wiklund, S.M. Gapstur, V.L. Stevens, J. Batra, J.A. Clements, N. Pashayan, J. Schleutker, D. Albanes, A. Wolk, L.A. Mucci, G. Cancel-Tassin, S. Koutros, K.D. Sørensen, L. Maehle, R.C. Travis, R. Hamilton, S.A. Ingles, B.S Rosenstein, Y.-J. Lu, G.G. Giles, A.S. Kibel, A. Vega, M. Kogevinas, J.Y. Park, B.G. Nordestgaard, H. Brenner, C. Maier, J. Kim, E.M. John, M.R. Teixeira, S.L. Neuhausen, A. Razack, R.P. Kaneva, N. Usmani, P. Townsend, M.J. Roobol, K.-T. Khaw, H. Pandha, R.M. Martin

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): R.A. Eeles, A. Wolk, K.D. Sørensen, Y.-J. Lu, M.R. Teixeira, R.P. Kaneva, M.G. Dominguez, K.-T. Khaw, R.M. Martin

Study supervision: H. Gronberg, Y.-J. Lu, A.S. Kibel, M.G. Dominguez, R.M. Martin

Other (coding): D.L. Santos Ferreira

Other (PRACTICAL consortium member; contributed samples): K.L. Penney

Other (provided primary sample material for analysis): P. Townsend

Other (implementation of PRAGGA study in Galicia, Spain): M.G. Dominguez

This work was supported by Cancer Research UK (CRUK; C18281/A19169) and the Medical Research Council Integrative Epidemiology Unit at the University of Bristol, which is supported by the Medical Research Council (MRC) and the University of Bristol (MC_UU_12013/2; author C. Adams). R.C. Richmond is supported by CRUK (C18281/A19169) and MRC (MC_UU_12013/2). W. Spiller is supported by MRC (MC_UU_12013/1) and a Wellcome Trust studentship (108902/15/Z). V.Y. Tan is supported by CRUK (C18281/A19169). P. Wurtz is funded by the Academy of Finland (grant nos: 312476 and 312477). D.E. Neal is funded by (C11043/A4286, C18281/A8145, and C18281/A11326). J.A. Lane is supported by CRUK (C18281/A19169). C.L. Relton is supported by MRC (MC_UU_12013/2) and CRUK (C18281/A19169). R.A. Eeles is supported by CRUK (C5047/A17528). Z. Kote-Jarai is supported by CRUK (C5047/A17528). K.-T. Khaw with EPIC Norfolk is supported by grants from the MRC (MR/N003284/1; G1000143) and CRUK (14136). R.C. Travis is supported by CRUK (C8221/A19170), CRUK (14136 for EPIC-Norfolk and C570/A16491 for EPIC-Oxford), and the MRC (1000143 for EPIC-Norfolk and MR/M012190/1 for EPIC-Oxford. R.M. Martin is supported by CRUK (C18281/A19169). The National Institute for Health Research Biomedical Research Centre supported R.M. Martin, C.L. Relton, G.D. Smith. Genotyping of the OncoArray was funded by the US NIH [U19 CA 148537 for ELucidating Loci Involved in Prostate cancer SuscEptibility (ELLIPSE) project and X01HG007492 to the Center for Inherited Disease Research (CIDR) under contract number HHSN268201200008I]. Additional analytic support was provided by NIH NCI U01 CA188392 (principal investigator: F.R. Schumacher). The PRACTICAL consortium was supported by Cancer Research UK Grants C5047/A7357, C1287/A10118, C1287/A16563, C5047/A3354, C5047/A10692, C16913/A6135, European Commission's Seventh Framework Programme grant agreement no. 223175 (HEALTH-F2-2009-223175), and The NIH Cancer Post-Cancer GWAS initiative grant: no. 1 U19 CA 148537-01 (the GAME-ON initiative). We would also like to thank the following for funding support: The Institute of Cancer Research and The Everyman Campaign, The Prostate Cancer Research Foundation, Prostate Research Campaign UK (now Prostate Action), The Orchid Cancer Appeal, The National Cancer Research Network UK, The National Cancer Research Institute (NCRI) UK. We are grateful for support of NIHR funding to the NIHR Biomedical Research Centre at The Institute of Cancer Research and The Royal Marsden NHS Foundation Trust.

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.