Effects of diets on risks of cancer and the mediating role of metabolites

Cancer imposes a major global health burden, accounting for millions of new cases and deaths annually. Nutritional factors are considered to have probable causal or protective associations with cancer risk. The Mediterranean diet (MedDiet) and the MIND diet (a MedDiet- and DASH-inspired pattern emphasizing plant-based foods, green leafy vegetables, and berries while limiting animal and high-saturated-fat foods) have been reported to protect against adverse health outcomes. However, evidence on their relationships with overall and site-specific cancer risks is mixed, and underlying mechanisms remain unclear. Prior prospective studies in specific populations reported null associations for MedDiet with overall cancer, while recent meta-analyses suggest risk reductions. Evidence for MIND diet and cancer risk is limited and largely from case-control studies. A hypothesized mechanism is that diet influences circulating and tumor microenvironment metabolites, which in turn modulate cancer cell metabolism. NMR-based metabolomics offers quantitative, reproducible profiling of metabolites. This study aims to evaluate associations between adherence to MedDiet and MIND diet with overall and 22 site-specific cancer risks in the UK Biobank, identify metabolites associated with overall cancer risk from 168 NMR metabolites, and assess whether these metabolites mediate the diet–cancer associations.

Prior research offers varying conclusions about diet–cancer relationships. Prospective cohorts in France and Sweden reported no significant association between MedDiet adherence and overall cancer, whereas an updated meta-analysis of 117 studies indicated reduced overall cancer risk with greater MedDiet adherence. For the MIND diet, evidence has primarily come from case-control studies suggesting inverse associations with specific cancers such as breast cancer and glioma, with a paucity of prospective data. Studies in different countries and populations limit generalizability due to varying dietary behaviors. Mechanistically, metabolic reprogramming is a cancer hallmark; diet-driven changes in metabolite levels may alter tumor microenvironment and cancer cell metabolism. NMR metabolomics can capture such dietary influences on systemic metabolism. Thus, more comprehensive, prospective, and mechanistic evaluations are warranted.

Design and data source: Prospective analyses using the UK Biobank, which enrolled >500,000 adults aged 37–73 years (2006–2010) across the UK. Cancer incidence follow-up through national registries and records up to Nov 30, 2022 (England), Aug 31, 2022 (Scotland), and May 31, 2022 (Wales). Ethics approval obtained; informed consent provided. Study populations: Three linked analyses with differing inclusion/exclusion criteria due to data availability: - Diet–cancer associations: n=187,485 after excluding prevalent cancer, loss to follow-up, missing diet before cancer onset, or >80% missing covariates. - Metabolite selection for overall cancer risk: n=202,303 after excluding prevalent cancer, loss to follow-up, lipid-lowering drug use before blood draw, missing baseline metabolites, or >80% missing confounders. - Mediation analysis (shared subset having both diet and metabolite data): n=85,669. Exposure assessment: - Dietary intake via repeated web-based 24-hour recalls (Oxford WebQ): baseline (Apr 2009–Sep 2010) plus up to 4 follow-ups every 3–4 months (Feb 2011–Jun 2012). Intake recorded for 200 foods and 32 beverages. Average intake across assessments used to derive adherence scores. - MEDAS score (PREDIMED-based; binary scoring across 14 components; olive oil and sofrito not available, so maximum score=12). - MIND score (15 components: 10 beneficial, 5 adverse; concordance scores 0/0.5/1; olive oil unavailable, maximum score=14). Metabolomics: - High-throughput NMR-metabolomics of baseline EDTA plasma in ~250,000 participants, yielding absolute concentrations for 168 metabolites (fatty acids, glycolysis-related metabolites, ketone bodies, amino acids, lipoproteins, etc.). Covariates: - Demographics and behaviors: age, sex, full-time education, household income, Townsend deprivation index, family history of cancer, smoking status, alcohol drinking status, physical activity, sleep duration, BMI, waist circumference/WHR, systolic blood pressure, and dietary energy intake. Data from questionnaires and physical measures; dietary energy from Oxford WebQ. Data handling and transformations: - Multiple imputation for covariates with <20% missingness. Metabolite values assumed missing due to detection limits and imputed as half the minimum detectable value. Metabolites ln(x+1)-transformed and standardized to z-scores. Statistical analyses: - Diet–cancer associations: Cox proportional hazards models with age as timescale; diet scores analyzed as continuous and tertiles (T1 reference). Models: (1) adjusted for dietary energy; (2) further adjusted for sex, education, income, Townsend index, family history of cancer; (3) additionally adjusted for smoking, alcohol, physical activity, sleep duration, BMI, WHR, SBP. Results as HRs and 95% CIs. Stratified analyses and interaction tests by sex, education, family history, smoking, alcohol, and BMI via likelihood ratio tests. - Metabolite selection for overall cancer: Random split into training (n=121,382) and testing (n=90,921). In training, Cox models for each metabolite with FDR control (Benjamini–Hochberg). Elastic net models (10-fold CV) to select metabolites among correlated features across cut-offs (0.25–1.00). Gradient boosting decision trees to assess variable importance. Metabolites identified by ENM and supported by feature importance considered final. - Mediation analyses: Cox-based multiple mediation using the mrm package to estimate natural direct and indirect effects of diets on overall cancer via identified metabolites. Reported proportional mediation (PM) and HRs for indirect effects, individually and jointly. Sensitivity analyses: - Alternative scoring (baseline-only diet; participants with ≥2 diet recalls; alternative Mediterranean adherence score), exclusion of individual diet components, leave-one-cancer-out for overall cancer definition, additional adjustment for smoking intensity and alcohol frequency, adjustment for macronutrients and fiber, lag analysis excluding cancers within 2 years of last diet assessment, exclusion of participants with poorer overall health ratings.

- Cohort and follow-up: Among 187,485 participants, 26,391 incident cancers occurred over a median 13.2 years (range 0.1–15.7). Those who developed cancer were older at baseline and more likely to have a family history of cancer, smoke, drink alcohol, and have higher BMI/WHR. - Overall cancer and diet adherence: - Higher MEDAS and MIND scores were associated with lower overall cancer risk across models. In model 3 (fully adjusted), tertile 3 vs tertile 1 HRs (95% CI): MEDAS 0.818 (0.789–0.849); MIND 0.808 (0.784–0.832), with significant trends. Similar inverse associations when modeled continuously. - Dietary components: Greater intake of vegetables, fruits, berries, nuts, whole grains, seafood, legumes, and wine associated with lower overall cancer risk; higher intake of butter/margarine/cream, cheese, red meat and products, fast/fried foods, pastries, and sweets associated with higher risk, aligning with MedDiet/MIND recommendations. - Site-specific cancers: - Significant inverse associations observed for multiple cancers. The study reports significant associations for 14 (MEDAS) and 13 (MIND) specific cancers overall. Both diets showed protective associations for colorectal, lung, breast, and prostate cancers at tertiles 2 and 3. Additional significant associations included head and neck, esophageal, malignant melanoma, kidney, thyroid, non-Hodgkin lymphoma, multiple myeloma, and leukemia (details in Supplementary Data 8). - Subgroup interactions: MIND score by sex for overall cancer (Pinteraction=0.018), stronger in males; multiple interactions for specific cancers, e.g., stomach (diet×sex), colorectal (diet×sex and diet×smoking), liver (MEDAS×BMI), pancreas (MIND×sex), lung (MIND×smoking), Hodgkin lymphoma (MEDAS×education), leukemia (MIND×education). - Metabolites associated with overall cancer (top 10): - Total lipids in VLDL: HR 0.972 (0.953–0.991) - Total choline: HR 0.924 (0.904–0.939) - Omega-3 fatty acids: HR 0.880 (0.864–0.897) - Tyrosine: HR 0.935 (0.918–0.952) - Glucose: HR 0.955 (0.938–0.972) - Citrate: HR 0.902 (0.885–0.918) - Creatinine: HR 0.976 (0.957–0.995) - Albumin: HR 1.051 (1.029–1.073) - Free cholesterol in IDL: HR 0.907 (0.889–0.924) - Total lipids in large HDL: HR 0.908 (0.888–0.928) Many of these metabolites showed consistent associations across multiple specific cancers. - Mediation analysis (n=85,669): - Joint mediation by identified metabolites explained 19.7% of the MEDAS–overall cancer association and 20.8% of the MIND–overall cancer association. - Individual mediators (illustrative estimates): omega-3 fatty acids showed the largest protective indirect effect; total choline and albumin partially masked effects (HR>1 for indirect path). For MEDAS, indirect HRs (examples): omega-3 0.972; total lipids in large HDL 0.975; free cholesterol in LDL/IDL ~0.991; total choline 1.011; albumin 1.007; total lipids in VLDL 1.011. Direct effects remained protective (e.g., MEDAS direct HR ~0.92; MIND direct HR ~0.80).

The study demonstrates that greater adherence to MedDiet and MIND diet patterns is associated with reduced risk of overall cancer and multiple site-specific cancers in a large, well-characterized cohort with long follow-up. These findings align with evidence that plant-forward dietary patterns rich in vegetables, fruits, legumes, whole grains, fish/seafood, and nuts, and low in high-saturated-fat animal products and ultra-processed foods, are protective against carcinogenesis. The consistency across subgroups and sensitivity analyses supports robustness, although some interactions indicate effect modification by sex, smoking, education, and BMI for certain cancers. Mechanistically, identified circulating metabolites—particularly omega-3 fatty acids, citrate, and lipoprotein-related measures—likely reflect diet-induced metabolic states that influence tumor biology. Mediation analyses suggest that a meaningful fraction (~20%) of the protective effect of these diets operates through changes in metabolite profiles, with omega-3 fatty acids showing the strongest mediating contribution. Some metabolites (e.g., albumin, total choline) exhibited masking effects, underscoring complex, potentially bidirectional metabolic pathways in carcinogenesis. Differences in protective strength between MedDiet and MIND for certain cancers (e.g., kidney, brain, thyroid) may derive from distinct dietary emphases (e.g., broader fruit and dairy emphasis in MedDiet vs. specific focus on leafy greens and berries in MIND). Overall, the integration of diet, metabolomics, and cancer outcomes advances understanding of diet–metabolism–cancer pathways.

In this large UK Biobank analysis, higher adherence to Mediterranean and MIND dietary patterns was associated with significantly lower risks of overall and multiple specific cancers. A set of 10 NMR-derived metabolites was associated with overall cancer risk, and together these metabolites mediated approximately one-fifth of the protective diet–cancer associations, with omega-3 fatty acids contributing prominently. These results support dietary guidance emphasizing plant-based foods, fish/seafood, and reduced intake of animal fats and ultra-processed foods for cancer prevention. Future research should include causal inference approaches (e.g., intervention studies, Mendelian randomization), repeated longitudinal metabolomics, detailed dietary component assessments (including olive oil), and exploration of effect modification across populations and cancer subtypes to refine precision nutrition strategies.

- Causality cannot be inferred from observational design; potential for residual confounding and reverse causation (e.g., preclinical disease influencing diet). - Temporal misalignment between last dietary assessments (up to 2012) and cancer follow-up (through 2022) could introduce exposure misclassification if diets changed. - Many participants (≈40.2%) had only a single dietary recall contributing to the average score, which may not reflect habitual intake; all dietary data were self-reported, introducing potential measurement error. - Lack of olive oil and sofrito information limited full MEDAS/MIND scoring; however, sensitivity analyses suggested minimal impact in this UK cohort. - Generalizability may be constrained by UK Biobank participation and population characteristics; site-specific analyses varied in power due to differing case numbers.