Associations of dietary patterns with brain health from behavioral, neuroimaging, biochemical and genetic analyses

Food-liking reflects hedonic responses that drive food choices and intake. With abundant food options, people develop diverse dietary patterns that influence chronic disease risks and brain health. Prior studies link diet to cognition and psychiatric disorders, and neuroimaging shows diet-related brain structural and functional differences. However, inconsistencies in definitions of traditional dietary patterns (e.g., Western, Mediterranean, prudent, vegetarian) and limited sample sizes lead to mixed findings, especially for vegetarian diets. There is a need for a robust, data-driven classification of dietary patterns at population scale and an integrated assessment of their associations with mental health, cognition, biomarkers, brain imaging, and genetics. This study uses UK Biobank food-liking data to identify naturally developed dietary patterns and examine their associations with brain health using behavioral, biochemical, neuroimaging, and genetic analyses.

Evidence indicates diet is associated with cognitive function and mental health: higher simple carbohydrates and saturated fats relate to poorer cognition, while protein intake may enhance executive function and working memory. Unhealthy or Western-type diets are linked to depression, anxiety, bipolar disorder, stroke, sleep problems, and Alzheimer’s disease. Mediterranean-type diets are associated with preserved brain volume and cortical thickness and reduced neurodegenerative risk. Traditional dietary patterns (Western, Mediterranean, prudent, vegetarian/plant-based) have variable associations with mental health, with contradictory evidence on vegetarian diets (some show higher depression/anxiety, others the opposite or no effect). Mechanisms include gut–brain axis perturbations, inflammation, and oxidative stress; neuroimaging links dietary adherence to structural and functional brain differences. These mixed findings underscore the need for large-scale, standardized, data-driven dietary pattern classification to clarify associations with brain health.

Data: UK Biobank participants who completed the food-liking questionnaire; final sample n=181,990 (mean age 70.7 ± 7.7 years; 57.08% female). Food-liking: 140 food/beverage items grouped into 10 categories (alcohol, beverages, dairy, flavorings, fruits, fish, meat, snacks, starches, vegetables). Responses on a 9-point hedonic scale; missing values imputed using kNN (k=7). Subtype identification: z-score normalization; within-category PCA to reach ≥80% explained variance (robustness checked at 70% and 90%); hierarchical clustering (Euclidean distance; minimum variance algorithm) on 83 PCs. Cluster validity assessed via silhouette criterion (optimal not exceeding four). Four subtypes identified. Alignment of liking with consumption: compared subtype-wise averages of corresponding food frequency items showing congruence. Phenotypic analyses: One-way ANCOVA with covariates (age, sex when applicable, BMI, education, Townsend deprivation index; imaging additionally controlled for scanning site; GMV additionally for intracranial volume). Equality of variances checked with Levene’s tests. Multiple comparisons corrected (Bonferroni for mental health, cognition, biomarkers, PRS; FDR for neuroimaging). Mental health: 8 symptom domains from MHQ (n=118,616), items scaled to (0,1); higher indicates more symptoms except well-being. Cognitive tests: six tasks including fluid intelligence, trail making, symbol-digit substitution, pairs matching, reaction time, numeric memory (sample sizes up to n≈179,740). Blood/metabolic biomarkers: 30 blood biochemistry and 31 blood counts categorized; NMR-based 168 metabolic biomarkers; analysis sample n=42,665. Neuroimaging: GMV from T1 MRI (n=32,715) using CAT12; AAL2 atlas (94 regions). Diffusion MRI (n=31,195): FA and MD in 48 JHU ICBM-DTI-81 tracts; preprocessing with Eddy and FSL TBSS. Longitudinal risks: Cox proportional hazards models for 11 mental/neurological disorders (e.g., anxiety, depression, stroke), subtype 4 as reference; Schoenfeld residuals to test proportional hazards; stratification by age or BMI when needed; HRs with 95% CIs; FDR correction. Structural Equation Modeling (SEM): three models contrasting each subtype (1–3) versus subtype 4; latent variables for mental health, cognitive function, and brain MRI traits constructed from measures significantly differing in post hoc tests; confirmatory factor analysis; model fit via RMSEA; P values via Wald tests with FDR. Genetics: Case-control GWAS (PLINK 2.0) comparing each subtype (1–3) vs subtype 4; logistic regression adjusted for sex, age, BMI, top 10 ancestry PCs, batch; QC on UKB v3 imputed data; participants of recent British ancestry; 337,199 participants after QC; current study used genetic data for 181,551 individuals. Gene mapping and enrichment via FUMA (SNP2GENE, GENE2FUNC) with GTEx v8 tissue expression and GWAS Catalog enrichment (BH FDR). Polygenic Risk Scores (PRS): UKB optimized PRS for AD, PD, CVD, ischemic stroke, bipolar disorder, schizophrenia (n=176,465), plus PRSice-2 computed PRS for depression and suicide attempt (n=126,895) at p-threshold 0.05; analyses adjusted for PRS genetic PCs.

Dietary subtypes: Four data-driven subtypes identified among 181,990 participants with proportions: subtype 1 (starch-free/reduced-starch) 18.09%, subtype 2 (vegetarian) 5.54%, subtype 3 (high protein and low fiber) 19.39%, subtype 4 (balanced) 56.98%. Liking-consumption alignment: Subtype-wise relative scores for selected items showed close correspondence between food-liking and food-consumption. Mental health and cognition (ANCOVA, Bonferroni-corrected): Significant main effects across subtypes for 7 mental health measures: anxiety (F=41.5, P=8.9×10^-22), depression (F=71.4, P=3.9×10^-46), mental distress (F=62.1, P=4.0×10^-40), psychotic experience (F=17.4, P=2.6×10^-11), self-harm (F=116.8, P=1.6×10^-75), trauma (F=155.4, P=1.6×10^-100), well-being (F=256.8, P=3.8×10^-166). Subtype 4 had the lowest symptom scores and highest well-being. Subtypes 2 and 3 tended to have higher symptom scores and lower well-being. Cognitive functions: fluid intelligence (F=15.0, P=8.9×10^-10), pairs matching (F=6.6), reaction time (F=20.1, P=5.2×10^-13), symbol-digit substitution (F=18.6, P=4.9×10^-12). Subtype 4 had shortest reaction time; subtype 3 had highest symbol-digit correct and second-lowest reaction time. Longitudinal risks (Cox PH, subtype 4 reference, FDR-corrected): Anxiety risk higher in subtypes 1 (HR 1.09, 95% CI 1.00–1.17, adj P=0.03), 2 (HR 1.26, 1.14–1.41, adj P=6.2×10^-5), 3 (HR 1.23, 1.15–1.31, adj P=3.2×10^-6). Depression risk higher in subtypes 2 (HR 1.18, 1.04–1.33, adj P=0.03) and 3 (HR 1.22, 1.13–1.30, adj P=3.8×10^-4). Stroke risk higher in subtypes 1 (HR 1.13, 1.03–1.24, adj P=0.03) and 3 (HR 1.21, 1.11–1.31, adj P=2.3×10^-5). Eating disorder risk elevated in subtypes 1 (HR 1.86, 1.45–2.38, adj P=9.1×10^-10), 2 (HR 2.68, 2.00–3.58, adj P=9.7×10^-10), 3 (HR 1.96, 1.48–2.59, adj P=2.0×10^-5). Blood and metabolic biomarkers: 167/229 biomarkers differed by subtype (Bonferroni P<0.05/229). Strongest differences in fatty acids (e.g., DHA F=312.5, P=1.1×10^-100; omega-3 F=232.4, P=1.4×10^-149), amino acids (glycine F=122.1, P=1.0×10^-80), renal function (urea F=114.8, P=5.2×10^-74), HDL-related lipid measures. Post hoc vs subtype 4: subtype 3 differed on 127 biomarkers (mostly lower; e.g., DHA t=−25.7, d=−0.3, P=3.7×10^-144; omega-3 t=−21.3, d=−0.3, P=6.4×10^-100; HDL cholesterol t=−14.3, d=−0.2, P=1.7×10^-40). Subtype 1 differed on 49 biomarkers (many higher; e.g., degree of unsaturation t=8.7, d=0.1, P=4.0×10^-15), and some lower phospholipids and saturated fats. Subtype 2 differed on 72 biomarkers (e.g., urea t=−18.7, d=−0.4, P=2.5×10^-77; glycine t=18.9, d=0.4, P=3.1×10^-77; omega-3 t=−14.4, d=−0.3, P=8.7×10^-47). Neuroimaging: GMV: 23/94 AAL2 regions differed by subtype (FDR). Subtype 3 vs 4: 16 regions differed, 11 lower in subtype 3 (e.g., postcentral gyrus, parahippocampal gyrus, inferior parietal gyrus). Subtype 1 vs 4: 7 regions (putamen, caudate, pallidum, paracentral lobule). Subtype 2 vs 4: 4 regions (thalamus, precuneus, paracentral lobule) higher in subtype 2. White matter: FA differed in 8 ROIs (FDR). Subtype 3 vs 4: 7 ROIs lower FA (e.g., medial lemniscus, uncinate fasciculus, external capsule). Subtype 1 vs 4: cingulum hippocampus higher FA. Subtype 2 vs 4: uncinate fasciculus higher FA. MD differed in 11 ROIs. Subtype 3 vs 4: all 11 higher MD (e.g., external capsule, anterior limb internal capsule, superior fronto-occipital fasciculus, hippocampal cingulum). Subtype 1 vs 4: 3 ROIs higher MD. Subtype 2 vs 4: cerebral peduncle higher MD. PRS differences (ANCOVA Bonferroni P<0.05/8): significant for AD (F=8.2), ischemic stroke (F=8.3), PD (F=6.7), CVD (F=6.0), bipolar disorder (F=34.1), schizophrenia (F=72.4), depression (F=11.5), suicide attempt (F=6.4). Subtype 2 showed higher genetic predisposition to multiple mental disorders; subtype 3 higher ischemic stroke PRS; subtype 4 lower PRSs overall. SEM: Subtype 3 vs 4 model showed food preference associated with mental health (β=0.052, Padj=5.5×10^-6), brain MRI traits (β=−0.037, Padj=4.6×10^-4), and cognition (β=0.077, Padj=3.5×10^-3). Brain MRI predicted mental health (β=−0.058, Padj=1.5×10^-7) and cognition (β=0.098, Padj=9.2×10^-11); mental health predicted cognition (β=−0.117, Padj=1.1×10^-12). Similar directional associations for subtypes 1 and 2 vs 4. Genetics (GWAS): 1,266 genome-wide significant SNPs (P<5×10^-8) distinguished subtype 3 from subtype 4, concentrated on chromosomes 2, 3, 13, 17 (e.g., rs36164224 OR=1.07, P=8.6×10^-10; rs62250502 OR=0.93, P=1.3×10^-11; rs3124402 OR=0.92, P=2.5×10^-11; rs2532387 OR=1.07, P=1.2×10^-10). Two SNPs differentiated subtype 1 vs 4; none for subtype 2 vs 4. Mapping yielded 16 genes (e.g., MAPT, MVB12B, NSF, CADM2, CRHR1, MEIS1, PLEKHM1, KANSL1) with high expression in multiple brain regions (ACC, frontal cortex, amygdala, hippocampus, cerebellum) and enrichment for traits related to mental disorders and cognition (e.g., alcohol use disorder Padj=1.8×10^-10; PD Padj=5.5×10^-8; AD in APOE ε4 carriers Padj=4.3×10^-7; reaction time Padj=3.5×10^-8; brain morphology Padj=3.5×10^-8).

Using a data-driven approach on a large UK cohort, the study identified four naturally occurring dietary patterns that relate systematically to mental health, cognition, biomarkers, brain structure, and genetics. Individuals with a balanced dietary pattern exhibited better mental health, faster processing speed, and more favorable biomarker profiles, gray matter volumes, and white matter integrity. The high protein/low fiber subtype showed less favorable brain structure (lower GMV, lower FA, higher MD), worse mental health risks, and biomarker profiles suggestive of inflammation and lipid alterations, aligning with mechanisms involving gut–brain axis and inflammation. Vegetarian subtype showed higher mental health symptom scores and lower well-being but also higher genetic susceptibility (PRSs) to several psychiatric and neurodegenerative disorders, suggesting genetic factors may partly underlie observed phenotypes. SEM indicated directional associations: food preference relates to brain MRI traits and mental health, which in turn influence cognitive function, highlighting a pathway from diet to brain structure to mental health and cognition. GWAS and gene expression analyses implicated brain-expressed genes (e.g., MAPT, CRHR1, KANSL1) enriched for neurological and psychiatric traits, providing biological substrates that may modulate diet–brain health links.

This large-scale, data-driven analysis delineated four dietary subtypes from food-liking data and demonstrated comprehensive associations with mental health, cognition, biomarkers, neuroimaging, and genetics. A balanced dietary pattern was linked to better brain health profiles and lower mental disorder risks. Genetic findings highlight brain-expressed genes and pathways related to cognition and mental disorders differentiating subtypes, especially high protein/low fiber versus balanced. The study offers a robust classification framework for dietary patterns and underscores the potential of food preferences as markers for targeting interventions to promote brain health. Future research should test causality, include detailed nutrient measures (e.g., tryptophan, omega-3/6 pathways), validate mental health assessments with comprehensive scales, examine generalizability beyond UK Biobank, and extend analyses to younger populations to evaluate lifespan effects.

Dietary patterns were derived from food-liking rather than direct consumption data, although liking-consumption alignment was observed. UK Biobank’s healthy volunteer bias may limit generalizability. Responders to the food-liking questionnaire differed demographically from nonresponders. Key biochemical contributors to serotonin synthesis (e.g., tryptophan) were not measured, and detailed involvement of omega-3/omega-6 in patterns was not captured. Mental health measures were simplified composites rather than full clinical scales. Causal inferences cannot be drawn from observational analyses. Findings are in primarily older adults; applicability to adolescents and middle-aged cohorts requires further study.