Large-scale GWAS of food liking reveals genetic determinants and genetic correlations with distinct neurophysiological traits

Dietary behavior profoundly impacts health, and in modern food environments individual choice increasingly shapes consumption. Understanding the biological drivers of food preference (liking) is important because liking reflects hedonic responses closely tied to biology and may guide targeted dietary interventions and the development of acceptable nutritious foods. Food-liking is influenced by genetics, biology, psychology, and environment; twin studies show moderate heritability (~50% in children, persisting into adulthood with non-shared environments). Prior genetic studies have mostly examined food consumption or candidate genes for specific tastes with mixed results, and genome-wide studies of liking have been limited in size/scope. This study asks: what are the genetic determinants of food-liking across many specific items; how are liking traits structured; and how do they relate genetically to consumption behaviors and neurophysiological traits? The authors perform a large-scale GWAS of food and beverage liking in UK Biobank, derive a hierarchical structure of liking dimensions, and assess genetic correlations with consumption, other complex traits, and brain morphology/connectivity.

Twin and family studies indicate food preferences are moderately heritable, with shared environment prominent in childhood and non-shared factors in adulthood. Previous GWAS primarily targeted food consumption behaviors, identifying numerous loci linked to dietary habits. Candidate gene studies on liking (e.g., TAS2R bitter receptors and coffee liking) yielded mixed and sometimes inconsistent findings. More recent genome-wide approaches have identified loci for liking of sweet foods and specific items (e.g., cilantro/coriander), but many studies were underpowered or focused on narrow taste domains. Thus, a comprehensive, well-powered genome-wide analysis of diverse food-liking traits has been lacking, and the broader genetic architecture and neurophysiological correlates of liking remained unclear.

Study population: UK Biobank participants (project 19655), aged 37–73 at recruitment (2006–2010), with written informed consent and ethical approvals. Inclusion: participants of European descent who completed a 2019 online food-liking questionnaire. Genotyping: UK Biobank or UK BiLEVE Axiom arrays; standard imputation, principal components, and QC as per UKB protocols. Food-liking phenotypes: Online questionnaire extended from prior instruments, 152 items total; the present study analyzed 139 food and beverage items (excluding non-food behaviors). Liking scored on a 9-point hedonic scale (1=Extremely dislike to 9=Extremely like), with options for never tried/prefer not to answer. Coffee and tea liking were measured with and without sugar; derived measures included maximum liking (sweetened vs unsweetened) and difference scores. GWAS: For each of 144 initial food-liking traits (later model comprised 119 items), raw scores were regressed on age, sex, first 10 genetic PCs, array type, and batch. Genetic relatedness accounted for using GRAMMAR+ residuals (fastGWA). Association testing used regscan under an additive model for SNPs with MAF>0.001. Study-wide significance was set at p<1.47×10−9, derived from 34 independent components explaining >95% genetic variance (5×10−8/34). Associations at p<5×10−8 were also considered if within a locus harboring a study-wide hit. Hierarchical factor analysis and model construction: Pairwise genetic correlations among items estimated via LD Score Regression (ldsc) using HapMap3 reference SNPs. Hierarchical clustering (Ward’s D2) defined correlated groups; factors were fit using GenomicSEM with model fit criteria CFI>0.9 and SRMR<0.1, removing poorly fitting items as needed. Factors with rg>0.9 were merged. Iteration yielded a multi-level hierarchical structure with up to four highest-order factors (e.g., F-Highly palatable, F-Low caloric, F-Acquired, and a minor F-Caffeinated Sweet Drinks group). SNP effects on latent factors were estimated as weighted linear combinations of SNP effects on items using GenomicSEM loadings (per Tsepilov et al. 2020) for computational efficiency. Liking vs consumption: Genetic correlations and SNP-heritability (LDSC) compared between liking and corresponding consumption traits using Pan-UKBB GWAS where matched or similar items existed. Genetic correlations with other traits: For the three main highest-order factors, genetic correlations were computed against a curated set of socio-economic, anthropometric, biochemistry, and behavior traits via LD Hub; results summarized across 31 representative traits. Locus definition and colocalization: SNPs with p<1×10−5 were grouped by proximity; loci were separated when adjacent SNPs were >250 kb apart; a locus required at least one SNP at p<1.47×10−9 to be significant. Overlapping loci were merged. HyPrColoc was applied to define sub-loci, evaluate regional colocalization probabilities, and identify likely causal SNPs per trait cluster. Replication and meta-analysis: Replication attempted in up to 26,154 individuals across 11 cohorts (mostly European ancestry): ALSPAC, INGI (CARL, VB, FVG), CROATIA (Korcula, Vis), NTR, Silk Road, TwinsUK, VIKING, etc. Meta-analysis focused on 54 overlapping traits with ≥10,000 samples. Results were rescaled to a 0–1 liking scale; summary-level QC via EasyQC; inverse-variance meta-analysis with METAL. Only 235 SNP–trait associations could be tested for replication due to trait availability. Direct vs mediated SNP effects: Using GenomicSEM, the effect of each sentinel SNP was fit simultaneously across nodes of the hierarchical model to distinguish direct from mediated associations. A two-step approach estimated SNP effects on latent traits and residual item effects, using a dummy latent variable highly correlated (0.99) with the true latent factor to achieve identifiability. For complex hierarchies, the model was split into smaller subtrees, fixing loadings from the construction phase. Competing conditional estimates were compared, choosing the estimate with the smallest absolute Z-score as the best representation of direct effect. Differences between original and conditional estimates were tested by the Clogg et al. method; effects with p>0.05 were considered “direct effect only”. Gene prioritization: For each sub-locus sentinel SNP, HaploReg was queried (r2≥0.8). Genes were prioritized by: (1) sentinel/LD proxy being non-synonymous in gene; (2) sentinel/LD proxy coding variant; (3) top SNP intronic or UTR in gene; (4) top SNP in strong LD with intronic variant; (5) nearest gene. Enrichment analyses: Genes near loci associated at p<5×10−8 were used for tissue enrichment (FUMA, GTEx general/specific tissues) and Gene Ontology term enrichment (clusterProfiler enrichGO). Brain MRI genetic correlations: GWAS summary stats for 3,260 imaging-derived phenotypes (morphology and resting-state ICA100 networks) were obtained from BIG40. IDPs with low heritability or high uncertainty were removed (remaining 2,329 IDPs). Genetic correlations with the three main liking factors were estimated via high-definition likelihood (HDL); FDR correction applied across 6,987 tests (q<0.05).

- Hierarchical structure of food-liking: Three main dimensions were identified: F-Highly palatable (e.g., desserts, meat, savory), F-Low caloric (e.g., vegetables, fruit, whole grains), and F-Acquired (e.g., unsweetened coffee, alcohol, cheese, strong-tasting vegetables). A minor F-Caffeinated Sweet Drinks group also emerged. - Genetic correlations among liking dimensions: F-Low caloric and F-Acquired were moderately correlated (rg=0.59). F-Highly palatable was largely independent from the other two (rg≈0.05 and 0.16). F-Caffeinated Sweet Drinks had weak positive correlation with F-Highly palatable (rg=0.39) and weak negative correlations with F-Acquired (rg=−0.30) and F-Low caloric (rg=−0.25). - Liking vs consumption: Strong genetic correlations were observed between liking and corresponding consumption traits (all rg>0.7 except beer rg=0.4 and white bread rg=0.1). Mean SNP heritability for liking (~0.08) was about twice that of consumption (~0.04), with liking higher for nearly all items. - Genetic correlations with other complex traits: F-Highly palatable correlated with higher BMI and body fat percentage and with indicators of lower socio-economic status; F-Low caloric and F-Acquired showed opposite patterns (lower obesity indices, better lipid profiles). F-Acquired associated with higher educational attainment and sedentary jobs; F-Low caloric showed no correlation with educational attainment but positive correlation with non-sedentary jobs. - Brain MRI genetic correlations: F-Acquired and F-Low caloric showed negative correlations with cortical thickness (frontal, parietal, occipital regions) and positive correlations with cortical surface area (peri-central sulcus, temporal fusiform, insula). F-Highly palatable showed negative correlations with striatal volumes (putamen, caudate) and few connectivity associations (notably positive with rostral frontal-parietal networks). These distinct patterns suggest different neurophysiological underpinnings across liking dimensions. - GWAS hits and pleiotropy: 1,401 significant food-liking associations across 173 loci were identified; 143/173 loci (~82%) were pleiotropic (mean ~8 traits per locus). Highly pleiotropic loci included MHC II (58 traits), FTO (51), and CADM loci (59 total), suggesting non-specific effects on food-liking. - Replication: Of 235 testable SNP–trait associations, 61 (26%) replicated at one-tailed p<0.05 with concordant direction; 194 (82.5%) showed consistent direction of effect (binomial p=5×10−25), indicating underpowered replication rather than lack of true effects. - Colocalization: Within 143 loci, 138 had at least one cluster of traits colocalizing to the same variant(s), yielding 203 distinct clusters. 225 of 1,270 locus–trait associations did not colocalize with others. An exception was the chromosome 17 inversion region (MAPT) where colocalization failed. - Direct vs mediated effects: Among 1,261 associations testable within the hierarchy, 495 were inferred as direct effects, highlighting that many associations operate through higher-order factors. - Gene prioritization and notable genes: 250 likely causal genes prioritized. Approximately 43.8% of associations were intragenic; ~7% non-synonymous; ~6% UTR; ~1% synonymous. Twelve prioritized genes were taste (4) or olfactory receptors (8). Strongest single-item association: OR4K17 with onion liking (beta=0.31 on 9-point scale; p=4×10−71). TAS2R38 variants associated with liking of salty foods, alcoholic beverages, horseradish, and grapefruit; taste and olfactory receptor associations were confined to F-Acquired and F-Low caloric, not F-Highly palatable. Variants near FGF21 associated with higher sweet liking and lower liking of strong-tasting foods (e.g., fish, eggs, mayonnaise, fatty foods). A non-synonymous GIPR variant associated with higher liking of low-caloric foods and lower liking of fatty foods (e.g., mayonnaise, cheese, cream) and aligns with lower BMI associations.

The study demonstrates that food-liking has a measurable, biologically grounded genetic architecture. Liking dimensions derived from genetic correlations reveal that highly palatable, energy-dense foods are genetically distinct from low-caloric and acquired tastes, suggesting different underlying biological processes. High genetic correlations between liking and consumption, despite measurement ~10 years apart, imply stable, overlapping genetic influences, while differences likely reflect environmental and behavioral factors. Distinct neurophysiological signatures further validate the model: F-Highly palatable correlates negatively with striatal morphology (putamen, caudate), consistent with prior evidence of altered reward circuitry in energy-dense food exposure and adiposity, whereas F-Low caloric and F-Acquired relate to morphology and connectivity in frontal, parietal, and occipital regions involved in sensory processing, identification, and decision making. Enrichment of neural processes (notably glutamatergic and GABAergic synapses) and subcortical brain tissues supports a central role of brain circuits—especially basal ganglia—in shaping food-liking. The hierarchical framework clarifies pleiotropy and mediation, distinguishing variants with broad influences (e.g., MHC II, FTO, CADM) from those acting on specific items or factors. The observation that taste and olfactory receptor genes associate with acquired and low-caloric items but not highly palatable foods suggests potential evolutionary pressures maintaining preference for energy-dense foods while modulating acceptance of aversive or complex flavors. The inconsistent replication of some candidate gene associations (e.g., limited evidence for sweet/umami receptor genes) emphasizes the value of large-scale, hypothesis-free approaches. Findings for FGF21 and GIPR illustrate mechanisms shifting preferences between sweet/fatty and lower-caloric options, potentially influencing diet quality and body weight. Overall, the results provide a comprehensive map linking genetic variation, structured liking dimensions, consumption behavior, and neurophysiological traits, informing future studies on dietary behavior, causal inference (e.g., Mendelian randomization), and intervention development.

This large-scale GWAS of food-liking in over 150,000 individuals maps a hierarchical structure with three principal dimensions—Highly palatable, Low caloric, and Acquired—and identifies 1,401 associations across 173 loci, many novel and pleiotropic. Liking is strongly genetically correlated with consumption yet exhibits roughly double the SNP heritability, indicating a robust biological component. Distinct genetic correlations with brain morphology and connectivity underscore differing neurophysiological underpinnings, particularly implicating basal ganglia in preferences for highly palatable foods. The study refines interpretation of prior associations, highlights roles for genes such as FGF21 and GIPR in shifting preferences, and provides resources (summary statistics, prioritized genes) for future work. Future research should extend analyses beyond European ancestry, improve replication power, investigate causal pathways to health outcomes, explore evolutionary hypotheses around taste/olfaction genes and palatable foods, and test interventions targeting preference modulation to enhance dietary quality.

Primary limitations include selection bias in UK Biobank, whose participants are healthier, older, and more educated than the general population, potentially biasing estimates—especially genetic correlations. Additional bias may arise from participation in optional follow-up questionnaires. Average liking rankings suggest possible reverse causality (e.g., lower reported liking for typically palatable items), similar to observations for consumption traits. Replication was constrained by limited overlap of traits across cohorts and modest sample sizes, resulting in lower formal replication rates despite strong directional concordance. Many liking traits lacked matched consumption GWAS, limiting direct validation. Finally, the chr17 inversion region complicated colocalization analyses.