Diet is a modifiable risk factor for cardiometabolic diseases, yet establishing causality between diet and disease outcomes remains challenging(1). Most evidence in nutritional epidemiology stems from observational studies, which are susceptible to confounding and reverse causation, and many observational associations have failed to replicate in large-scale randomised controlled trials (RCTs). Although RCTs are considered the gold standard for causal inference, they often face practical barriers in nutrition research. Mendelian randomisation (MR) offers a complementary alternative by using genetic variants as instrumental variables (IVs) to proxy modifiable exposures such as diet(2). As genetic variants are randomly assorted at conception and remain fixed throughout life, MR can be thought of as nature’s RCTs to infer the long-term causal effect of an exposure. However, identifying valid IVs for dietary exposures remains difficult. Conventional significance-based approaches often select variants that exhibit pleiotropy, i.e. affecting multiple traits beyond the dietary exposure of interest. This violates a core MR assumption that IVs only affect the outcome through the exposure and may give misleading conclusions about the true effects of diet. To address this, we used genetic IVs related to taste and smell perception, which are biological drivers of food preferences and intake. Anchoring IVs in the sensory biology of dietary behaviours can minimise pleiotropic bias and improve the reliability of diet-related MR findings(3). MR was applied to assess the causal effects of four dietary patterns (DPs) (Unhealthy, Healthy, Meat-based, Pescatarian) from the UK Biobank (n = 258,758-449,210) on twelve cardiometabolic outcomes (n = 149,006-1,320,000), including body mass index, coronary artery disease (CAD), high-density lipoprotein cholesterol, low-density lipoprotein cholesterol, total cholesterol, triglycerides (TG), systolic blood pressure, diastolic blood pressure, type 2 diabetes (T2D), haemoglobin A1c, fasting insulin, and fasting glucose (FG). Findings included robust evidence (Bonferroni-corrected p < 0.00185) that the Unhealthy DP causally increased risk of CAD (odds ratio [95% confidence interval] = 3.23 [1.89, 5.52] per SD increase in Unhealthy DP score, p = 1.86×10-5) and TG levels (beta = 0.45 ± 0.13 SD per SD increase in Unhealthy DP score, p = 5.23×10-4), supporting existing mechanistic understanding between diet and cardiovascular disease risk. Suggestive evidence (p < 0.05) was found for an Unhealthy DP increasing T2D risk, a Healthy DP lowering FG in individuals without T2D, and a Meat-based DP elevating CAD risk. Notably, our chemosensory IVs did not replicate implausible associations identified using conventional significance-based IVs, such as a Healthy DP increasing BMI, suggesting that these may reflect residual bias or pleiotropy. Overall, this study demonstrated the utility of MR in strengthening causal inference and triangulating evidence in nutritional epidemiology, while emphasising the crucial importance of using biologically informed IVs in MR studies on dietary exposures.