Abstract
What we eat shapes not only our metabolic health but, increasingly, appears to influence brain structure and function across the lifespan. Dietary patterns and specific nutritional exposures are among the most readily modifiable risk factors for cognitive decline, yet the neural mechanisms through which diet affects the brain remain incompletely understood. A further complication is that genetic risk for Alzheimer’s disease may modify these relationships: the same dietary pattern might be neuroprotective in one genetic background and neutral in another. Disentangling these interactions requires large-scale datasets that combine dietary data, neuroimaging, genetic information, and longitudinal cognitive outcomes.
Using UK Biobank data, this research program investigates the relationships between dietary factors and brain health through multiple complementary studies. Two published studies focused on the associations between specific dietary exposures and neural connectivity. The first examined coffee and tea consumption and found that their associations with neural network connectivity patterns were modulated by genetic risk factors for Alzheimer’s disease — suggesting that the neuroprotective or neutral effects of these common beverages depend in part on an individual’s genetic background (Nutrients, 2024). The second study took a broader view, examining how overall dietary patterns associate with resting-state functional connectivity and demonstrating that AD genetic influences significantly impact these diet–brain relationships (Nutrients, 2023).
An ongoing study extends this work from functional connectivity to structural outcomes, exploring the relationship between nutrition and brain volumetrics, cortical thickness, and gray-white matter contrast. A methodological innovation in this study is the use of large language model (LLM)-based embeddings to learn a representative space that captures dietary pattern structure from high-dimensional nutritional data. Traditional approaches to dietary pattern analysis (principal components, clustering on nutrient intakes) impose strong assumptions about how foods relate to one another. LLM-based embeddings offer a more flexible alternative, learning dietary representations directly from the data in a way that can capture complex, nonlinear relationships between foods and eating behaviors. This approach has the potential to reveal dietary patterns that conventional methods miss, and to link them more precisely to neuroimaging outcomes.