Abstract-Project 2: Causal Relationship Disentangler for Precision Nutrition Predicting individual responses to food and dietary patterns, the stated goal of the National Institutes of Health (NIH) Common Fund’s Nutrition for Precision Health program, requires uncovering the causal connections between diet and health. Despite the importance of diet for treating and reducing risk of many chronic diseases, guidelines often rely on associations rather than causal relationships. Establishing a causal model (set of causal relationships) is vital to provide accurate dietary guidelines to individuals and help them balance priorities. The key obstacles to a comprehensive model of causes and effects of diet have been a lack of methods to translate findings to new populations and a lack of data suitable to learn about causes. The first major challenge is understanding to whom and under what conditions a finding applies. There are no existing methods that can identify causal relationships between diet and other factors and can determine when these findings apply. A second core obstacle is that dietary studies often capture different sets of variables due to the cost and challenge of collecting data on the many causes and effects of nutrition, and many studies rely on food logs kept by participants. This leads to missing variables and missing values, and both can confound causal inference. Many methods exist for imputing missing values but they may lead to unacceptable errors for individuals based on patterns of missingness in real-world data. Single imputation methods provide a single value for each missing instance. Thus, given the type of missingness we face in nutrition (both missing at random [MAR] and missing not at random [MNAR]) and the importance of establishing causal relationships rather than correlations, there is a significant need for new imputation methods. To address this, we introduce new approaches for handling missing data that preserve causal structure. In the Causal Relationship Disentangler for Precision Nutrition we propose new methods for causal generalizability that learn when and why causal relationships are true. Our methods are applicable to all health outcomes and timescales. Learning how to transfer causal knowledge and doing so with missing data is critically important for realizing the potential of nutrition for precision health. Precision health requires knowing what conclusions we can draw about both populations and individuals and being able to systematically predict what interventions will work for an individual. Our automated approaches to generalizing causal models will provide the critical link between data and actions, allowing the knowledge created to generalize beyond All of Us. Our investigative team has for over a decade developed new methods that learn causal models from observational data and provide automated causal explanations, as well as statistics, data science, and biostatistics. Aim 1 will develop methods ...