Bruce Kristal received a BS in Life Sciences from MIT in 1986 and his PhD (Virology) from Harvard University in 1991. He rose from post-doc to Research Assistant Professor at the University of Texas Health Science Center at San Antonio (1991-1996), then moved to Burke Medical Research Institute (1996) and the Departments of Biochemistry (1997) and Neuroscience (1998) at Weill Medical College of Cornell University, becoming an Associate Professor in Neuroscience in 2004. He joined Brigham and Women’s Hospital's Department of Neurosurgery in April 2007 and the Department of Surgery at Harvard Medical School in 2008 and transitioned to the Division of Sleep and Circadian Disorders in the Department of Medicine in 2016. Dr. Kristal was the founding secretary (2004-2008) and member of the Board of Directors of the Metabolomics Society (2004-2011). His laboratory currently has active projects in developing blood tests that predict future disease risk for preventable disorders (e.g., diabetes, breast and colon, heart disease), and the development of computer-based approaches to better deliver personalized medical care.
Talk Title: Long-Range Disease Risk Prediction -- Algorithmic Information Fusion in a Life and Death Environment
Over-nutrition and suboptimal dietary macronutrient choices are arguably the major environmental stressor in individuals living in Western societies. Obesity and poor diet are estimated to cause or contribute to as many as 25% of all cancers, in addition to cardio- and cerebro-vascular disorders (hypertension, heart attack, stroke) and metabolic diseases such as diabetes. One of the most clear examples of this, dietary or caloric restriction (CR), is the most potent and reproducible known means of increasing longevity and reducing morbidity (including cancer, diabetes, etc) in mammals. As one example, risk of breast cancer is generally decreased by more than 90% in CR rodents, and the CR-mediated effects are usually dominant to those induced by genetic risk factors, carcinogens, or co-carcinogens. The robust observations of reduced morbidity in CR animals is directly analogous to studies in humans that link obesity with poor health outcomes, including increased risk of neoplastic disease. We therefore proposed to test the general concept that biomarkers of diet in rats will predict risk of future disease in humans.
Preliminary work with plasma small molecules (metabolomics) shows promise, but long-term clinical utility would be enhanced by consideration of other factors (blood proteins, genetics, demographic considerations, etc. Among other complications, these data are collected from different sources (e.g., humans and animals), different data structures (e.g., qualitative, discrete, continuous, bimodal), and can be static or temporally shifting (e.g., genetic vs environmental), etc... Thus, this is a classic -- and nasty -- fusion problem that will require work at all levels -- starting by simply defining the ground rules for success and failure and by trying to consider in advance the potential problems. Major issues involve deciding if, when, and at what level (e.g., data, latent variable, decision model) to conduct algorithmic fusion as well as the criteria by which we can evaluate in the absence of infinitely large training and test sets. We will present our modeling approaches, the models, and their ability to distinguish sera based on caloric intake, as well as data from the initial application of these markers. We will then address some of the checks and cross-checks used to evaluate these data and lay out a path towards fusion-based approaches.