This collaborative project with IBM Research aims to make breakthrough improvements in methodologies for building models from observational data that can both predict and explain biological phenomena. A fundamental challenge in the life and health sciences is explaining hidden physiological and disease mechanisms. These hidden mechanisms shape experimental and clinical observations, and medicine strives to improve health by influencing them with new therapies such as drugs. Machine learning can now make stunningly accurate predictions of biological phenomena based on observed data from variable sources. These predictions, however, are difficult to interpret and exploit, because they usually do not address the underlying physiological mechanisms from which the data and predictions derive. On the other hand, mechanistic models replicate features of the experimental and clinical data and address their causes with model parameters that represent the underlying physiology. But these models usually fail to address inherent cell-to-cell and patient-to-patient variability. This project will develop a hybrid deep learning/mechanistic modeling framework that can capture and explain the inherent variability in biological data through identification of parameter sets that result in model outputs consistent with data. The framework is intended to be versatile enough to find input parameters of a model for multiple conditions distinguished by some factor (e.g., treatment, age, or disease state) simultaneously; such 'intervention' scenarios are common in practice. The framework will advance the state of the art by enabling researchers to incorporate additional constraints based on prior knowledge about the nature of an intervention. This project will provide interdisciplinary industrial research experiences to community college and graduate students.
This project will develop and apply novel hybrid modeling architectures that use generative adversarial networks, a class of machine learning algorithms in which two artificial neural networks compete, to map distributions of experimental observations to distributions of biophysical model parameters. The system will tackle a set of important biological questions involving the electrophysiology of circadian clock neurons and aging, cardiac arrhythmias, and Alzheimer's disease using datasets provided by experimental collaborators. First, the system will be employed to identify which ion channel conductances are involved in the age-related decline of circadian rhythm amplitude in suprachiasmatic nucleus neurons and the altered excitability properties of hippocampal neurons in mouse models of Alzheimer's disease. Second, the system will be employed on human electrocardiogram data to test the hypothesis that circadian rhythms in cardiac excitability can affect the efficacy of drugs used to treat cardiac arrhythmias. Students from Essex County College, an open-access, two-year college that is federally designated as a minority serving institution, will perform summer research mentored by a team of researchers from academia (New Jersey Institute of Technology and Purdue University) and industry (IBM) with complementary expertise in artificial intelligence and deep learning, biophysical modeling and simulation, and dynamical systems theory. ECC students, as well as NJIT graduate students, will gain exposure to the industrial research environment through interactions with IBM's T.J. Watson Research Center.
This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.
|Effective start/end date||7/15/22 → 6/30/25|
- National Science Foundation: $464,102.00