Integrating AI and Mechanistic Modeling to Improve Real Forecasting of Biological Systems
PI:
TJ Sego, PhD
Luis Sordo Vieira, PhD
Project Summary
Computational modeling is now a well-established methodology to act as a surrogate for biological systems. Computational modeling in the biological domain is largely subdivided into mechanistic mathematical modeling and data-driven modeling. Mechanistic modeling of biological systems has led to significant advances in our understanding and applications in fields as diverse of developmental biology, cancer, immunology, and cardiovascular medicine. However, mathematical modeling is largely limited when mechanisms are unknown, or when data is sparse. Recent advances have brought data-driven methodologies (e.g., artificial intelligence/machine learning) to the forefront of computational modeling. These methods excel at pattern recognition when data is plentiful. On the other hand, data-driven methodologies are limited in their ability to infer mechanisms beyond data available. Due to their relative strengths and limitations, scientific consensus now calls for hybrid modeling, combining mathematical modeling with data-driven methodologies. Such nascent hybrid methodologies are starting to be used in the field of biomedicine. Of relevance are methodologies that infer mathematical structures from datasets (e.g., generative modeling and symbolic regression) and data-driven methodologies to improve the performance of mechanistic modeling when mechanisms are unknown. Although these methods have proven their potential in physics, cosmology, and quantum mechanics, their potential in the biomedical domain remains largely untested, and their potential synergy with established modeling approaches remains unexplored. Unlike engineered systems, in which many of these methods have been developed and deployed on, biological systems are complex, spanning various spatiotemporal scales, with no obvious hierarchy of communication between scales. As such, it remains unclear whether these methodologies can be deployed in biomedicine.
We propose to develop and evaluate two complementary strategies that integrate mechanistic modeling with machine learning. The goal is to create hybrid models that are predictive, interpretable, and capable of extrapolation, enabling more robust forecasting and control across biological scales. Our first objective is to test the capabilities of symbolic regression to generate mathematical models of multiscale biological systems. We will generate an ODE model from target data using neural differential equations with regularized evolution via symbolic regression and attempt to parametrize them to increasingly difficult data. We will begin with targeting synthetic data from a multi-scale agent-based model of the immune response to an influenza infection for which an ordinary differential equation (ODE) model representation is known. We will then move to synthetic data from an agent-based model with unknown solutions, and then real-world intensive care unit data. In all cases, we will compare the ability of the hybrid model from each modeling method to parameterize to a training dataset as a measure of predictive capability, and to predict a test dataset with the parameterized model as a measure of extrapolatory capability. Our second objective is to test hybrid modeling methods that emphasize model interpretability, and methods for which AI/ML can enhance the potential of mathematical modeling by inferring unknown mechanisms. We will achieve this by constructing ODE models with an additive Gaussian Process stochastic model term which can adaptively learn from data. We hypothesize that the additive stochastic model term includes information that could inform the nature of the missing mechanisms and enhance the model’s extrapolatory potential.
By making mechanistic models more data-driven while enhancing AI models to reason beyond its training data, this project will advance the role of computational biology to inform about the principals underlying biological systems organization and emergent function.