Title: Deep Neural Networks for Prediction and Control of Dynamical Models
Abstract: Dynamical models of biological systems are often interesting due to the high-level qualitative behavior that they display. However, the mathematical tools for analyzing qualitative dynamics, such as bifurcation theory, have trouble scaling to high dimensional problems. In this talk we will discuss a method which utilizes neural networks coupled with numerical differential equation solvers to understand the behavior of phenomenological biological models. Using a modification of methods from Generative Adversarial Networks (GANs), we will demonstrate a novel inversion technique with the ability to predict parameters in which biological ecosystems are constrained to cycle in a chosen domain and predict reaction rates such that Turing patterns will occur. An application will be discussed showing how this method can be coupled with pharmacokinetic/pharmacodynamic (Pk/Pd) estimation to predict intervention doses for neonatal opioid addiction, offering a fully automated pipeline for personalized treatment of a common high-risk condition.