Steering T-Cell Adaptation Using Opponent Exploitation Algorithms and Computational Game Theory (Research Area 10.3)
Abstract
Living organisms adapt to challenges through evolution and adaptation. These survival mechanisms have proven to be a key difficulty in developing therapies, since the challenged organisms develop resistance. It would be desirable to be able to harness evolution/adaptation for therapeutic, technological, and scientific goals. For example, could we steer a person s own T cell population to a state that causes the immune system to better treat the disease at hand, e.g., cancers or autoimmune diseases? We propose the wild idea of steering evolution/adaptation strategically using computational game theory and opponent exploitation techniques. A sequential contingency plan for steering evolution/adaptation is constructed computationally for the setting at hand. For example, for therapeutics, we propose modeling this as a (zero-sum) imperfect-information game between a treater and a disease, with potentially both sequential and simultaneous moves. Solving the game model for Nash equilibrium (or its refinements) provides an optimal treatment plan assuming the disease plays optimally, that is, in the worst possible way for the treater. The scalability of algorithms for imperfect-information games has increased by orders of magnitude over the last ten years. The leading approach involves running an abstraction algorithm to construct a smaller, strategically similar game, then computing a (near-) equilibrium of the abstract game, and then mapping the computed strategies to the original game. We will adapt and enhance algorithms for finding game-theoretic equilibria for this biological domain. The game-theoretic approach does not need a probabilistic model of disease behavior: it assumes the opponent behaves in the worst possible way for us. This is safe in the zero-sum setting: if the opponent actually does not behave in this way, that can only benefit us. However, the purely game-theoretic approach may sometimes be overly conservative: the disease may not behave optimally. We propose that opponent exploitation techniques be used to take advantage of the disease s suboptimal play. In essence, an opponent model predicts what the opponent would do - typically probabilistically - in various points in the game. In the proposed work, several opponent exploitation approaches will be adapted for this domain and tested. Biological opponents have a distinct weakness that will further be exploited here. Evolution and adaptation is myopic: it does not look ahead in the game tree. An example of this is evolving a disease into a trap where it can be easily attacked so that it is destroyed or becomes less powerful (e.g., less virulent, less contagious, or less able to evolve in bad ways). More generally, the task is to compute a strategy for ourselves that yields low utility to the myopic opponent. In this project we will apply this novel approach to steering T cell adaptation in order to I) carefully down-regulate overactive T cells in autoimmune diabetes, and 2) up-regulate T cells to battle cancer. T cells have the capacity to differentiate into one of at least eight different Th subsets, each with its own molecular signature and effector functions that evolved to deal with specific pathogens but which can also cause immunopathology and disease. Autoimmunity and cancer represent two cases in which inappropriate T cell responses contribute to disease. The sequential plans generated via the proposed computational approaches will be evaluated in vitro and in vivo. This will close the loop, enabling us to go back and refine our models and algorithms, and repeat. High-level summary of work across years: Year l: Better opponent exploitation algorithms; better action sets and observables; first in vitro tests. Year 2: Enhanced game-solving algorithms; hybrid algorithms; scalable game representations; in vitro and in vivo tests. Year 3: Advanced algorithms; in vitro and in vivo tests.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Sep 11, 2018
- Source ID
- W911NF1710082
Entities
People
- Tuomas Sandholm
Organizations
- Army Contracting Command
- Massachusetts Institute of Technology
- United States Army