DNA binding economies

This thesis develops a new scalable modeling framework at the interface of game theory and machine learning to recover economic structures from limited slices of data. Inference using economic models has broad applicability in machine learning. Economic structures underlie a surprisingly broad array of problems including signaling and molecular control in biology, drug development, neural structures, distributed control, recommender problems, social networking, as well as market dynamics. We demonstrate the framework with an application to genetic regulation. Genetic regulation determines how DNA is read and interpreted, is responsible for cell specialization, reaction to drugs, metabolism, etc. Improved understanding of regulation has potential to impact research on genetic diseases including cancer. Genetic regulation relies on coordinate binding of regulators along DNA. Understanding how binding arrangements are achieved and their effect on regulation is challenging since it is not always possible to study regulatory processes in isolation. Indeed, observing the action of regulators is an experimental and computational challenge. We need causal genome-wide models that can work with existing high-throughput observations. We abstract DNA binding as an economy and develop fast algorithms to predict average binding arrangements as competitive equilibria. The framework supports viewing regulation as a succession of regulatory states. We complete the framework with algorithms to infer causal structure from high-throughput observations. Learning here deviates from work in learning in games, it is closer to the economic theory of revealed preferences. Our algorithms predict the effect of experimental perturbations and can be used to refine experimental hypotheses. We show that the economic approach reproduces known behavior of a genetic switch (λ-phage), and that it can complete the map of coordinate binding in yeast. Thesis Supervisor: Tommi Jaakkola Title: Associate Professor