Structure from Dynamics: Machine Learning and Modeling in Cells Undergoing Multi-scale Self-organization

Abstract

Living things are complex, self-assembling systems capable of generating structured patterns that span nanosecond-scale behavior of amino acids to trillions of cells organized in a decades-long process of development. With computational models we can represent biological processes as mathematical relationships, and simulations allow us to observe and predict their behavior. Here we confront several challenges in the process of constructing and making sense of computational biological models. We present a vision for how emerging computational tools and frameworks can enable investigators to construct models that are both larger and more detailed than previously possible. We demonstrate that these energy-rule based structural models can explain how changes in individual protein conformations lead to adaptive resistance to targeted cancer therapy. Further, we envision a road map for the next generation of models, combining computationally predicted protein structures, machine reading to encode molecular interactions, biophysical constraints, and automated mining of massive in-silico experiments to provide unprecedented biological understanding. We then focus on applications of machine learning tools to the challenge of understanding the biological regulatory circuits that govern early development. We present MC-Boomer, a reinforcement learning algorithm that automatically constructs Boolean models with specified behaviors. We describe a model analysis framework that identifies families of structurally similar models and extracts the interactions that drive their behavior. We then introduce TangleFlow, a deep learning architecture that models the dynamics of cells on the developmental landscape. Applied to a mouse model of Cornelia de Lange syndrome (CdLS), TangleFlow accurately captures subtle differences in early cell fate commitment that underlie large-scale phenotypic differences in CdLS patients. We further present a novel analysis technique, relying on the sparsity of the TangleFlow architecture, that allows us to identify multi-gene regulatory circuits guiding cell fate commitment. Together, this work aims to both propose and implement a novel vision for constructing and analyzing data-driven models of biological systems. We aim to build computational bridges from atomistic biophysical descriptions of protein conformations to tissue-scale cell fate organization, helping us to manipulate biological systems to restore their health and function.