Engineering Cellular Self-Organization

Abstract

Understanding how local cell-cell interactions orchestrate global organization of tissues, organs and entire body plans is a central challenge in developmental biology. Despite decades of progress, it remains difficult to (i) connect mechanistic models of individual cells with observed collective behaviors of tissues; (ii) to develop such models with experimentally accessible “tuning knobs”; and (iii) to use these models to predict and design what happens in complex assemblies of cells. In this thesis, we aim to develop a unified framework, based on differentiable programming, to both engineer and infer multicellular behaviors within physically grounded models of development. We present three such examples in Chapters 2-4 that operate at the interface of theory and experiments. We model tissues as collections of interacting cells in 2D or 3D that can communicate through morphogen diffusion, mechanical forces, and gene regulatory networks governing local decision-making about proliferation and morphogen secretion. By embedding these physical processes within an end-to-end differentiable simulation, we enable the use of gradient-based optimization to learn interpretable rules that drive emergent multicellular organization. To handle stochastic or non-differentiable elements in the simulation, such as in cell division, we employ policy-gradient methods to “estimate” a gradient signal from a batch of simulations. In Chapter 2, using this approach, we show that simple gene regulatory circuits can generate complex morphogenetic behaviors—including directional growth, homeostasis, and shape formation—through purely local interactions. We infer gene regulatory mechanisms to elongate or branch a cluster of cells, to establish homeostasis between cell types via growth factor signalling, and to couple mechanics to proliferation. In Chapter 3, we examine the potential of such approaches to interpret and inform experiments, by quantitatively modeling cell–cell adhesion in engineered cells. By engineering cell types with different combinations and expression levels of cadherins, we use high-throughput sorting assays to fit physical potentials to pairs of interacting cell types. We demonstrate that we can invert these models to identify physical potentials that can assemble multicellular structures with a target spatial organization. We argue that such methods hold promising potential to elucidate the principles of multiple-cadherin cell types, as is prevalent in development. Finally, in Chapter 4, we apply our approach to real cell tracking data (by Guignard et al.) from light sheet microscopy of a developing ascidian embryo, imaged from the 64-cell stage to the early neurula stage. We constrain the initial state of simulations to the actual morphogen expression pattern at the 64-cell stage, and show that we can infer gene regulatory mechanisms that recapitulate the spatiotemporal sequence of cell division throughout the experimental trajectory. The learned mechanism shows conserved morphogen expression patterns across embryos, and unique to each morphogen. These results illustrate the potential of differentiable models as data-driven simulators of development. Together, this work establishes differentiable programming as a powerful bridge between theory and experiments, enabling both the design of novel multicellular systems and the inference of regulatory mechanisms in developing organisms. These results point toward a future in which the principles of self-organization can be systematically uncovered and harnessed to engineer living systems with predictive control. With the rapid increase in GPU-accelerated frameworks for large MD simulations, and rich transcriptomic datasets at the single-cell scale, we can infer complex mechanistic models to recapitulate observables; these can be refined with more perturbation experiments. This opens up doors to precise spatiotemporal control over biological systems like organoids.