UCSF

Living within eukaryotic cells, the endosymbiont organelles mitochondria organize into complex networks and undergo constant reorganizations. With an almost purposeful self-directing quality, mitochondrial movements have puzzled cell biologists since its earliest observation under the microscope. Today, most understand mitochondrial dynamics by their fission and fusion events, in which mitochondria fragment and merge with one another. However, these two classes of dynamics fail to explain many other movements that involve more complex shape changes and may not be regulated by known molecular machinery. With access to a newly invented microscope, SNOUTY, ideal for imaging fast-moving organelles, this thesis investigates a previously underappreciated set of dynamics regulated by physical forces. In addition, we develop the early stages of a unique computational platform that seeks to combine live-cell microscopy with image tracking, agent-based simulations, and generative modeling to unify the diverse movements of mitochondria into a single in silico environment. The first section of this dissertation establishes a biophysical mechanism for an emerging class of dynamics, mitochondrial pearling. Using high-speed light-sheet microscopy we capture spontaneous and transient pearling events occurring in many physiological contexts including neuronal action potentials, T cell activation, and replicative senescence. We design several chemical, genetic, and mechanical methods for inducing pearling which establish three primary physical forces driving pearling events. These forces highlight how mitochondria's unique membrane properties of electrochemical coupling and inner membrane folding regulate these dynamical transitions. We conclude by placing mitochondrial pearling as a primary class of dynamics with broad implications for mitochondrial biology. The second section begins to implement a procedural modeling system to simulate mitochondrial networks using 3D computer graphics. This model is used to generate synthetic representations of mitochondria with modular subsystems to represent each known class of mitochondrial dynamics. Next, a generative deep learning algorithm is developed to interface simulation and microscopy data. Each dataset is run through an image tracking and feature extraction pipeline, nellie, and then fed into a novel deep learning generative algorithm, and subsequently propagated back to train simulation parameters. This training scheme is used to identify sets of model architectures that correctly represent empirical data. Trained simulations thereby can be used to establish which sets of mitochondrial processes are sufficient to model their dynamics and generate synthetic datasets composed of realistic animations of mitochondria. The dissertation concludes with placing our novel biophysical mechanisms and computational methods into a larger conceptual context. This includes future utilities for biotechnology and reframing important physiological events with mitochondrial dynamics in mind.