Machine learning for functional connectomics in Caenorhabditis elegans

Thesis

Santiago Ram n y Cajal first traced the microscopic intricacies of individual neurons in the late 19th century. His work revolutionised neuroscience, and earned him the moniker “the father of modern neuroscience.” Cajal was a major contributor to the newly established neuron doctrine, outlining that even the most complex organism behaviours are simply the result of the interactions between networks of small, discrete units, called neurons. Since then, neuroscience research has measured, codified, and modelled these neural interactions in progressively more detail.

However, these neural models are rarely applied at whole-brain scales, and are even less frequently used as mechanistic models of neural activity in which Bayesian inference is performed. This offers an exciting and truly ground-breaking opportunity. Application of Bayesian inference to whole-connectome neural models would allow previously unobservable physiological states and parameter values to be inferred from readily observable data at an unprecedented scale. It would also allow interpretable neural models to be refined, tested and even learned directly from observed data without manual input. Finally, tuned whole-brain simulators promise in silico neuroscience experiments, where hypotheses can be investigated without the need for biological specimens. Such experiments are not hindered by the logistical requirements of obtaining measurements in vivo, and can be easily, cheaply and reliably reproduced.

In this thesis we develop tools to perform this estimation. Specifically, we study performing this estimation in the nematode roundworm Caenorhabditis elegans, the only organism for which the entire connectome has been fully mapped. At its core, the work we present is the first attempt to infer, at whole-connectome scale, neural states and parameters from limited observations. We introduce the machine learning and computational neuroscience tools required to make this tractable, and explore extensions arising directly from this work. We conclude by suggesting further extensions of this work that are within reach, and discussing more wide-reaching ideas and implications of this work.

Authors: Warrington, A

• DOI: 10.5287/ora-ganrdbjke
• Type of award: DPhil
• Level of award: Doctoral
• Awarding institution: University of Oxford
• Language: English
• Keywords: C. elegans; Connectomics; Neuroscience; Machine Learning; Caenorhabditis elegans
• Subjects: Connectomics; Computational neuroscience; Machine learning; Caenorhabditis elegans