Using deep learning to understand computations in neural circuits with Connectome-constrained Mechanistic Models

Advances in experimental techniques yield detailed wiring diagrams of neural circuits in model-systems such as the Drosophila melanogaster. How can we leverage these complex connectomes, together with targeted recordings and perturbations of neural activity, to understand how neuronal populations perform computations underlying behavior? Achieving a mechanistic understanding will require models that are consistent with connectomes and biophysical mechanisms, while also being capable of performing behaviorally relevant computations. Current models fail to address this need: Mechanistic models satisfy anatomical and biophysical constraints by design, but we lack methods for optimizing them to perform tasks. Conversely, deep learning models can be optimized to perform challenging tasks, but fall short on mechanistic interpretability.To address this challenge, we will provide a machine learning framework that unifies mechanistic modeling and deep learning, and will make it possible to algorithmically identify models that link biophysical mechanisms, neural data, and behavior. We will use our approach to study two key neural computations in D. melanogaster. We will build large-scale mechanistic models of the optic lobe and motor control circuits which are constrained by connectomes and physiological measurements, and optimize them to solve specific computational tasks: Extracting behaviorally relevant information from the visual input, and coordinating leg movements to achieve robust locomotion. Our methodology for building, interpreting and updating these `deep mechanistic models' will be applicable to a wide range of neural circuits and behaviors. It will serve as a powerful hypothesis generator for predicting neural tuning and optimizing experimental perturbations, and will yield unprecedented insights into how connectivity shapes efficient neural computations in biological and artificial networks.