Date of Award:

5-2022

Document Type:

Thesis

Degree Name:

Master of Science (MS)

Department:

Computer Science

Committee Chair(s)

Nicholas S. Flann

Committee

Nicholas S. Flann

Committee

Gregory J. Podgorski

Committee

Vladimir Kulyukin

Abstract

Deep learning continues to solve significant scientific and engineering problems, but the solutions found are neural networks with thousands of parameters that provide no scientific or engineering insights. A solution to this problem, explored in this work, is to learn mathematical models that represent mechanisms that can be interpreted by scientists and engineers.

A challenging learning problem is to discover the genetic regulatory mechanisms that drive pattern formation during early biological development. Using known mathematical models of these processes, consisting of coupled ordinary differential and partial differential equations, we aim to identify the model parameters that describe the biological mechanisms at play.

To guide learning, we use raw gene expression data sampled from the model organism Drosophila melanogaster, a fruit fly, which is normalized through a series of processing steps before learning. Our learning method applies the powerful techniques of algorithmic differentiation and gradient descent that underlie deep-learning advances.

The results of this study reveal multiple genetic regulatory solutions capable of producing genetic expression patterns that match those observed in the fruit fly embryo. Cluster analysis of these solutions identifies a set of discrete genetic regulatory networks that more closely match those that function in the actual embryo.

Checksum

ac023836e37be4cc681f2a7c656e5dc1

Share

COinS