Date of Award
Summer 8-4-2023
Degree Name
Doctor of Philosophy (PhD)
Degree Type
Dissertation
Abstract
Artificial intelligence has made substantial progress in resolving key issues in a variety of scientific and technical domains, particularly in cross-disciplinary sciences. Deep learning is becoming ubiquitous, particularly in biology, healthcare, and drug discovery. Deep learning techniques offer chances to advance science with the potential to significantly outperform state-of-the-art approaches in many scientific fields. There are, however, notable challenges. Each discipline in which deep learning could be applied has a unique manner of communicating concepts, which we describe as interpretability, critical for use in science, both in industry and academia. Interpretability can be defined as decisions made during visualization generation or presenting a model and results paired with reasoning behind those results. This work closely examines translating deep learning in scientific domains to improve interpretation in distinct domains where deep learning is not widely used. We expect significant gains by adapting deep learning to specific scientific domains. We expect to see deep learning more suited to advancing scientific knowledge, novel deep learning advancements adopted within these scientific domains, and accuracy gains against current state-of-the-art methods used throughout these distinct domains. This research specifically aims to develop novel deep learning methods that assist domain experts in having methods that are: (1) potentially more accurate, (2) interpretable to scientists, and (3) informative, leading to potential discoveries in three scientific domains: bioactivation, quantum chemistry, and clinical studies. First, we examine to what extent subgraph mining algorithms can be used to systematically mine for novel structural alerts. Next, we explore how physical structures can be represented in a way that allows predictions of both low- and high-level properties with deep learning. Finally, we cover how effective deep learning is in clinical studies and how deep learning can improve clinical decision-making. This dissertation shows that computer science methodology can be adapted for domain-specific knowledge and successfully communicate results in domain-specific languages.
Language
English (en)
Chair
S. Joshua Swamidass