ORCID

0000-0001-6658-6591

Keywords

mRNA, AI, Geometric Deep Learning, DTI, Representation Learning, fairness

Abstract

The rapid emergence of global health crises, exemplified by the Coronavirus Disease 2019 (COVID-19) pandemic, underscores the profound limitations of conventional drug discovery methods, which are inherently protracted and resource-intensive. A robust, accelerated response framework demands the sophisticated deployment of Artificial Intelligence (AI) and Machine Learning (ML) across the entire therapeutic pipeline. We propose a unified approach centered on \textbf{Structural and Representation Learning} to capture the intricate topological and chemical nuances of biological systems, thereby transforming complex molecular and genomic data into effective vector embeddings suitable for deep learning, enhancing predictive power, and crucially, interpretability. This dissertation details advancements across the entire therapeutic development cycle, strictly following the flow: Drug-Target Interaction (DTI) Drug-Target Affinity (DTA) De Novo Drug Generation mRNA Sequence Modeling. For the initial stages, we developed structure-augmented, attention-based models, such as \textit{AttentionSiteDTI} and \textit{FragXsiteDTI}, to refine repurposing efforts by integrating predicted binding site information and fragment-level interpretability. Moving to novel design, we established generative pipelines, including, leveraging diffusion models optimized via Multi-Objective Optimization (MOO) to synthesize novel, high-quality, target-specific ligands for SARS-CoV-2 targets. Finally, we addressed the design of next-generation prophylaxis by introducing advanced language models, such as \textit{Equi-mRNA}, which explicitly encode hierarchical and group-theoretic biological priors (e.g., equivariant codon symmetry) to optimize mRNA stability and translational efficiency for vaccine development. The entire framework was validated against the urgent demands of the COVID-19 crisis, providing a comprehensive toolkit for rapid drug repurposing, \textit{de novo} antiviral design, and accelerated mRNA vaccine optimization. We further generalize the utility of MOO through the \textit{FairBiNN} framework to manage trade-offs inherent in translational data science, such as balancing predictive accuracy with fairness metrics. This work presents a substantial contribution to computational drug design, demonstrating that explicitly encoding structural and biological knowledge through advanced representation learning techniques is paramount for building robust, interpretable, and generalizable AI systems capable of accelerating pandemic response.

Completion Date

2025

Semester

Fall

Committee Chair

Ozmen Garibay, Ozlem

Degree

Doctor of Philosophy (Ph.D.)

College

College of Engineering and Computer Science

Department

Computer Science

Format

PDF

Identifier

DP0029720

Document Type

Thesis

Campus Location

Orlando (Main) Campus

Download

Share

COinS