Award Date

12-15-2025

Degree Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Computer Science

First Committee Member

Mingon Kang

Second Committee Member

Laxmi Gewali

Third Committee Member

Fatma Nasoz

Fourth Committee Member

Bryar Shareef

Fifth Committee Member

Mira Han

Number of Pages

72

Abstract

Weakly supervised learning (WSL) has emerged as a paradigm for training machine learning models using datasets that are partially labeled, ambiguous, or incomplete. This approach addresses the challenges of acquiring fully annotated data, which is often time-consuming and resource-intensive. However, traditional WSL models face limitations in interpretability due to label ambiguity, limited granularity, data bias, and model complexity, making their application in sensitive domains such as healthcare, bioinformatics, and biological analysis less reliable. The lack of transparency and explainability highlights the need for novel interpretable WSL frameworks. To address these gaps, I propose a series of interpretable WSL frameworks that integrate biological relevance, model transparency, and predictive robustness. The first framework, Multilayered Self-Attention Mechanism (Multi-SA), leverages intermediate feature representations within neural networks to enhance Class Activation Maps (CAM) for weakly supervised semantic segmentation. By incorporating multi-layered self-attention mechanisms, Multi-SA captures complex object patterns from image-level labels, achieving state-of-the-art performance on datasets such as PASCAL VOC 2012. This framework effectively bridges the gap between pixel-level segmentation and image-level annotations, enhancing both interpretability and localization precision. Building on this foundation, the second framework, BIN-PU, introduces a Positive-Unlabeled (PU) learning strategy to predict bacterial compound-protein interactions using only positive samples. BIN-PU mitigates the absence of negative data by generating pseudo-positive and pseudo-negative samples through a binning strategy, enabling stable training and improved generalizability. Extensive experiments demonstrate that BIN-PU outperforms benchmark methods such as PUCPI, making a significant contribution to bioinformatics, particularly in cytochrome P450 interaction prediction and related drug discovery applications. Extending WSL interpretability to cellular-level analysis, we developed the Biologically Interpretable Contrastive and Attention Network using Holotomography (BICAN-HT) for cellular senescence classification. This model integrates empirical supervised contrastive learning with cross-attention mechanisms to capture nucleus–cytoplasm relationships, revealing biologically meaningful patterns consistent with senescence-associated morphological changes. The framework not only achieves high accuracy but also provides an interpretable visualization of cellular dependencies, enabling insight into the biological mechanisms of senescence. Finally, I present a comprehensive review of artificial intelligence applications in cellular senescence research, highlighting key computational advancements, challenges, and future opportunities. This synthesis bridges the gap between AI-based imaging analysis and biological understanding, guiding future directions toward integrative, interpretable, and biologically grounded AI systems. Overall, this dissertation advances the field of weakly supervised learning by embedding interpretability into model design across diverse domains--from visual scene understanding to molecular and cellular biology, enabling actionable, biologically reliable predictions.

Keywords

bacterial CYPs; Cellular senescence; Cytochrome P450; Image segmentation; Positive-unlabeled learning; Weakly supervised learning

Disciplines

Artificial Intelligence and Robotics | Bioinformatics | Computer Engineering

File Format

PDF

File Size

3100 KB

Degree Grantor

University of Nevada, Las Vegas

Language

English

Rights

IN COPYRIGHT. For more information about this rights statement, please visit http://rightsstatements.org/vocab/InC/1.0/

Download

Available for download on Tuesday, December 15, 2026


Share

COinS