Back to search

Georgia Institute of Technology

Modeling Single-Cell Data: Recursive Clustering for Resolving Cellular Heterogeneity in scRNA-seq and Label-Efficient Deep Learning for Subcellular Protein Localization in Fluorescence Microscopy

Abstract

dc:description.abstract

Single-cell technologies enable detailed characterization of cellular heterogeneity but generate large, noisy, high-dimensional datasets that are difficult to analyze. This dissertation develops computational methods for two single-cell data modalities: single-cell RNA sequencing (scRNA-seq) and protein localization in fluorescence microscopy. For scRNA-seq, this work introduces a recursive clustering strategy that iteratively performs dimensionality reduction and clustering within previously identified cell clusters. This approach is designed to identify fine-grained cell types that may not be resolved using features selected from the full dataset. Across four benchmark scRNA-seq datasets, recursive clustering improved cell-type detection relative to an equivalent single-pass clustering procedure, particularly for detailed cell-type annotations and higher-resolution clustering, while also improving clustering stability. These improvements were robust across clustering algorithms, feature-selection settings, and evaluation metrics. Applied to Crohn's disease scRNA-seq data, recursive clustering also produced a hierarchical representation that supported analysis of disease-associated changes in cell abundance and gene expression at multiple levels of cluster granularity. For fluorescence microscopy, this dissertation develops two label-efficient approaches for single-cell subcellular protein localization. The first adapts self-supervised learning to microscopy using tailored image augmentations and signal-aware crop and mask sampling that preserve sparse protein localization signals. The second introduces a weakly supervised multi-view, multi-branch multiple-instance learning architecture that uses field-of-view labels and cell segmentation masks to combine information across image scales and prediction pathways. Both approaches improved single-cell protein localization performance, with the multi-view multi-branch model outperforming previously published methods on the Human Protein Atlas Single-Cell Classification (HPA-SCC) public benchmark. Together, these studies show the value of designing computational methods around the structure of the biological data being analyzed. As single-cell datasets continue to grow in scale and complexity, approaches that account for this structure will be increasingly important for translating high-dimensional measurements into biological insight.

Degree

thesis:*
Name thesis:degree_name
Computer Science, PhD
Department thesis:degree_department
Computational Science and Engineering
Grantor
Georgia Institute of Technology
Year dc:date.issued
2026

Author and committee

dc:creator, dc:contributor.*
Author dc:creator
  • Squires, Michael Francis
Advisor dc:contributor.advisor
  • Zhang, Xiuwei
Committee members dc:contributor.committeemember
  • Qiu, Peng
  • Cherry, Elizabeth
  • Sinha, Saurabh
  • Luo, Yunan
ORCID person:identifier.orcid

Identifiers

dc:identifier.*
Handle dc:identifier.uri
https://hdl.handle.net/1853/82142

Chain of custody

source
Harvested from
Georgia Tech
Base URL
repository.gatech.edu/server/oai/request
Last updated
2026-11-19
Source record
OAI-PMH GetRecord
related terms
citation

Squires, Michael Francis. Modeling Single-Cell Data: Recursive Clustering for Resolving Cellular Heterogeneity in scRNA-seq and Label-Efficient Deep Learning for Subcellular Protein Localization in Fluorescence Microscopy. Georgia Institute of Technology, 2026. https://hdl.handle.net/1853/82142