Interpretable analytical methods for characterizing disease-associated phenotypic modulation of cells

Abstract

Realizing the analytical potential of single-cell technologies requires computational frameworks that address fundamental challenges in characterizing cellular heterogeneity and disease-associated phenotypes. High-throughput cytometry and spatial omics platforms have advanced substantially in their resolution and throughput, enabling simultaneous measurement of dozens of parameters across millions of cells. However, these technologies present analytical obstacles including technical artifacts, batch effects, and the complexity of spatial interactions that limit biological discovery and clinical application. Addressing challenges such as robust cell type identification, cross-cohort transferability, and spatial variance decomposition necessitates development of innovative computational approaches. To this end, this thesis contributes to single-cell computational biology by (1) establishing a multiview framework that harmonizes correlation and distance metrics through ensemble learning, overcoming similarity metric limitations in imaging cytometry to enable robust identification of cellular phenotypes, (2) developing transferable deep learning approaches combining batch-agnostic normalization with hierarchically-structured feature selection to enable crossinstitutional validation of cytometry-based biomarkers, and (3) creating variance decomposition methods for spatial transcriptomics that quantify cell type-specific interactions while correcting for lateral spillover, revealing how proximity modulates gene expression programs in breast cancer and melanoma. The frameworks outlined in this thesis provide validated computational tools advancing single-cell analysis and establish foundations for future development of interpretable methods characterizing disease-associated cellular modulation.