Title: Machine Learning Enhanced Feature Extraction and Biomarker Identification from Spatial Transcriptomics Data.

Date: April 13th, 2026

Time: 1:00pm - 3:00pm

Location: Krone Engineered Biosystems Building (EBB), room 1005, Children’s Healthcare of Atlanta Seminar room (First floor CHOA Room)

Zoom: https://gatech.zoom.us/j/91470654803?pwd=YyPkJ2xSurahERwLQR5tpoeEa7DP4p.1&from=addon

Zhou Fang

Machine Learning PhD Student

Department of Biomedical Engineering
Georgia Institute of Technology

Committee

1 Dr. Ahmet F. Coskun (Advisor), Department of Biomedical Engineering, Georgia Institute of Technology

2 Dr. Peng Qiu, Department of Biomedical Engineering, Georgia Institute of Technology

3 Dr. Cassie Mitchell, Department of Biomedical Engineering, Georgia Institute of Technology

4 Dr. Sara McCoy, Department of Rheumatology, University of Wisconsin, Madison

5 Dr. Xiuwei Zhang, School of Computational Science and Engineering, Georgia Institute of Technology

Abstract

Advances in spatial transcriptomics technologies have generated rich molecular data at high spatial resolution, revealing complex spatial organizations of biological samples. Current methods fail to balance capturing data complexity and result interpretability. This thesis presents three graph-based methods aimed at bridging the gap. First, Spatially Resolved Gene Neighborhood Network (SpaGNN) quantifies pairwise subcellular colocalization relationships among genes, producing spatially resolved features that outperform single-cell ribonucleic acid (RNA) expression in distinguishing similar cells. Second, 3-Dimensional Spatially Resolved Gene Neighborhood Network with Embedding (3D-SpaGNN-E) extends the SpaGNN pipeline to 3D spatial transcriptomics data and introduces a graph autoencoder to model relationships among subcellular regions, enabling the identification of cell-cell communication sites. The last part applies a graph attention network to characterize cellular neighborhoods in subtypes of Sjögren’s disease, learning a disease-relevant latent representation. Analysis of the latent space identified distinct cellular neighborhoods and gene expression associated with disease states. Together, these approaches represent a suite of graph-based frameworks for analyzing the spatial organization of biological samples while maintaining interpretability.