Date of Award

January 2025

Document Type

Thesis

Degree Name

Master of Public Health (MPH)

Department

School of Public Health

First Advisor

Xiting X. Yan

Abstract

Interactions among cells are fundamental to many biological processes, including tissue organization, homeostasis, and disease progression. Understanding these communication networks can illuminate how cells coordinate their functions and how disruptions in these networks may lead to pathology. The advent of spatial transcriptomics (ST) technologies that measure the spatial location and gene expression profile of cells enabled us to study CCC at an unprecedented level of detail. To achieve this goal, many methods have employed graph neural networks with attention mechanisms, treating the attention scores of edges as indicators of the likelihood of communication between corresponding pairs of cells. Despite the intuitive alignment between the model design and the structure of cell-cell communication networks, these methods typically compute attention scores based on a loss function aimed at predicting gene expression. However, it remains unclear whether attention scores optimized for prediction truly reflect communication likelihood. To address this gap in knowledge, this thesis simulates spatial transcriptomic data with known intercellular communication networks as the “ground truth” and uses these simulations to evaluate the performance of various graph neural networks in inferring cell-cell communication networks. Two distinct graph neural network models are evaluated : a GraphTransformer, which leverages global self-attention to capture broad gene expression correlations, and a Graph Attention Network (GAT), which confines attention to local neighborhoods. The results reveal a clear disconnection between gene expression prediction accuracy and cell-cell communication inference accuracy. While the GraphTransformer achieves higher scores for downstream gene prediction, it fails to recover the actual cell–cell communication in the simulated data. By contrast, the GAT model exhibits competitive predictive performance and better recovers the intercellular comminications, likely due to its focus on direct neighbor relationships. These findings highlight the crucial role of simulation studies in exposing whether the attention scores in graph neural networks truly capture biological communication, and underscores the value of rigorous simulation frameworks hen utilizing parameters from attention-based graph neural network for biological interpretation.

Comments

This thesis is restricted to Yale network users only. It will be made publicly available on 06/16/2027

Download

Share

COinS