Date of Award

4-3-2024

Publication Type

Thesis

Degree Name

M.Sc.

Department

Computer Science

Keywords

Cell-cell interaction prediction;Feature selection;Graph convolutional neural network;LSTM;Spatial transcriptomics

Supervisor

Luis Rueda

Abstract

Several studies are available that use gene expression data to infer cell-cell inter-actions. Nevertheless, most of these studies target intra-cellular interactions. The advent of spatial expression data paves the way for methodologies capable of deducing interactions, spanning both intra- and inter-cellular domains. However, spatial data also presents new challenges, including noisy and high-dimensional data and sparse representation. We propose a new model based on a graph neural network to predict cell-cell interaction from spatial data. Specifically, the study constructs a graph from the spatial data, forming the foundation for the model that combines the ability of Long Short-Term Memory (LSTM) and Graph Neural Network (GNN). The model’s unique ability capitalizes on LSTM’s sequence learning and GNN’s graph-based potential, designed to predict links within the spatial context. The model exhibits enhanced predictive capabilities through rigorous testing compared to simi-lar approaches. Our investigation demonstrates that integrating our pipeline with the backward search technique yields the highest area under the curve (AUC) score. Furthermore, we have conducted a comparative analysis, juxtaposing this performance against two alternative approaches, SEAL (learning from Subgraphs, Embeddings and Attributes for Link prediction) and GCNG (Graph Convolutional Neural Networks for Genes). Our results demonstrate that integrating our pipeline with the backward search technique yields the highest AUC score. The effectiveness of our approach is validated on two well-known datasets, seqFISH+ and Merfish, which capture the spatial intricacies of cellular communication.

Download

Share

COinS