UCLA

Graphs are a universal language for describing systems composed of interacting elements. In biomedicine, graph-structured data are pervasive, ranging from molecular and cellular interaction networks to healthcare knowledge graphs. Graph neural networks (GNNs) have initiated a new paradigm for graph representation learning. Moreover, the Transformer model, conceptually a specialized form of GNN, has been pushing the boundaries in the most impactful domains of machine learning over the past half-decade. These domains range from language modeling and computer vision to computational biology. However, the complex nature of biomedical data poses significant technical challenges for graph learning, including heterogeneity, multi-scale interactions, and limited data annotations. This dissertation attempts to address these obstacles across various biomedical applications, including pathology image analysis and heterogeneous data integration. We first introduce a bipartite graph convolutional network for drug repurposing through heterogeneous information fusion. Additionally, we present a temporal graph learning method that leverages heterogeneous electronic health record (EHR) to improve depression prediction. We further advance graph deep learning techniques for histopathology image analysis across multiple magnification levels, including patient-level survival prediction and tissue-level tumor microenvironment modeling. Lastly, inspired by recent advances in vector-quantized image modeling and Generative Pretrained Transformer (GPT), we propose a self-supervised learning framework tailored for computational pathology. Our approach explores pretraining and fine-tuning strategies to improve generalizable pathology image representation learning, thus facilitating the development of data-efficient deep learning models for various clinical workflows. Together, these methods demonstrate the versatility and effectiveness of graph deep learning in addressing complex biomedical challenges, paving the way for real-world applications in disease modeling, diagnosis, and treatment optimization.