Image

ABSTRACT
Artificial intelligence, particularly deep neural networks, has increasingly contributed to advances in biological data analysis. However, the high dimensionality, complex patterns, and challenging data distributions inherent to biological data continue to hinder accurate and efficient modeling. During my PhD, I aimed to address these challenges by developing end-to-end deep learning methods for mining predictive patterns from biological data, with the goal of supporting a wide range of tasks in biological and biomedical research.

Biological signals are often represented as images—either in dense forms (e.g., H&E-stained tissue slides) or sparse formats (e.g., mass spectrometry data). Building on this image-based perspective, my research explored predictive modeling problems that fall into two broad categories: mass spectrometry data and tissue images. Each study within these domains targeted a specific prediction task (e.g., flowering time prediction or gene expression prediction), applying deep neural networks to extract informative patterns and improve performance.

My research centered on two main domains:

Mass Spectrometry Data: I explored both dense and sparse image representations to tackle complex tasks such as flowering time prediction and de novo peptide sequencing. By transforming mass spectrometry signals into image-like formats, I developed models that effectively locate and utilize informative spectral regions, leading to improved prediction accuracy and greater adaptability compared to traditional methods.

Tissue Images: To address challenges in histopathological image analysis—such as limited annotations and vast image sizes—I proposed a self-supervised pretraining strategy for large pathology images, improving downstream tasks like gland segmentation. In addition, I developed a comprehensive framework for gene expression prediction from high-resolution whole slide images. This approach integrates:

• Exemplar-based learning to identify and leverage similar regions within tissue,
• Graph neural networks to capture spatial relationships across image regions, and
• Zero-shot learning to generalize gene expression predictions to previously unseen genes using semantic embeddings derived from gene function and phenotype.

Overall, my work establishes a unified deep learning framework for extracting biologically meaningful patterns from both dense and sparse image-based data. By addressing key modeling challenges in mass spectrometry and tissue image analysis, this research advances the scalability, precision, and applicability of deep learning in clinical genomics, computational pathology, and personalized medicine.

BIOGRAPHY
Yan Yang is a PhD candidate at the Biological Data Science Institute, The Australian National University. His research focuses on integrating self-supervised learning, vision-language models, and graph neural networks to develop interpretable and data-efficient methods for histopathological image analysis. He has received several accolades, including the WACV Best Paper Award (Application Track) and a CVPR Oral presentation. With nine A* and eight A-level publications, as classified by the Australian CORE ranking, his work reflects significant contributions to the community.