Doctoral Defense - Jiayuan Ding

About the Event

The Department of Computer Science & Engineering

Michigan State University

Ph.D. Dissertation Defense

July 2nd, 2025 at 1:30pm EST

EB 3105 & https://msu.zoom.us/j/6508740740

Passcode: Upon Request from Vincent Mattison or Advisor

From Task-Specific Tools to Foundation Models in Single-Cell Genomics

By: Jiayuan Ding

Advisor: Dr. Jiliang Tang

Single-cell transcriptomics enables high-resolution profiling of gene expression at the level of individual cells, offering unprecedented insights into cellular heterogeneity, development, and disease. However, the unique characteristics of single-cell data, including high dimensionality, sparsity, and technical noise, introduce significant computational challenges. To address these, I first develop task-specific deep learning models tailored for spatial transcriptomics. Notably, GNNDECONVOLVER leverages Graph Neural Networks to perform cell type deconvolution without relying on scRNA-seq reference data, while FOCUS introduces a semi-supervised graph contrastive learning framework that models the subcellular distribution of RNA transcripts to enhance spatial cell type annotation. To facilitate systematic benchmarking and broader usability, I develop DANCE, a unified deep learning library and benchmark platform for single-cell analysis. Building on this foundation, DANCE 2.0 further addresses the critical need for interpretable and automated preprocessing, transforming the traditionally opaque, trial-and-error approach into a transparent, data-driven workflow. Inspired by the recent success of large language models, I propose Tabula, a novel foundation model for single-cell genomics that preserves the tabular structure of the data while enabling self-supervised learning across distributed datasets with built-in privacy safeguards. Unlike prior models that repurpose NLP techniques, Tabula is specifically designed for the structure and scale of single-cell data. Collectively, these contributions trace a cohesive trajectory from narrowly focused, task-specific tools to a general-purpose, privacy-preserving foundation model, laying the groundwork for unified, scalable, and interpretable single-cell analysis.