2026 Theses Doctoral

A Multi-resolution Perspective of Modeling Non-linear Effects on Transcriptomic Cell State

Hong, Justin J.

Single-cell RNA sequencing technologies have advanced exponentially, now profiling tens of millions of cells with spatial and perturbational capabilities. Yet, analyzing these massive datasets remains challenging due to technical noise, complex biological interactions, and limited interpretability.

This thesis develops a suite of scalable, interpretable machine learning methods to understand transcriptomic cell states across three distinct biological scales. I introduce: (1) Stable Differentiable Causal Discovery (SDCD), a framework for inferring gene regulatory networks from Perturb-seq data at scale; (2) Attention-Mechanism Interpretation of Cell-cell Interactions (AMICI), a model for identifying cellular interactions in spatial transcriptomics using interpretable attention scores; and (3) Multi-resolution Variational Inference (MrVI), a hierarchical latent-variable model for meta-analysis across multi-patient datasets, capable of capturing both shifts in cell-type abundance and local sample-specific variation.

Together, these methods address the critical need for computational approaches that scale with growing data availability while providing interpretable insights ranging from intracellular gene regulation to patient-level disease patterns, enabling actionable understanding of cellular states across biological scales.