Year: 2026
Type: dissertation
Abstract: Recent advances in deep learning, combined with the growing availability of biomedical data, have enabled the development of complex models that approximate the processes underlying biology and disease. These models are increasingly used to predict outcomes under hypothetical perturbations to observed data. For example, they can estimate the impact of alternative public health policies on disease spread or predict how genetic mutations affect RNA function. High predictive performance in this task is particularly important in biomedical settings, where empirical testing of alternative scenarios is often infeasible due to ethical or financial constraints. In this dissertation, I present two biomedical applications where deep learning models are used to predict outcomes under hypothetical perturbations, with a focus on sequence-based tasks. First, I describe PAN-cODE, a Latent ODE model for forecasting epidemiological caseloads under alternative public health policies. Second, I detail the pre-training of Orthrus, a Mamba-based mRNA foundation model capable of predicting mRNA properties in response to genetic perturbations. While both approaches offer high accuracy on standard predictive tasks and realistic predictions in response to perturbation, formally grounding these predictions within a causal inference framework remains an important open challenge. To address this methodological gap, I introduce StrAF, a general method for encoding arbitrary conditional independencies into autoregressive normalizing flows. Leveraging StrAF, I obtain causally valid counterfactual queries that surpass existing baselines in accuracy, providing a principled framework for assessing hypothetical perturbations.
Source: TSpace (University of Toronto)
Author: Shi, Ruian
Institution: University of Toronto
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7204019411
