Mechanistic interpretation of tissue foundation models: Patient-by-patient endotyping to uncover disease mechanisms in chronic disease and cancer
Supervisors
Dr Ke Yuan, School of Cancer Sciences, University of Glasgow
Prof John Le Quesne, School of Cancer Sciences, University of Glasgow
Prof Ramanuj DasGupta, School of Cancer Sciences,
Dr Xiao Fu, School of Cancer Sciences, University of Glasgow
Summary
Artificial intelligence foundation models are now trained on extensive tissue repositories measuring multiplexed proteins and spatial transcriptomics. These models are increasingly deployed to suggest biological hypotheses within automated systems. When such a model appears to have discovered a biological pattern, how do we verify it is correct? Current interpretability methods rely on qualitative labels, identifying tissue patches that activate internal neurons and labelling them with descriptive terms like "immune infiltration", which remain fundamentally untestable and risk codifying bias. This project replaces subjective labels with testable hypotheses: compact, biophysically simulated tissue mechanisms (e.g., immune exclusion or hypoxia-driven invasion) governed by physical parameters like cell motility. By simulating tissue across parameter regimes and querying AI models directly, we quantify the evidence for competing mechanisms within each individual patient’s biopsy. We deploy this framework across patient cohorts representing major chronic diseases with substantial socioeconomic health disparities in Scotland: lung adenocarcinoma and progressive liver disease (tracing healthy tissue through MASH to hepatocellular carcinoma). This will reveal patient-specific active processes, test their prognostic value for clinical outcomes, uncover pre-malignant mechanistic transitions to support early intervention in high-risk populations, and establish transparent, equitable clinical AI auditing.
Supervisor(s) for each project will be holding a 30-minute Q&A session during the first two weeks of December.
If you have any questions about a project, you are welcome to attend the relevant session.
If you would like to receive the webinar details, please submit a request via the Webinar Link Request Form.