Cancer Calibration — Transformer Posterior Estimator (TPE)
Amortised Bayesian calibration of cancer microsimulation models using transformers.
Joint work led by Prof. Nora Pashayan (PI) at the Centre for Cancer Genetic Epidemiology, University of Cambridge. The project grew out of the CRUK ACED ABIDE Feasibility Study (Addressing the Base Error Impeding the Development of Biomarkers for Early Detection), led by Prof. Shonit Punwani (UCL) — an international collaboration across Cambridge, UCL, Stanford, Manchester, and Oregon.
The Transformer Posterior Estimator (TPE) is a transformer-based amortised posterior estimator with age-window masking — a capability unavailable in conventional sequential methods.
On a 48-cell oracle benchmark covering four cancers, three model complexities (4/6/8-state), and four methods (IMABC, BayCANN, PNPE, TPE), TPE matches or beats sequential methods at one-shot amortised cost: a single training run versus per-cell calibration.
The work also delivers a cross-method benchmark showing that identifiability is a model–data property, consistent across inference methods on the same oracle data — not a method artefact, and that identifiability degrades systematically with model complexity.
Paper in preparation.