Model-based Bayesian inference of cancer dynamics from heterogenous longitudinal data

Abstract

The kinetic parameters of cancer population dynamics are critical for developing reliable predictors of tumour growth patterns, extracting metrics for patient stratification and creating algorithms that can forecast clinically significant events. Here, we introduce a model-based Bayesian framework that leverages longitudinal phenotypic (e.g., tumour volume, cell counts) or genotypic (e.g., mutation frequency) data to infer critical parameters of tumour progression within a single patient. Our models uses population genetics to estimate probability distributions for tumour growth rates, initiation and extinction times, pinpointing abrupt shifts in tumour dynamics due to treatment response and revealing associations between drug resistance and pre-existing cancer cell populations. We apply our framework to address pivotal clinical questions across three major cancer types. In colorectal cancer, we use tumour markers data to identify extensive pre-existing RAS-linked resistance to cetuximab. In lung cancer, we use somatic mutation frequencies in citculating tumour DNA to determine prognostic growth rates and develop a test for monitoring minimal residual disease. In chronic lymphocytic leukaemia, we use white blood cell counts to stratify patients by growth patterns and predict time to treatment, advancing adaptive monitoring strategies.