Who’s in and who’s out: leveraging homogeneous preclinical data to extrapolate tumour growth outcomes across heterogeneous populations

High failure rates in preclinical and clinical studies remain a major obstacle in anti-cancer drug development. A key factor is the lack of heterogeneity in preclinical models, which typically use genetically identical mice or monoclonal cell lines that fail to reflect real-world variability. Additionally, preclinical data are often aggregated, obscuring important individual-level insights. Here, we introduce a computational framework specifically designed to address these challenges. Using a lung cancer xenograft experiment reporting averaged tumour volume and Kaplan-Meier survival data as a case study, we reconstruct virtual clones via Bayesian inference, grounded in a minimal modelling framework that uses established ordinary differential equations to simulate tumour growth. A key innovation is the explicit mechanistic linkage between tumour dynamics and individual survival probabilities. The reconstructed clones show excellent agreement with experimental data. We then apply standing variations modelling to generate heterogeneous virtual cohorts not included in the original study. These cohorts accurately recapitulate independent xenograft experiments not used in model calibration, thereby validating our approach. By capturing realistic variability at the preclinical stage, our method offers a practical framework to improve drug development pipelines, reduce costly experimental iterations and identify rare subpopulations most and least likely to benefit from treatment.