Abstract P2-10-29: Time dependent breast cancer metastasis prediction using novel biological imaging, clinico-pathological and genomic data combined with Bayesian modeling to reduce over-fitting and improve on inter-cohort reproducibility.

Abstract Background: Breast cancer heterogeneity demands that prognostic models must be biologically driven and recent clinical evidence indicates that future prognostic signatures need evaluation in the context of early versus late metastatic risk prediction. The aim of our work was to identify biologically validated quantitative imaging parameters with improved correlation to clinical outcome, and to address some of the remaining obstacles for a truly robust prognostic model in clinical use. Method: We identified 4 seed proteins (ezrin/radixin/moesin-cofilin), along with several kinases as biologically relevant subnetwork of proteins that control tumor cell motility and metastasis. Patient-derived breast cancer tumour samples were used to perform a combination of imaging methods such as Fluoresecence lifetime imaging microscopy, automated segmentation and co-localisation intensity analysis. A complexity optimized Bayesian proportional hazard regression model was performed on a total of 419 breast cancer patients to validate time dependent predictions using traditional clinicopathological, genomic and our novel optical imaging-derived parameters. An independent dataset of 300 patient samples from the Leeds Institute of Molecular Medicine is currently being evaluated, representing a large cross centre validation of our integrated model. Results: We demonstrate that the traditional gold standard clinico-pathological variables are poor predictors for patients that survive long periods, and that their predictive significance (in terms of hazard ratios) varies significantly between two temporal cohorts where the adjuvant treatments are vastly different. Moreover, we investigate the predictive accuracy of a combined imaging/clinicopathological model compared with genomic/clinicopathological models. We demonstrate how to reduce over-fitting to help improve the performance of prognostic models. Results of an integrated model combining genomic and imaging parameters are still awaited. Discussion: We have produced the first optical imaging-derived multivariate tumour metastatic signature, which measures underlying key biological variables involved in regulating cancer cell motility. Using Bayesian proportional hazards regression in a time-dependent manner, we highlight the inadequacies of existing prediction tools and present a model combining the clinicopathological parameters with our imaging-based metastatic signature, as an integrative reproducible prognostic tool across different temporal cohorts. Citation Information: Cancer Res 2012;72(24 Suppl):Abstract nr P2-10-29.