About the project
Motivated by the need to determine the right therapy for each patient, based on the patient's unique genomic profile and conditions, we develop new models and hybrid methods that combine mechanistic knowledge available on the disease and therapy with the need to capture unknown interactions and non-linear effects. Phrased as a prediction task (will this therapy work or not for this patient?, say), we construct a digital twin of the patient, diseases and therapy, that can be used in "what-if" mode. This synthetic copy of the patient is build in part by exploiting available knowledge (biological, clinical, physical, chemical etc) through the mathematics of differential equations, cellular automata, stochastic processes, etc, and in part by modelling interactions agnostically by means of neural nets, random forests, hierarchial models, etc.
For example, we are working on a new mathematical model which accurately predicts patient outcome following CDK4/6i plus endocrine therapy that marks a step toward more personalized treatments in patients with Luminal B breast cancer. Our mechanistic mathematical model accounts for protein signaling and drug mechanisms of action and is trained on extensive, publicly available data from breast cancer cell lines. The model allows to determine a patient-specific response score based on the expression of six genes (CCND1, CCNE1, ESR1, RB1, MYC, and CDKN1A). The model was validated in five independent cohorts of 148 patients in total with breast cancer treated with endocrine therapy and CDK4/6i.
Another recent work consists in predicting the space-time changes of a tumour under treatment. This is challenging due to high inter-patient heterogeneity, unclear importance of available biomarkers, and sparse and noisy longitudinal measurements. Mathematical models of longitudinal measurements can summarize cancer dynamics by means of a few interpretable model parameters. By using machine learning methods to predict the model parameters using baseline covariates, the relationships between the covariates and cancer dynamics under treatment can be discovered. However, this two-step approach does not account for the uncertainty of model parameter estimates in the first step. Instead, we suggest hierarchical Bayesian modeling to model the relationships from baseline covariates to longitudinal measurements through mechanistic parameters while incorporating the uncertainty of each part of the model.
Several new projects are underway, all merging knowledge via mechanistic models and machine learning by more data driven models.
Project team
Associate Professor Alvaro Köhn-Luque, University of Oslo, Institute of Basic Medical Sciences
Professor Arnoldo Frigessi, University of Oslo
Researcher Youness Azimzade, University of Oslo, Institute of Basic Medical Sciences
Researcher Leonard Schmiester, University of Oslo, Institute of Basic Medical Sciences