Joe Bender, Biomedical Engineering/ICM
A major obstacle in oncology drug development is the ineffectiveness of therapies in some patients due to inter-individual tumor heterogeneity. Some molecular biomarkers have proven useful, but these kinds of markers are still unable to accurately predict all responsive and nonresponsive patients. Computational models can aid in the development of predictive biomarkers by providing a source of simulated data that can be used to find robust, multivariate markers. Here, we present a method for generating and testing predictive biomarkers that incorporates patient-specific gene expression data into mathematical models of ligand-receptor interactions and drug pharmacokinetics. We applied the method to the vascular endothelial growth factor (VEGF) family using a model of the transport of two VEGF-A isoforms, two PlGF isoforms, and soluble VEGF-R1 throughout the body, along with the binding of these ligands to receptors (VEGFR-1 and VEGFR-2), co-receptors (Neuropilin-1), and extracellular matrix in normal and tumor tissue. We added population variability by varying the protein production rates according to gene expression data for 5 genes (VEGFA, PGF, FLT1, KDR, NRP1) from the breast, prostate, and kidney cancer TCGA datasets. In this way, we create a population of virtual patient models. We use this population to simulate the interactions with two antibody drugs: one that binds VEGF-A (Avastin), and one that binds NRP1. We used principal component analysis of the population model output to show that untreated baseline receptor binding variation is linearly associated with the input gene expression, but drug response variation has a more complex relationship. In general, the anti-VEGF-A antibody inhibits VEGF-A binding in tumors while enhancing the level of VEGF-A in normal tissues and the blood. The anti-NRP1 antibody blocks VEGF-A and PlGF from interacting with NRP1 in both tumor and normal tissues, but this effect may be negated in a subset of patients by compensatory increases in binding to VEGFR-1 and VEGFR-2. We compared existing gene expression-derived molecular subtypes to the predicted drug response and found several associations, particularly for the triple-negative subtype of breast cancer. We also used partial least squares and support vector machines to build regression models of drug response metrics. We found that the simulated baseline data performed better as inputs to the regression models than the gene expression data by itself. This suggests that the nonlinear transformation applied by the mathematical model to gene expression data creates better predictors of treatment response. Our results provide a framework that can be generalized to other drugs and other diseases; available matched gene expression and clinical outcome data can be used to validate the predicted biomarkers.