Computational Modeling of Drug Response Dynamics Using Patient-Specific Multi-Scale Data

Abstract

Inter-patient variability is one of the biggest challenges of drug response prediction because of multiple scales of complexity biological processes. The combination of patient specific multi-scale multi-genomic, molecular, and clinical data presents a bright opportunity to enhance drug response predictability and achieve precision medicine. The purpose of the study is to create a computational model of the prediction of the dynamics of continuous drug responses (e.g., IC50 values, tumor progression, or biomarker levels, etc.) with the use of patient-specific multi-scale data. A heterogeneous multi-scale computational architecture was created that fuses machine learning and mechanistic modeling in order to incorporate a variety of heterogeneous multi-scale data. The methods used to create informative representations were data preprocessing and feature engineering and dimensionality reduction. Cross-validation strategies were implemented to train and validate the model and the predictive performance of the model estimated based on regression measures such as Root Mean Square Error (RMSE), Mean Absolute Error (MAE) and the coefficient of determination (R 2). The proposed model had better predictive power than baseline methods and was found to have lower values of RMSE and MAE and a higher value of R 2 in a variety of datasets. Also, the incorporation of multi-scale characteristics was crucial in promoting robustness and generalizability of the model when modeling patient-specific drug response dynamics. The resulting computer model is a good predictor of continuous drug response when multi-scale patient data are used, demonstrating that the model could be useful in the future in support of personalized treatment and in the development of computational medicine.