Machine learning prediction of drug dynamics in dasatinib

A drug’s effect varies from person to person and is strongly influenced by its plasma concentration. Therefore, predicting the time course of plasma drug concentrations is important when investigating drugs. The objective of this study was to predict the time course of plasma concentrations of the anticancer drug dasatinib using information from dasatinib-treated patients, such as age, weight, and time after dose. Pharmacometrics models are conventional methods for predicting the time course of plasma concentrations; however, they often fail to adequately reflect individual differences between patients and are treated uniformly. This leads to the problem of suboptimal treatment for individual patients, resulting in different treatments and side effects for each patient. To overcome these problems, we propose two machine learning-based methods for predicting the time course of plasma dasatinib concentration. The first method uses Support Vector Regression (SVR) with all explanatory variables. To realize more accurate predictions, the second method performs feature selection using the contribution ratio. The performance of the proposed methods was experimentally demonstrated; a drug’s effect varies from person to person and is strongly influenced by the plasma concentration of the drugsed using a dataset of plasma dasatinib concentrations in 18 dasatinib-treated patients. The experimental results confirmed that SVR with all explanatory variables and feature selection achieved a median mean absolute error (MAE) of 19.6% and 19.0%, respectively.