Application of Recurrent Neural Networks for Pharmacokinetic Modeling and Simulation

Abstract

Pharmacometrics and the utilization of population pharmacokinetics play an integral role in model informed drug discovery and development (MIDD). Recently, there has been a growth in the application of deep learning approaches to aid in areas within MIDD. In this work, we aim to explore the potential of deep learning approaches towards drug concentration prediction and simulation. A total of 1,527 olanzapine drug concentrations sparsely sampled from 523 individuals along with eleven patient-specific covariates provided by the CATIE studies were used in model development, validation, and simulation. LSTM and LSTM-ANN with multiple inputs were investigated towards olanzapine drug concentration predictions. The LSTM-ANN model captured the relationships within a pharmacokinetic dataset and generated olanzapine drug concentration predictions with a lower RMSE than the LSTM model. Bayesian optimization was implemented to tune the hyperparameters of the LSTM-ANN model. The LSTM-ANN model had a RMSE of 29.566 in the validation set. A population pharmacokinetic model using NONMEM model was constructed as a reference to compare the performance of the LSTM-ANN model. The RMSE of the NONMEM model was 31.129. Permutation importance revealed age, sex and smoking were highly influential covariates in the LSTM-ANN model. The LSTM-ANN model showed potential in drug concentration prediction as it performed comparably to the NONMEM model. Future studies investigating clinical studies with varying sample sizes and sampling strategies are required to further examine the potential of a LSTM-ANN model towards drug concentration prediction. For olanzapine drug concentration simulations, three RNN cells within an RNN-ANN model with multiple inputs were studied. The GRU-ANN model resulted in the optimal RNN-ANN model with the lowest RMSE in the simulation data. Bayesian optimization was implemented to optimize the hyperparameters of the GRU-ANN model. The optimized GRU-ANN model resulted in a simulation RMSE of 24.844. Visual inspection revealed the simulated olanzapine drug concentrations were lower than their respected observed olanzapine drug concentrations. Exploratory data analysis revealed the underperformance may have been a result of dosing levels between the CATIE studies having little overlap. A comprehensive clinical trial study is required to fully explore the potential of a GRU-ANN model towards drug concentration simulations.