Machine Learning for Pharmacokinetic/Pharmacodynamic Modeling.

Area of Science:

• Pharmacokinetics and Pharmacodynamics
• Computational Biology
• Machine Learning in Drug Development

Background:

• Traditional pharmacokinetic/pharmacodynamic (PK/PD) modeling faces challenges with irregularly sampled data and predicting outcomes for novel dosing scenarios.
• Recurrent Neural Networks (RNNs) offer potential for advanced PK/PD modeling due to their ability to handle sequential data.

Purpose of the Study:

• To evaluate novel RNN architectures (ODE-LSTM, Phased LSTM, CTGRU, GRU-D) for modeling and predicting PK/PD data with irregular sampling.
• To assess the capability of these RNNs to extrapolate to unseen dosing regimens and dose levels.

Main Methods:

• Simulated PK/PD data using a one-compartment absorption model and an Indirect PK/PD model I, incorporating inter-individual variability and measurement errors.
• Trained and tested four advanced RNN models on data with 6 or 12 time points per day, including daily (QD) and twice-daily (BID) dosing regimens.
• Evaluated model performance in predicting PK/PD for both familiar and unseen dosing regimens and dose levels.

Main Results:

• All four RNNs successfully modeled QD and BID PK/PD data, accurately capturing fluctuations and return to baseline.
• Extrapolation to twice-daily (BID) dosing regimens based on BID PK data was achieved.
• Predicting PK/PD for dose levels outside the training range proved challenging for most RNNs, with GRU-D showing reasonable performance for 3- and 10-fold higher doses.

Conclusions:

• Novel RNNs, particularly GRU-D, demonstrate significant promise for advancing PK/PD modeling and prediction, especially for irregularly sampled data.
• These models overcome limitations of previous RNNs and highlight the potential of integrating neural networks into PK/PD analysis.
• Further research is needed to improve the extrapolation capabilities of RNNs for significantly unseen dose levels.