MMPK: A Multimodal Deep Learning Framework to Predict Human Oral Pharmacokinetic Parameters

Area of Science:

• Pharmacokinetics and Drug Metabolism
• Computational Chemistry and Cheminformatics
• Machine Learning in Pharmacology

Background:

• Accurate prediction of in vivo pharmacokinetic (PK) profiles is essential for drug development, impacting safety, efficacy, and dosage optimization.
• Machine learning offers a promising approach to accelerate PK prediction, reducing the time and resources required for drug discovery.
• Existing methods often struggle to capture the complex multiscale molecular information relevant to PK behavior.

Purpose of the Study:

• To develop a novel deep learning framework, MMPK, for predicting human oral pharmacokinetic (PK) parameters.
• To integrate diverse molecular representations, including molecular graphs, substructure graphs, and SMILES sequences, for comprehensive feature extraction.
• To enhance model efficiency and robustness through multitask learning and data imputation techniques.

Main Methods:

• Construction of a large human oral PK dataset comprising over 1,200 compounds and 5,000+ compound-dose combinations.
• Development of the MMPK multimodal deep learning framework integrating graph and sequence-based molecular representations.
• Implementation of multitask learning and data imputation to optimize learning from the PK dataset.

Main Results:

• MMPK demonstrated superior performance compared to baseline models in predicting eight key PK parameters.
• Achieved an average geometric mean fold error (GMFE) of 2.895 and root mean squared logarithmic error (RMSLE) of 0.599.
• The model's effectiveness highlights the benefit of integrating multiscale molecular information for PK prediction.

Conclusions:

• The MMPK framework provides a powerful and accurate tool for predicting human oral PK profiles.
• This approach has significant potential to streamline drug development by improving the efficiency of PK assessments.
• The MMPK model and its underlying dataset are made publicly available to facilitate further research and application.