Combining Machine Learning and Molecular Dynamics to Predict P-Glycoprotein Substrates.

Area of Science:

• Pharmacology
• Computational Chemistry
• Drug Discovery

Background:

• P-glycoprotein (P-gp) is a key efflux transporter limiting drug bioavailability.
• Accurate prediction of P-gp substrates is crucial for early-stage drug discovery.
• Existing in silico models rely on structural and physicochemical descriptors.

Purpose of the Study:

• To evaluate molecular dynamics fingerprints (MDFPs) as novel descriptors for P-gp substrate classification.
• To compare the performance of MDFP-based machine learning models against traditional descriptors.
• To assess the generalization ability of models on diverse chemical spaces.

Main Methods:

• Generation of MDFPs from short molecular dynamics simulations in various environments.
• Training and evaluation of machine learning classifiers using MDFPs and traditional 2D descriptors.
• Validation using in-house and public datasets (ChEMBL), including external and prospective analyses.

Main Results:

• All tested classifiers achieved high accuracy on chemically diverse subsets.
• Tree-based machine learning models demonstrated strong interpolation capabilities.
• Only models trained on MDFPs or property-based descriptors exhibited robust generalization to new chemical spaces.

Conclusions:

• MDFPs offer a valuable orthogonal descriptor for P-gp substrate prediction.
• Tree-based machine learning models combined with MDFPs or property-based descriptors are most effective for reliable in silico drug screening.
• This approach enhances the prediction of drug candidates with improved bioavailability early in the drug discovery pipeline.