A multimodal contrastive learning framework for predicting P-glycoprotein substrates and inhibitors.

Area of Science:

• Pharmacology and Toxicology
• Computational Chemistry
• Drug Discovery

Background:

• P-glycoprotein (P-gp) significantly impacts drug ADMET properties.
• Accurate prediction of P-gp inhibitors/substrates is vital for drug development.
• Existing models lack comprehensive molecular information, limiting predictive accuracy.

Purpose of the Study:

• To develop an advanced computational model for predicting P-gp inhibitors and substrates.
• To overcome limitations of existing models by utilizing multimodal molecular data.
• To enhance the accuracy and reliability of P-gp related drug assessments.

Main Methods:

• Compiled a large dataset of 5,943 P-gp inhibitors and 4,018 substrates.
• Developed a multimodal graph contrastive learning (GCL) model (MC-PGP).
• Integrated SMILES, molecular fingerprints, and graph features using attention-based fusion and GCL.

Main Results:

• MC-PGP achieved superior performance compared to 12 state-of-the-art methods.
• Demonstrated AUC-ROC improvements of 9.82% for inhibitors and 10.62% for substrates on external datasets.
• Interpretability analysis revealed key functional groups involved in P-gp interactions.

Conclusions:

• MC-PGP offers a significant advancement in predicting P-gp inhibitors and substrates.
• The model provides valuable, chemically intuitive insights for drug design.
• This approach aids in optimizing drug candidates and improving toxicological assessments.