Machine Learning Modeling for ABC Transporter Efflux and Inhibition: Data Curation, Model Development, and New...

Area of Science:

• Pharmacology
• Computational Chemistry
• Biochemistry

Background:

• Machine learning models are increasingly used to predict ATP-binding cassette (ABC) transporter interactions.
• Previous models often suffered from limited applicability due to small training datasets.

Purpose of the Study:

• To curate a comprehensive database of ABC transporter bioactivity data.
• To develop and validate robust quantitative structure-activity relationship (QSAR) models for predicting substrate binding and inhibition of key ABC transporters.

Main Methods:

• Manual curation of over 24,000 bioactivity records for P-gp, BCRP, MRP1, and MRP2 from literature and databases.
• Development of QSAR models using eight datasets, four machine learning algorithms, and three chemical descriptor sets.
• 5-fold cross-validation and external validation using DrugBank compounds.

Main Results:

• Eight curated datasets comprising approximately 8800 unique chemicals were generated.
• QSAR models achieved high performance with an average correct classification rate (CCR) of 0.764 for substrate binding and 0.839 for inhibition.
• Model predictions correlated with xenobiotic brain exposure, with predicted P-gp and BCRP substrates showing reduced brain penetration.

Conclusions:

• A large, curated database for ABC transporter computational modeling has been established.
• Validated QSAR models can accurately predict transporter substrate binding and inhibition.
• These models can inform predictions of drug distribution, including brain exposure and tissue penetration.