Predicting phospholipidosis using machine learning.

Area of Science:

• Drug Discovery and Development
• Toxicology
• Computational Chemistry

Background:

• Phospholipidosis (PLD) is a cellular disorder caused by cationic amphiphilic drugs, characterized by phospholipid accumulation and lamellar inclusion bodies.
• PLD can impede drug development, highlighting the need for accurate predictive computational methods.
• Previous studies indicate machine learning models achieve high performance in predicting PLD.

Purpose of the Study:

• To enhance the prediction of drug-induced phospholipidosis using an expanded dataset.
• To compare the efficacy of different molecular descriptors (circular fingerprints, E-Dragon) and machine learning models for PLD prediction.

Main Methods:

• Utilized a larger dataset of drug-induced phospholipidosis cases mined from scientific literature.
• Employed machine learning models, including Random Forest and Support Vector Machines.
• Evaluated the performance of circular fingerprints and E-Dragon descriptors, individually and in combination.

Main Results:

• Circular fingerprints significantly outperformed E-Dragon descriptors in building predictive models for phospholipidosis.
• No substantial performance difference was observed between Random Forest and Support Vector Machine models.
• The expanded dataset and optimized feature selection improved predictive accuracy.

Conclusions:

• Circular fingerprints are superior to E-Dragon descriptors for developing accurate computational models of drug-induced phospholipidosis.
• Random Forest and Support Vector Machines offer comparable predictive performance for this toxicological endpoint.
• These findings contribute to more efficient drug safety assessment and development.