A large descriptor set and a probabilistic kernel-based classifier significantly improve druglikeness classification.

Area of Science:

• Computational chemistry
• Medicinal chemistry
• Drug discovery

Background:

• Developing accurate predictive models for molecular druglikeness is crucial for efficient drug discovery.
• Existing methods for predicting druglikeness often require refinement for improved accuracy.

Purpose of the Study:

• To develop and validate a highly accurate druglikeness prediction model using machine learning.
• To identify characteristic molecular features indicative of druglike and nondruglike compounds.

Main Methods:

• Utilized a probabilistic support vector machine (SVM) algorithm.
• Employed Extended Connectivity Fingerprints (ECFP_4) for molecular representation.
• Trained and tested the model on large datasets from the World Drug Index (WDI) and Available Chemical Directory (ACD).

Main Results:

• Achieved a high correct classification rate of 92.73% on a dataset of 341,601 compounds.
• Demonstrated significant improvement in prediction accuracy compared to previously published methods.
• Visualized key molecular features distinguishing druglike from nondruglike molecules.

Conclusions:

• The developed probabilistic SVM model with ECFP_4 is a powerful tool for predicting molecular druglikeness.
• The identified characteristic features can serve as valuable guidance for de novo and fragment-based drug design strategies.
• This approach enhances the efficiency of early-stage drug discovery pipelines.