Pharmacophore features for machine learning in pharmaceutical virtual screening.

Area of Science:

• Computational chemistry
• Cheminformatics
• Machine learning in drug discovery

Background:

• Traditional 3D molecular alignment methods often assign equal importance to all pharmacophore features.
• Identifying the relative importance of specific pharmacophore features is crucial for accurate molecular superposition and prediction.

Purpose of the Study:

• To develop novel molecular features based on the overlap volume of pharmacophore features for improved machine learning models.
• To investigate the impact of weighting pharmacophore features by their overlap volume on predictive performance.

Main Methods:

• Derived overlap volume of pharmacophore features from 3D molecular alignment.
• Incorporated these overlap volumes as weighted features into machine learning models.
• Validated the approach using the DUD-E dataset.

Main Results:

• Machine learning models utilizing overlap volume-based pharmacophore features achieved high performance.
• Median Area Under the Curve (AUC) reached approximately 0.98.
• Recall rate was observed to be nearly 0.8.

Conclusions:

• Weighted pharmacophore features derived from overlap volume offer a significant advancement over traditional methods.
• This approach enhances the accuracy and effectiveness of machine learning models in molecular alignment and drug discovery tasks.