Probabilistic Approach for Virtual Screening Based on Multiple Pharmacophores.

Area of Science:

• Computational Chemistry
• Cheminformatics
• Machine Learning in Drug Discovery

Background:

• Pharmacophore modeling is traditionally a non-probabilistic virtual screening technique.
• Molecular correspondence to a pharmacophore is typically the sole indicator of bioactivity.

Purpose of the Study:

• To introduce a probabilistic framework for pharmacophore modeling.
• To develop and evaluate new consensus prediction schemes for pharmacophore-based virtual screening.

Main Methods:

• Pharmacophores were conceptualized as one-class machine learning models.
• Two probability calculation schemes (Max and Mean) were proposed for consensus predictions.
• Approaches were benchmarked on ChEMBL datasets and validated on DUD-E datasets.

Main Results:

• The proposed Max and Mean approaches provide probabilistic confidence scores for pharmacophore models.
• These methods enable ranking of compounds identified by multiple pharmacophore models.
• Superior early enrichment was observed compared to common consensus approaches.

Conclusions:

• A probabilistic, well-performing, and easy-to-implement alternative for pharmacophore-based virtual screening has been developed.
• The new methods enhance the utility of pharmacophore models in drug discovery pipelines.