Prediction of Promiscuity Cliffs Using Machine Learning

Area of Science:

• Medicinal Chemistry
• Computational Drug Discovery
• Pharmacology

Background:

• Polypharmacology, enabled by promiscuous compounds, is a growing area in drug discovery.
• The molecular basis of compound promiscuity remains poorly understood.
• Structural analogs with varying promiscuity are frequently identified.

Purpose of the Study:

• To elucidate the molecular origins of drug promiscuity.
• To develop predictive models for compound promiscuity using machine learning.
• To compare the predictive power of compound pairs versus individual compounds.

Main Methods:

• Systematic extraction of matched molecular pairs (MMPs) with varying promiscuity from public data.
• Elimination of compounds with potential false-positive activity annotations.
• Development and comparison of machine learning models for predicting promiscuity cliffs (PCs) and non-PC MMPs.

Main Results:

• Promiscuity cliffs (PCs) contain richer structure-promiscuity relationship information than individual promiscuous compounds.
• Machine learning models trained on PCs outperformed models using individual compounds for promiscuity prediction.
• Feature analysis identified key molecular descriptors contributing to promiscuity prediction.

Conclusions:

• Analyzing compound pairs (PCs) is a more effective strategy for understanding and predicting drug promiscuity.
• This approach enhances the understanding of structure-promiscuity relationships.
• The findings support the development of more targeted polypharmacological drugs.