Prediction of compound potency changes in matched molecular pairs using support vector regression

Area of Science:

• Medicinal Chemistry
• Computational Chemistry
• Drug Discovery

Background:

• Matched molecular pairs (MMPs) represent structural transformations within compounds sharing biological activity.
• These transformations can indicate bioisosteric replacements or activity cliffs, crucial for understanding structure-activity relationships.
• Existing methods struggle to predict the magnitude and direction of potency changes associated with these transformations.

Purpose of the Study:

• To develop a predictive model for potency changes encoded by matched molecular pairs (MMPs).
• To differentiate this prediction task from traditional quantitative structure-activity relationship (QSAR) analyses.
• To leverage MMP analysis combined with advanced machine learning for improved prediction accuracy.

Main Methods:

• Introduction of direction-dependent MMPs to capture transformation specifics.
• Integration of MMP analysis with support vector regression (SVR) modeling.
• Exploration of novel kernel functions and molecular fingerprint descriptors within SVR.

Main Results:

• Support vector regression (SVR) models achieved accurate predictions of MMP-encoded potency changes across diverse datasets.
• Shared key structural context was identified as critical for prediction accuracy.
• SVR models demonstrated superior performance compared to random forest (RF) and MMP-based averaging controls.

Conclusions:

• The developed SVR models effectively predict potency changes derived from matched molecular pairs.
• Kernel characteristics significantly influence prediction accuracy, more so than SVR optimization specifics.
• This approach offers a valuable tool for drug discovery by predicting structure-activity relationships.