Coupling Matched Molecular Pairs with Machine Learning for Virtual Compound Optimization.

Area of Science:

• Medicinal Chemistry
• Computational Chemistry
• Drug Discovery

Background:

• Matched molecular pair (MMP) analysis is crucial for understanding structure-activity relationships (SAR) in compound optimization.
• Traditional statistical methods for MMP analysis can be augmented by machine learning (ML) to predict novel compounds.

Purpose of the Study:

• To introduce and evaluate a novel MMP/ML method for automated SAR decomposition and prediction.
• To assess the prediction capabilities and model transferability of the MMP/ML approach in different compound optimization scenarios.

Main Methods:

• Developed a fragment-based MMP implementation integrated with various machine learning algorithms.
• Tested the method on two scenarios: 'new fragments' and 'new static core and transformations'.
• Utilized deep neural networks (DNNs) as a primary ML approach for prediction.

Main Results:

• All tested ML methods showed good performance, particularly in the 'new fragments' scenario.
• Deep neural network models demonstrated superior performance, enabling reliable predictions even in the 'new static core and transformations' scenario with limited SAR knowledge.
• Models trained on broader datasets exhibited enhanced generalizability and could predict beyond the training data's chemical space.

Conclusions:

• The integration of MMP analysis with deep neural networks offers a powerful and promising strategy for high-quality predictions in drug discovery.
• This MMP/ML approach facilitates automated SAR decomposition and prediction, aiding in the exploration of novel chemical entities and optimization projects.