A chemical class-based approach to predictive model generation.

Area of Science:

• Computational Chemistry
• cheminformatics
• Machine Learning in Drug Discovery

Background:

• Predictive models are crucial for analyzing diverse screening data in drug discovery.
• Current default methods may lack interpretability and reliability for complex datasets.

Purpose of the Study:

• To quantitatively compare a novel class-based modeling approach against a default method.
• To evaluate the interpretability and reliability of predictive models generated by each approach.

Main Methods:

• Recursive partitioning models were generated using both default and class-based approaches.
• Scaffold-based partitioning was employed for the class-based method.
• Models were validated on hold-out data using consensus scoring and multiple descriptor sets across five random trials.

Main Results:

• Both approaches demonstrated similar predictive performance.
• The class-based approach showed superior interpretability for extracting structure-activity relationships.
• The class-based approach provided enhanced reliability for virtual screening applications.

Conclusions:

• The class-based approach offers significant advantages in model interpretability and reliability over the default method.
• This approach can improve confidence in applying predictive models to novel, unseen data.
• Enhanced insights into structure-activity relationships can accelerate drug discovery efforts.