Scaffold-based evaluation metrics for fair comparison of molecular generators

Area of Science:

• Medicinal Chemistry
• Computational Chemistry
• Drug Discovery

Background:

• Molecular generators explore chemical space for novel compounds.
• Current metrics for molecular generators focus on validity and novelty, not biological activity.
• Evaluating generator performance is challenging due to molecule diversity and volume.

Purpose of the Study:

• Introduce novel scaffold-based metrics for evaluating molecular generators.
• Enable fair comparison of molecular generators for drug discovery applications.
• Address limitations of existing metrics in assessing biological relevance.

Main Methods:

• Developed scaffold Recovery Score (RS), SEt scaffold Diversity (SED), and Absolute SEt scaffold Recall (ASER) metrics.
• Applied metrics to compare Molpher, DrugEx, REINVENT, and a Graph-based genetic algorithm.
• Utilized scaffold recovery from input sets to assess generator performance.

Main Results:

• Scaffold-based metrics provide a realistic framework for generator evaluation.
• Metrics facilitate comparison of generators for drug discovery.
• Scaffold recovery is a key indicator of biological relevance.

Conclusions:

• Scaffold-based metrics improve the evaluation of molecular generators.
• These metrics are crucial for optimizing generators in drug discovery.
• The proposed metrics aid in designing focused virtual chemical libraries.