Coverage Score: A Model Agnostic Method to Efficiently Explore Chemical Space.

Area of Science:

• Computational Chemistry
• Drug Discovery Informatics
• Machine Learning in Chemistry

Background:

• Effective compound selection is crucial for efficient drug discovery.
• Existing methods like clustering and diversity sampling have limitations in maximizing information gain.
• Active learning often requires computationally intensive batch selection.

Purpose of the Study:

• Introduce Coverage Score, a novel subset-based selection method.
• To balance representation and diversity for selecting maximally informative compound subsets.
• To evaluate Coverage Score against existing model-dependent and model-independent techniques.

Main Methods:

• Coverage Score combines Bayesian statistics and information entropy.
• The method allows influence from prior selections and desired properties.
• Comparison of subsets selected by Coverage Score against random and diversity selections across multiple datasets.

Main Results:

• Coverage Score consistently selects subsets yielding more accurate predictions in drug-like chemical space.
• Random Forest models built on Coverage Score subsets showed up to 12.8% lower root-mean-square-error than random selections.
• Selected subsets retained up to 99% of the structural dissimilarity compared to diversity-based selections.

Conclusions:

• Coverage Score offers a superior approach for selecting informative compound subsets in drug discovery.
• The method effectively balances representation and diversity, improving predictive model performance.
• Coverage Score provides a computationally efficient and effective alternative to existing selection strategies.