Active Learning FEP: Impact on Performance of AL Protocol and Chemical Diversity

Area of Science:

• Computational chemistry
• Machine learning in drug discovery

Background:

• Active learning using free energy perturbation (AL-FEP) is a strategy for developing machine learning models to predict biochemical potency.
• This approach is particularly useful in early-stage lead optimization when experimental data is scarce.

Purpose of the Study:

• To evaluate the performance of AL-FEP in generating predictive models for bromodomain inhibitor series.
• To retrospectively analyze the AL-FEP workflow parameters for routine application in drug discovery projects.

Main Methods:

• AL-FEP was applied to two distinct bromodomain inhibitor series, varying whether the core structure was constant or included in compound idea generation.
• Model performance was assessed using measured biochemical potency data.
• A retrospective evaluation covered compound selection strategies, explore-exploit ratios, and cycle-based compound numbers.

Main Results:

• Well-performing predictive models were generated within a few active learning cycles, particularly when the core structure remained constant.
• Significant performance variations (model enrichment, R-squared) were observed based on workflow parameters.
• The study identified specific parameter recommendations for AL-FEP based on deployment context (potency maximization vs. broad accuracy).

Conclusions:

• AL-FEP is a viable method for generating predictive models in early-stage drug discovery.
• Optimizing AL-FEP workflow parameters is crucial for maximizing model performance and achieving project-specific goals.
• The findings provide guidance for the routine implementation of AL-FEP in medicinal chemistry projects.