Impact of Applicability Domains to Generative Artificial Intelligence

Area of Science:

• Computational Chemistry
• Artificial Intelligence in Drug Discovery

Background:

• Molecular generative artificial intelligence (AI) shows promise in drug design.
• Generative models often produce unrealistic or unsynthesizable molecular structures.
• A need exists for methods to guide generative models toward drug-like chemical spaces.

Purpose of the Study:

• To define and evaluate applicability domains for generative AI models in drug design.
• To constrain generative models to produce molecules within specific chemical spaces.
• To enhance the generation of drug-like molecules using generative approaches.

Main Methods:

• Empirical examination of various applicability domain definitions for generative models.
• Generation of novel structures using generative methods constrained by applicability domains.
• Assessment of generated molecules based on structural similarity, physicochemical properties, unwanted substructures, and drug-likeness metrics.

Main Results:

• Applicability domain definitions significantly influence the drug-likeness of generated molecules.
• Identified specific applicability domain definitions that optimize the generation of drug-like molecules.
• Validated generative model outputs against quantitative structure-activity relationship (QSAR) predictions.

Conclusions:

• Defined effective applicability domains for generative models in drug discovery.
• Demonstrated that tailored applicability domains improve the quality and drug-likeness of generated molecules.
• This work facilitates the industrial application of generative AI for novel drug design.