AI-Designed Molecules in Drug Discovery, Structural Novelty Evaluation, and Implications

Area of Science:

• Medicinal Chemistry
• Computational Chemistry
• Drug Discovery

Background:

• Achieving structural novelty is a key challenge in modern drug discovery.
• Artificial intelligence (AI) excels at analyzing molecular structure-activity relationships but its potential for exploring novel chemical spaces is not fully understood.
• Evaluating the novelty of AI-generated compounds is crucial for advancing drug design.

Purpose of the Study:

• To systematically assess the structural novelty of AI-designed active compounds across diverse drug discovery cases.
• To compare the novelty generated by ligand-based versus structure-based AI approaches.
• To identify factors influencing novelty in AI-driven drug discovery workflows.

Main Methods:

• Analysis of 71 published cases of AI-designed active compounds.
• Evaluation of structural novelty using similarity metrics (e.g., Tanimoto coefficient).
• Comparison of outcomes from ligand-based and structure-based AI modeling strategies.

Main Results:

• Ligand-based AI models frequently produced molecules with low novelty (58.1% of cases with Tcmax > 0.4).
• Structure-based AI approaches demonstrated superior performance in generating novel compounds (17.9% with Tcmax > 0.4).
• Screening workflows and target characteristics significantly impact the novelty of AI-generated molecules.

Conclusions:

• Systematic novelty assessment and manual validation are essential to prevent structural homogenization in AI drug discovery.
• Optimizing AI-driven drug discovery requires diverse training data, scaffold-hopping aware similarity metrics, and judicious use of similarity filters.
• Interdisciplinary collaboration is vital to balance the generation of novel chemical structures with desired biological activity.