Achieving structural novelty in drug discovery remains a critical challenge. Artificial intelligence (AI) has demonstrated remarkable potential in deciphering the complex relationships between molecular structures and activities from vast amounts of chemical and biological information. However, its ability to explore novel chemical spaces is underexplored. This study evaluates the structural novelty of AI-designed active compounds across 71 cases published in recent years. Ligand-based models often yield molecules with relatively low novelty (Tcmax > 0.4 in 58.1% of cases), whereas structure-based approaches exhibit better performance (17.9% with Tcmax > 0.4). Screening workflows significantly influence the novelty, with underexplored targets benefiting from structure-based methods. However, fingerprint-based similarity metrics may fail to detect scaffold-level similarities. Systematic novelty assessment and manual verification are essential to avoid structural homogenization. This Review provides insights for optimizing AI-driven drug discovery and underscores the need for interdisciplinary collaboration to balance novelty and activity. Specifically, we recommend the use of diverse training data sets, scaffold-hopping aware similarity metrics, and careful consideration of similarity filters in AI-driven drug discovery workflows.
Keywords: AI-driven drug discovery (AIDD); Artificial intelligence (AI); Ligand-based drug design (LBDD); Structure-based drug design (SBDD); Tanimoto coefficient (Tc).