Tackling the issue of confined chemical space with AI-based de novo drug design and molecular optimization

The search for molecular novelty frequently collides with the fact that drug candidates with the best translational prospects are confined to - or concentrated in - defined regions of chemical space. The new possibilities of AI, particularly retrosynthesis prediction and generative AI, allow for the automated or semi-automated exploration of less restricted and unexplored areas of chemical space.

The notion of novelty in drug discovery is discussed, and representative examples of AI-guided de novo drug design, optimization, and retrosynthesis prediction are presented, with a focus on reports on open-source tools published in the last 3 years (2022-2025). Scopus was used to search relevant literature.

Modern deep learning architectures have been adapted for the de novo design and molecular optimization. These technologies, and especially those based on conditional generation, will possibly have a great impact on expanding the regions of chemical space that are exploited therapeutically. However, there are some persistent challenges in the field that are gradually being addressed, including how to assess the synthetic accessibility of designed molecules without compromising the generation of structural novelty; the need to increase the availability and diversity of benchmark datasets; and the relative scarcity of large-scale experimental validation of the designs.