Optimizing Drug Design by Merging Generative AI With Active Learning Frameworks

• DOI: 10.48550/arxiv.2305.06334
• Publication Date: 2023-01-01
• Authors: Isaac Filella-Mercè; Alexis Molina; Marek Orzechowski; Lucía Díaz; Yang Ming Zhu; Julia Vilalta Mor; Laura Malo; Ajay Yekkirala; Soumya Ray; Vı́ctor Guallar
• Keywords: Chemical space; ADME; Cheminformatics

Traditional drug discovery programs are being transformed by the advent of machine learning methods. Among these, Generative AI methods (GM) have gained attention due to their ability design new molecules and enhance specific properties existing ones. However, current GM limitations, such as low affinity towards target, unknown ADME/PK properties, or lack synthetic tractability. To improve applicability domain methods, we developed a workflow based on variational autoencoder coupled with active steps. The designed iteratively learns from molecular metrics, including likeliness, synthesizability, similarity, docking scores. In addition, also included hierarchical set criteria advanced modeling simulations during final selection step. We tested our two model systems, CDK2 KRAS. both cases, generated chemically viable high predicted toward targets. Particularly, proportion high-affinity inferred was significantly greater than that in training data. Notably, uncovered novel scaffolds dissimilar those known for each target. These results highlight potential explore chemical space targets, thereby opening up possibilities endeavors.