Designing selective and potent kinase inhibitors remains a significant challenge in drug discovery. This work introduces a powerful generative AI workflow that leverages 3D-equivariant deep learning to accelerate the de novo structure-based design of novel kinase inhibitor candidates. Our approach utilizes a fine-tuned DiffLinker model, which operates directly within the protein's binding pocket to generate diverse chemical linkers between known, active ligand fragments. This strategy effectively preserves the essential pharmacophoric anchors of established compounds while systematically exploring new chemical space for the linking moiety. Every generated molecule is subjected to a rigorous, multi-stage validation pipeline to further assess its therapeutic potential. We first assess 3D structural integrity using PoseBuster, eliminating candidates with steric clashes or invalid geometries. Survivors are then filtered based on a suite of 2D cheminformatic properties, including Lipinski's Rule of Five, synthetic accessibility (SA), and the quantitative estimate of druglikeness (QED). This funneling process ensures the final library is not only structurally sound and diverse but also composed of chemically viable molecules with favorable drug-like properties. A critical validation of our workflow is the preservation of key protein-ligand interactions. As a proof of concept, novel molecules designed based on CDK1 inhibitors successfully re-established crucial hydrogen bonds with the kinase hinge region, a hallmark of potent inhibition. This result underscores the model's capacity to learn and apply complex 3D structural and chemical constraints, confirming the biological relevance of its designs. Our study highlights the powerful synergy between generative AI and structure-based drug design, presenting an open-source framework that promises to streamline the discovery of next-generation kinase inhibitors.
Kelich et al. (Sun,) studied this question.