In this project, we work with the protein kinase family, since they are involved in many diseases which make them an interesting drug target. The aim is to generate novel ligands as potential drug candidates for kinases using generative deep learning. The model includes structural information on the kinase binding pocket, since it is well studied and well conserved across protein kinases. We tackle the problem of ligand generation in a fragment-based approach by including domain-knowledge to increase kinase affinity and synthesizability in general.

Starting from the KinFragLib methodology, which fragments kinase ligands and assigns each fragment to a subpocket, we augment the library with docked ligands from kinodata. We currently explore graph-based deep learning for the recombination of fragments in a step-by-step manner. With this approach, we incorporate subpocket information in the model to generate potential kinase-focused ligands.