Enhancing accuracy of virtual kinase profiling via application of graph neural network to 3D pharmacophore ensembles

Kinase profiling is an essential step in both hit identification and selectivity evaluation. Since in vitro testing of large chemical libraries is costly and time-consuming, a computational approach can be applied to narrow down the reasonable chemical space. In this work, we collected data from several sources and prepared a curated, comprehensive database for training machine learning (ML) models to predict selectivity towards 75 kinases. We demonstrated the usefulness of this database by preparing several ML models with various molecular representations and model architectures. Among these, a graph neural network-based model enhanced by utilizing 3D pharmacophore ensembles showed the best performance. Finally, the developed model was applied to a library of in-stock compounds to facilitate kinase-focused drug discovery. © The Author(s), under exclusive licence to Springer Nature Switzerland AG 2025.