Total: 1
The key to successful drug design lies in the correct comprehension of protein-ligand interactions. Within the current knowledge paragm, these interactions can be described from both thermodynamic and kinetic perspectives. In recent years, many deep learning models have emerged for predicting the thermodynamic properties of protein-ligand interactions. However, there is currently no mature model for predicting kinetic properties, primarily due to lack of kinetic data. To tackle this problem, we have developed a graph neural network model called STELLAR-koff (Structure-based TransfEr Learning for Ligand Activity Regression) to predict protein-ligand dissociation rate constant. Unlike traditional protein-ligand property prediction models, which typically use a single complex conformation as input, STELLAR-koff employs transfer learning to transform multiple ligand conformations within the protein into a protein ligand interaction landscape, and uses this landscape as the primary input for the model. In addition, we expanded the PDBbind koff dataset from 680 to 1,197 entries and employed the augmented dataset for model training and testing. When tested through five-fold cross-validation, STELLAR-koff achieved Pearson correlation coefficient of 0.729 surpassing or being on pair with most of the published prediction methods. Tested on external set, STELLAR-koff demonstrated strong predictive performance on unseen protein, achieving a Pearson of 0.838 on the focal adhesion kinase in particular. Experimental validation on cyclin-dependent kinase also demonstrated the effectiveness of STELLAR-koff in real drug discovering scenarios. We believe this study provides an effective tool for predicting protein-ligand dissociation rate constant and offers new insight for the future development of this field.