Atomic Convolutional Networks for Predicting Protein-Ligand Binding Affinity

Empirical scoring functions based on either molecular force fields or descriptors are widely used, in conjunction with molecular, during the early stages of drug discovery to predict potency and affinity of a drug-like molecule to a given target. These models expert-level knowledge of physical chemistry and biology to be encoded hand-tuned parameters or features rather than allowing the underlying model select features in a data-driven procedure. Here, we develop a general3-dimensional spatial convolution operation for learning atomic-level chemical directly from atomic coordinates and demonstrate its application structure-based bioactivity prediction. The atomic convolutional neural is trained to predict the experimentally determined binding affinity of protein-ligand complex by direct calculation of the energy associated with complex, protein, and ligand given the crystal structure of the binding. Non-covalent interactions present in the complex that are absent in the-ligand sub-structures are identified and the model learns the strength associated with these features. We test our model by the binding free energy of a subset of protein-ligand complexes in the PDBBind dataset and compare with state-of-the-art cheminformatics machine learning-based approaches. We find that all methods achieve accuracy and that atomic convolutional networks either outperform perform competitively with the cheminformatics based methods. Unlike all protein-ligand prediction systems, atomic convolutional networks are-to-end and fully-differentiable. They represent a new data-driven,-based deep learning model paradigm that offers a strong foundation for improvements in structure-based bioactivity prediction.