Modeling interactions with biological targets is a necessary step to begin reasoning about the therapeutic potential of a novel molecule in the drug discovery process. Molecular docking aids drug discovery researchers by searching over potential binding ‘poses’ of a drug molecule, ranking these poses according a scoring function, and using this information to identify the binding mode of a protein-small molecule complex, allowing for the quantification of the strength of the molecules binding interaction with their intended protein targets. While these simulations are able to scale on large computing clusters, the results that are produced are not generally accurate. Often the ‘best’ pose as identified by the docking algorithm is not actually the ‘correct’ pose. The score of the ‘best’ pose is also not highly correlated with an experimental binding affinity when it is known. The issues that arise with scoring functions in many molecular docking pipelines make this step of in silico prediction of binding strength even more difficult. A productive step forward would suggest a scoring function that is able to address these issues. Deep learning methods such as Convolutional Neural Networks (CNNs) and Graph Convolutional Neural Networks (GCNNs) have shown promise in providing researchers frameworks from which it is possible to model binding affinity with reasonable predictivity on experimental datasets. As part of an ongoing collaboration effort between Lawrence Livermore National Laboratory and the American Heart Association towards developing a human-protein drug atlas, we are investigating the effectiveness of convolutional neural networks and spatial graph methods for improving the accuracy and throughput of molecular docking pipelines to help address scaling challenges in the development of this resource.
Cite this work
Researchers should cite this work as follows: