Refining Binding Mode Predictions with Graph Neural Networks and Detailed Contact Maps

The Empirical scoring functions are widely used for pose and affinity prediction by docking softwares. But the correct prediction of binding affinity is still a challenging task and crucial for the success of structure-based VS experiments (Guedes, et. al, 2018). All of the docking programs are able to generate ligand conformations similar to crystallographically determined protein/ligand complex structures for at least one of the targets. However, scoring functions are less successful at distinguishing the crystallographic conformation from the set of docked poses. In that sense, a significant limitation in molecular docking arises from the lack of confidence in scoring functions' ability to provide accurate binding energies (Ramírez and Caballero, 2018). In this project, we selected a set with 284 complexes protein-ligand from the core data set of PDBbind (Liu et al., 2015), and carried out docking calculations, using AutoDock Vina (Eberhardt et al., 2021). Then, we took the first ten poses for each one and performed contact analysis with BINANA software. Using  the contact information, the representation of protein-ligand complexes is done in such a way that protein atoms whose minimum distance to ligand atoms is greater than 4 Å are excluded. Each pose and its receptor  is mapped as a  contact map and then transformed into a bipartite  graph with descriptors. The graphs are  labeled as the positive or negative class according to a combination of RMSD and native contacts.  A Graph Attention Neural Network model  was used to learn ligand-protein contact information.