Protein-protein docking using learned three-dimensional representations

Protein-protein interactions are determined by a number of hard-to-capture features related to shape complementarity, electrostatics, and hydrophobicity. These features may be intrinsic to the protein or induced by the presence of a partner. A conventional approach to protein-protein docking consists in engineering a small number of spatial features for each protein, and in minimizing the sum of their correlations with respect to the spatial arrangement of the two proteins. To generalize this approach, we introduce a deep neural network architecture that transforms the raw atomic densities of each protein into complex three-dimensional representations. Each point in the volume containing the protein is described by 48 learned features, which are correlated and combined with the features of a second protein to produce a score dependent on the relative position and orientation of the two proteins. The architecture is based on multiple layers of SE(3)-equivariant convolutional neural networks, which provide built-in rotational and translational invariance of the score with respect to the structure of the complex. The model is trained end-to-end on a set of decoy conformations generated from 851 nonredundant protein-protein complexes and is tested on data from the Protein-Protein Docking Benchmark Version 4.0.