InterPepRank: Assessment of Docked Peptide Conformations by a Deep Graph Network.

Peptide-protein interactions between a smaller or disordered peptide stretch and a folded receptor make up a large part of all protein-protein interactions. A common approach for modeling such interactions is to exhaustively sample the conformational space by fast-Fourier-transform docking, and then refine a top percentage of decoys. Commonly, methods capable of ranking the decoys for selection fast enough for larger scale studies rely on first-principle energy terms such as electrostatics, Van der Waals forces, or on pre-calculated statistical potentials. We present InterPepRank for peptide-protein complex scoring and ranking. InterPepRank is a machine learning-based method which encodes the structure of the complex as a graph; with physical pairwise interactions as edges and evolutionary and sequence features as nodes. The graph network is trained to predict the LRMSD of decoys by using edge-conditioned graph convolutions on a large set of peptide-protein complex decoys. InterPepRank is tested on a massive independent test set with no targets sharing CATH annotation nor 30% sequence identity with any target in training or validation data. On this set, InterPepRank has a median AUC of 0.86 for finding coarse peptide-protein complexes with LRMSD < 4Å. This is an improvement compared to other state-of-the-art ranking methods that have a median AUC between 0.65 and 0.79. When included as a selection-method for selecting decoys for refinement in a previously established peptide docking pipeline, InterPepRank improves the number of medium and high quality models produced by 80% and 40%, respectively. The InterPepRank program as well as all scripts for reproducing and retraining it are available from: http://wallnerlab.org/InterPepRank.

SpatialSim: Recognizing Spatial Configurations of Objects With Graph Neural Networks.

An embodied, autonomous agent able to set its own goals has to possess geometrical reasoning abilities for judging whether its goals have been achieved, namely it should be able to identify and discriminate classes of configurations of objects, irrespective of its point of view on the scene. However, this problem has received little attention so far in the deep learning literature. In this paper we make two key contributions. First, we propose SpatialSim (Spatial Similarity), a novel geometrical reasoning diagnostic dataset, and argue that progress on this benchmark would allow for diagnosing more principled approaches to this problem. This benchmark is composed of two tasks: "Identification" and "Discrimination," each one instantiated in increasing levels of difficulty. Secondly, we validate that relational inductive biases-exhibited by fully-connected message-passing Graph Neural Networks (MPGNNs)-are instrumental to solve those tasks, and show their advantages over less relational baselines such as Deep Sets and unstructured models such as Multi-Layer Perceptrons. We additionally showcase the failure of high-capacity CNNs on the hard Discrimination task. Finally, we highlight the current limits of GNNs in both tasks.