A deep dilated convolutional residual network for predicting interchain contacts of protein homodimers

Abstract

Deep learning has revolutionized protein tertiary structure prediction recently. The cutting-edge deep learning methods such as AlphaFold can predict high-accuracy tertiary structures for most individual protein chains. However, the accuracy of predicting quaternary structures of protein complexes consisting of multiple chains is still relatively low due to lack of advanced deep learning methods in the field. Because interchain residue-residue contacts can be used as distance restraints to guide quaternary structure modeling, here we develop a deep dilated convolutional residual network method (DRCon) to predict interchain residue-residue contacts in homodimers from residue-residue co-evolutionary signals derived from multiple sequence alignments of monomers, intrachain residue-residue contacts of monomers extracted from true/predicted tertiary structures or predicted by deep learning, and other sequence and structural features. Tested on three homodimer test datasets (Homo_std dataset, DeepHomo dataset, and CASP14-CAPRI dataset), the precision of DRCon for top L/5 interchain contact predictions (L: length of monomer in a homodimer) is 43.46%, 47.15%, and 24.81% respectively, which is substantially better than two existing deep learning interchain contact prediction methods. Moreover, our experiments demonstrate that using predicted tertiary structure or intrachain contacts of monomers in the unbound state as input, DRCon still performs reasonably well, even though its accuracy is lower than when true tertiary structures in the bound state are used as input. Finally, our case study shows that good interchain contact predictions can be used to build high-accuracy quaternary structure models of homodimers.

A deep reinforcement learning approach to reconstructing quaternary structures of protein dimers through self-learning

“…In the optimal scenario true interchain contacts are extracted from the native quaternary structures of protein dimers and used as input to guide the assembly of the true tertiary structures of monomers in the bound state into quaternary structures. An interchain contact is a pair of residues from the two chains in a dimer in which the shortest distance between their heavy atoms is less than or equal to 6Å (Quadir, Roy, Halfmann, et al 2021;Roy et al 2021) . In the suboptimal scenarios, the predicted interchain contacts together with the true tertiary structure of monomers in the bound state are used as input.…”

“…Despite the notable progress in tertiary structure prediction, the prediction of quaternary structure of protein complexes is still in the early stage of development (Zeng et al 2018; J. Hou et al 2020; Quadir, Roy, Soltanikazemi, et al 2021; Quadir, Roy, Halfmann, et al 2021, 2; Yan and Huang 2021; Roy et al 2021; Xie and Xu 2021). The methods for quaternary structure prediction can be subdivided into two categories: ab-initio methods(Lyskov and Gray 2008; Pierce et al 2014; Quadir, Roy, Soltanikazemi, et al 2021; Evans et al 2021; Park et al 2021) and template-based methods(Tuncbag et al 2012; Guerler, Govindarajoo, and Zhang 2013).…”

“…Recently, a gradient descent structure optimization method - GD (Soltanikazemi et al 2022) and a simulated annealing method - ConComplex (Quadir, Roy, Soltanikazemi, et al 2021) have been developed to build quaternary structures from predicted interchain residue-residue contacts and tertiary structures of protein chains. Due to the recent development of more accurate deep learning methods for interchain contact prediction such as DRCon(Roy et al 2021) and GLINTER(Xie and Xu 2021), these approaches can generate high-quality quaternary structures for some proteins. The latest significant advance in the field is the extension of the end-to-end deep learning method of predicting tertiary structures to quaternary structure prediction(Baek et al 2021; Jumper et al 2021).…”

Multi-Head Attention-based U-Nets for Predicting Protein Domain Boundaries Using 1D Sequence Features and 2D Distance Maps

“…We use the GPU computing resource on the Summit supercomputer provided by Oak Ridge National Laboratory (ORNL) to train the deep learning network above. The Summit cluster [40, 41] provides many compute nodes each having 6 GPUs and 16 GB of memory, which enables the distributed deep learning training. We train 2D U-Net, 1D U-Net, and the multi-head attention layer on three separate GPU nodes.…”

Prediction of inter-chain distance maps of protein complexes with 2D attention-based deep neural networks

“…Following the deep learning revolution in the prediction of intra-chain residue-residue distances and tertiary structures, recently some deep learning methods were developed to predict the inter-chain residue-residue contact map of homodimers and/or heterodimers, such as DeepHomo (Yan and Huang, 2021), DRcon (Roy, et al, 2021), and GLINTER (Xie and Xu, 2022) that predicts the contact map for both homodimers and heterodimers using as input a graph representation of protein monomer structure and the row attention maps generated from multiple sequence alignments (MSAs) by the MSA transformer (Rao, et al, 2021). The attention map calculated by the MSA transformer is a kind of residue-residue co-evolutionary feature extracted from MSAs.…”