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

• DOI: 10.1093/bioinformatics/btac063
• Publication Date: 2022-01-31T20:13:34Z
• Authors: Raj S Roy; Farhan Quadir; Elham Soltanikazemi; Jianlin Cheng
• Keywords: Protein tertiary structure; Residue (chemistry); Protein quaternary structure; CASP

Deep learning has revolutionized protein tertiary structure prediction recently. The cutting-edge deep methods such as AlphaFold can predict high-accuracy structures for most individual chains. However, the accuracy of predicting quaternary complexes consisting multiple chains is still relatively low due to lack advanced in field. Because interchain residue-residue contacts be used distance restraints guide modeling, here we develop a dilated convolutional residual network method (DRCon) homodimers from co-evolutionary signals derived sequence alignments monomers, intrachain monomers extracted true/predicted or predicted by learning, and other structural features.Tested on three homodimer test datasets (Homo_std dataset, DeepHomo dataset CASP-CAPRI dataset), precision DRCon top L/5 contact predictions (L: length monomer homodimer) 43.46%, 47.10% 33.50% respectively at 6 Å threshold, which substantially better than DNCON2_inter similar Glinter. Moreover, our experiments demonstrate that using unbound state input, performs well, even though its lower true bound are input. Finally, case study shows good build models homodimers.The source code available https://github.com/jianlin-cheng/DRCon. https://zenodo.org/record/5998532#.YgF70vXMKsB.Supplementary data Bioinformatics online.