DRCon

DRCon predicts interchain residue-residue contacts in protein homodimers to support quaternary structure modeling.

Key Features:

• Deep Learning Architecture: DRCon uses a deep dilated convolutional residual network architecture.
• Input Data Utilization: The method leverages residue-residue co-evolutionary signals from multiple sequence alignments of monomers, intrachain residue-residue contacts (from true or predicted tertiary structures or predicted by deep learning), and additional sequence and structural features.
• Performance Metrics: On the Homo_std, DeepHomo, and CASP14-CAPRI test sets, DRCon attains precision for top L/5 interchain contacts (L = monomer length) of 43.46%, 47.15%, and 24.81%, respectively.
• Robustness to Input Variability: DRCon maintains reasonable performance when using predicted tertiary structures or intrachain contacts from the unbound state, with higher accuracy using true bound-state tertiary structures.

Scientific Applications:

• Quaternary Structure Modeling: Predicted interchain contacts can be used to model quaternary structures of protein complexes.
• Protein Engineering and Drug Design: Interchain contact predictions can inform design of proteins and development of drugs that target specific protein–protein interactions.

Methodology:

DRCon is trained on datasets of known homodimer structures by integrating co-evolutionary signals, intrachain contacts, and sequence/structural features, using dilated convolutions to capture long-range residue dependencies and residual connections to mitigate vanishing gradients.