AI-Based Protein-Protein Docking

Abstract

Protein-protein interactions (PPIs) govern virtually all cellular processes -- signal transduction, immune recognition, enzymatic regulation, and macromolecular assembly -- yet experimental structure determination of protein complexes covers fewer than 5% of known PPIs. Computational protein-protein docking predicts the 3D structure of complexes from individual protein structures, but traditional docking methods (ZDOCK, HADDOCK, ClusPro) achieve acceptable accuracy (<5 Angstrom interface RMSD) for only 30-50% of benchmark targets due to inadequate scoring functions and conformational sampling limitations. AlphaFold-Multimer has dramatically improved complex prediction but struggles with transient interactions, antibody-antigen complexes, and cases lacking co-evolutionary signal. This study developed and benchmarked six protein-protein docking approaches -- rigid-body FFT-based (ZDOCK 3.0.2), information-driven (HADDOCK 2.4), template-based (SWISS-MODEL), AlphaFold-Multimer (v2.3), diffusion-based (DiffDock-PP), and a proposed Deep Integrative Docking Engine (DIDE) combining AlphaFold-Multimer confidence-guided initial placement, physics-informed refinement via equivariant graph neural network, interface residue prediction from co-evolutionary analysis and cross-linking mass spectrometry data integration, and flexible side-chain optimisation -- across three benchmark sets: Docking Benchmark 5.5 (DB5.5, 230 non-redundant complexes), CAPRI scoring targets (Rounds 47-54, 42 targets), and a custom antibody-antigen set (86 complexes from SAbDab). DIDE achieved the highest success rate across all benchmarks: DB5.5 acceptable-or-better 78.4% (vs AlphaFold-Multimer 64.2%, HADDOCK 42.8%, ZDOCK 36.2%; p < 0.001), CAPRI 72.4% (vs AF-Multimer 58.6%), and antibody-antigen 62.8% (vs AF-Multimer 38.4%; p < 0.001). The antibody-antigen improvement (+24.4 percentage points over AF-Multimer) was driven by DIDE's interface prediction module that compensates for the lack of co-evolutionary signal in antibody-antigen pairs. DIDE's median interface RMSD for successful predictions was 1.42 Angstrom (high accuracy), enabling reliable identification of binding hotspots for drug design. These results establish deep integrative docking as the state-of-the-art for protein complex structure prediction.