ColabDock: inverting AlphaFold structure prediction model for protein-protein docking with experimental restraints

Abstract

Prediction of protein complex structures and interfaces potentially has wide applications and can benefit the study of biological mechanisms involving protein-protein interactions. However, the surface prediction accuracy of traditional docking methods and AlphaFold-Multimer is limited. Here we present ColabDock, a framework that makes use of ColabDesign, but reimplements it for the purpose of restrained complex conformation prediction. With a generation-prediction architecture and trained ranking model, ColabDock outperforms HADDOCK and ClusPro not only in complex structure predictions with simulated residue and surface restraints, but also in those assisted by NMR chemical shift perturbation as well as covalent labeling. It further assists antibody-antigen interface prediction with emulated interface scan restraints, which could be obtained by experiments such as Deep Mutation Scan. ColabDock provides a general approach to integrate sparse interface restraints of different experimental forms and sources into one optimization framework.

Fast and accurate modeling and design of antibody-antigen complex using tFold

“…We compare tFold-Ag with currently available end-to-end methods, including AlphaFold-Multimer [5], Uni-Fold MuSSe [23], and RoseTTAFold2 [28], as well as docking-based methods such as conventional docking methods like ZDock [9], Clus-Pro [8], HDock [10], and deep-neural-network-involved docking methods such as EquiDock [11], dyMEAN [12], and ColabDock [13]. tFold-Ag and other end-to-end methods (sequences-to-complex structure) can directly predict the complex structure using input sequences, whereas docking-based methods (structures-to-complex structure) necessitate the structures of individual components to predict the final structure of the complex.…”

“…Other works [7,61] have applied these restraints to sift through and select the most plausible structures from a pool of candidates. More recently, with the advent of AlphaFold, researchers [13] have been exploring how to combine these restraints with pre-trained models to enhance prediction accuracy.…”

“…On one hand, experimental techniques such as X-ray crystallography, cryo-electron microscopy, nuclear magnetic resonance (NMR), and mutagenesis analysis for investigating antibody-antigen complex structures are usually expensive, time-consuming, and exhibit a low success rate, hindering their potential as high-throughput approaches [7]. On the other hand, computational methods like molecular docking for modelling antibody-antigen structures often result in high false-positive rates and seldom yield unique solutions, necessitating known structures of individual components as inputs [8][9][10][11][12][13]. Moreover, docking-based methods struggle to identify epitopes when protein binding sites are ambiguous and substantial conformational changes occur during binding, a common issue in antibody-antigen interactions.…”

Highly conserved brain vascular receptor ALPL mediates transport of engineered viral vectors across the blood-brain barrier

“…We docked the full VP3 structure of our capsid onto the ALPL receptor structure with the constraints that the interaction pairs of residues identified using our AlphaFold-multimer predicted structure. The docking procedure was performed using local implementation ColabDock 56 .…”

A Perspective on the Prospective Use of AI in Protein Structure Prediction

“…On the other hand, the recent development of predictors relying on large language models (LLM) opens new perspectives in that field. , In particular, the lower quality of the ESMFold predictions might be compensated by the considerable decrease of the calculation time required by language models. Finally, ColabDock has been proposed to perform a form of docking restrained by experimental data . Several studies have addressed the ability of AF2 to predict protein–peptide complexes .…”