Abstract:Proteins are the 'workhorse' molecules carrying out nearly all biological functions within living organisms. Most of the functions of a protein are performed through its interactions with other proteins, known as protein−protein interactions (PPIs). Common instances include the interactions between antibodies and antigens, which bolster an organism's capacity to identify and combat external pathogens, the binding of ligand and receptor proteins that instigate cellular signalling cascades, and the interplay between enzyme and substrate proteins that facilitate metabolic processes. The important functional roles played by PPIs render them pivotal targets in numerous contemporary drug design initiatives.1 Significant efforts have been made to determine the three-dimensional structures of PPI complexes, which could provide a geometric and physical landscape to facilitate biological function annotation and drug discovery efforts targeting PPIs. Although structural biology techniques such as X-ray crystallography and Cryo-Electron Microscopy offer the most precise structural insights into PPIs, they frequently demand significant resources and lack scalability for resolving molecular structures across the entire proteome. The most recent success of artificial intelligence (AI) algorithms, such as AlphaFold2,2, 3 has enabled the computational prediction of protein structures with remarkable accuracy, opening avenues for obtaining high-quality structures of proteins and PPI complexes through computational means. Most AI methods train models from large-scale experimental structures built on co-evolutionary information obtained from multiple sequence alignments (MSAs), as the latter can offer critical structural information of the target protein. Specifically, during evolution when a mutation occurs at one residue site and disrupts its interaction with other residues, the protein may become unstable, making it difficult for species with such mutation to survive. However, if the interacting residues mutate at the same time and stabilize the protein structure, the species can continue to survive. This phenomenon refers to 'co-evolution'. Since proteins in current organisms have all undergone the rigours of co-evolution over hundreds of millions of years, aligning a vast array of protein sequences in MSAs can effectively deduce information about protein spatial distances between residues (see Figure 1A). Although MSA and the co-evolutionary information have been successfully utilized by various AI-based 3D structure prediction approaches,2, 4 due to the lack of large-scale PPI sequence databases, however, constructing comprehensive MSAs and deducing reliable quaternary co-evolutionary structure insights remain a major bottle-neck problem for PPI complex structure predictions. To tackle the challenge, we developed the DMFold5, 6 algorithm, in which we utilized a simple protocol to construct PPI MSAs and deduce inter-chain co-evolutionary information by pairing monomer-chain MSAs that can benefit from abundant sequence resources from metagenome repositories7 (see Figure 1B). As outlined in Figure 1C, DMFold first constructs monomer MSAs for each component chain through iterative dynamic programming and Hidden Markov Model searches against multiple metagenome sequence libraries.8 To ensure orthologous sequence alignments, which is critical for inter-chain co-evolutionary map deduction, DMFold selectively links the homologous sequences of different chains coming from the same species, where a new AI-driven MSA scoring strategy has been introduced to rank the MSAs prior to sequence pairing. The assembled PPI MSAs are subsequently fed into an end-to-end deep-learning neural network module to generate PPI complex structure predictions. To enhance the biomedical usefulness, the online DMFold server (https://zhanggroup.org/DMFold/) creates multiple outputs from primary sequences in addition to the PPI complex structure models, including the top 10 experimental structures closest to the target and structure-based functional annotations on gene ontology, enzyme commission and binding ligands, respectively. The Critical Assessment of protein Structure Prediction (CASP) is a community-wide blind experiment which holds biennially and aims at providing an objective benchmark of the state-of-the-art structure prediction technologies. DMFold participated (as 'Zheng') in the most recent CASP15 and ranked as the best protein complex structure prediction method by the official CASP assessors,9 where the overall Z-score of DMFold (35.4) was 18% higher than the second best method (29.9) and 2.9 times higher than that of the standard version of AlphaFold2 (12.3) (Figure 2A). Figure 2B lists the results of DMFold on six large-size protein complexes containing > 1500 residues, where TM-scores of the DMFold models are .98, .91, .94, .93, .93 and .85, respectively. Here, TM-score is a standard measure of structure prediction accuracy with larger values indicating higher prediction accuracy, where a TM-score > .9 generally corresponds to models with medium-to-high experimental structure accuracy.10 Notably, the two largest targets (H1111 and H1114) contain > 7900 residues both being heteromeric complexes with stoichiometry variable of 'A9B9C9' and 'A4B8C8', respectively. These results underscore the remarkable capacity of DMFold to model large-size PPI complexes, towards the solution of a longstanding challenge faced by traditional quaternary structure modelling approaches. A particularly exciting application of DMFold is on modelling antibody- or nanobody-antigen complexes, a type of PPI that plays pivotal roles in the defence of our body against the invasion of external pathogens. Figure 2C presents a comparison of structural models by DMFold and AlphaFold2 on three nanobody-antigen targets in CASP15, which represent three typical interaction modes of nanobodies with mouse CNPase. While AlphaFold2 failed to model the PPI orientations with TM-score below .7 for all three complexes, DMFold demonstrates exceptional prediction power with three models achieving TM-scores of .92, .95 and .99, respectively. The capability to accurately model functional PPIs such as antibody−antigen complex structures can help significantly expedite the development process of antibody-based therapeutics. In summary, we developed DMFold, an AI-based pipeline for atomic-level PPI structure prediction. The major advancement of DMFold lies in the construction of precise multi-chain MSAs by iteratively collecting and pairing high-quality monomeric MSAs from huge metagenome sequence databases, which allows for the derivation of reliable cross-chain co-evolutionary information and, therefore, enables AI network models to derive accurate quaternary structural patterns and 3D PPI conformations. We expect that the improved power of AI-based complex structure predictions will significantly enhance the accuracy of large-scale protein function annotations and the effectiveness of structure-based drug discovery processes targeting various PPI-related human diseases. Wei Zheng and Qiqige Wuyun wrote the initial version of the manuscript and prepared the figures. Yang Zhang revised and submitted the manuscript. This work is supported in part by the National Institute of General Medical Sciences (GM136422 and S10OD026825 to Y.Z.), the National Institute of Allergy and Infectious Diseases (AI134678 to Y.Z.), the National Science Foundation (IIS1901191, DBI2030790 and MTM2025426 to Y.Z.) and the National University of Singapore Start-up Grant (A-80001129-00-00 to Y.Z.). The authors declare no competing interests. Not applicable.

Predicting the structure of large protein complexes using AlphaFold and Monte Carlo tree search

Predicting the structure of large protein complexes using AlphaFold and Monte Carlo tree search

Predicting the structure of large protein complexes

CombFold: predicting structures of large protein assemblies using a combinatorial assembly algorithm and AlphaFold2

Accurate prediction of protein assembly structure by combining AlphaFold and symmetrical docking

Evaluation of AlphaFold-Multimer prediction on multi-chain protein complexes

Comprehensive Evaluation of AlphaFold-Multimer, AlphaFold3 and ColabFold, and Scoring Functions in Predicting Protein-Peptide Complex Structures

Deep Reinforcement Learning for Modelling Protein Complexes

Predicting direct physical interactions in multimeric proteins with deep learning

Enhancing AlphaFold-Multimer-based Protein Complex Structure Prediction with MULTICOM in CASP15

Impact of AlphaFold on structure prediction of protein complexes: The CASP15‐CAPRI experiment

AFM-RL: Large Protein Complex Docking Using AlphaFold-Multimer and Reinforcement Learning

AF2Complex predicts direct physical interactions in multimeric proteins with deep learning

Improved prediction of protein-protein interactions using AlphaFold2

Improved protein complex prediction with AlphaFold-multimer by denoising the MSA profile

Highly accurate protein structure prediction for the human proteome

Multi-domain and complex protein structure prediction using inter-domain interactions from deep learning

One Step Forward Towards Deep‐learning Protein Complex Structure Prediction by Precise Multiple Sequence Alignment Construction

Benchmarking AlphaFold for protein complex modeling reveals accuracy determinants

Highly accurate protein structure prediction with AlphaFold

Protein Complex Structure Prediction Powered by Multiple Sequence Alignments of Interologs from Multiple Taxonomic Ranks and AlphaFold2