MSA clustering enhances AF-Multimer's ability to predict conformational landscapes of protein-protein interactions

Abstract

Motivation: Understanding the conformational landscape of protein-ligand interactions is critical for elucidating the binding mechanisms that govern these interactions. Traditional methods like molecular dynamics (MD) simulations are computationally intensive, leading to a demand for more efficient approaches. This study explores how multiple sequence alignment (MSA) clustering enhance AF-Multimer's ability to predict conformational landscapes, particularly for proteins with multiple conformational states.

Results: We verified this approach by predicting the conformational landscapes of chemokine receptor 4 (CXCR4) and glucagon receptor (GCGR) in the presence of their agonists and antagonists. In our experiments, AF-Multimer predicted the structures of CXCR4 and GCGR predominantly in active state in the presence of agonists and in inactive state in the presence of antagonists. Moreover, we tested our approach with proteins known to switch between monomeric and dimeric states, such as lymphotactin, SH3, and thermonuclease. AFcluster-Multimer accurately predicted conformational states during oligomerization, which AFcluster with AlphaFold2 alone fails to achieve. In conclusion, MSA clustering enhances AF-Multimer's ability to predict protein conformational landscapes and mechanistic effects of ligand binding, offering a robust tool for understanding protein-ligand interactions.

Availability and implementation: Code for running AFcluster-Multimer is available at https://github.com/KhondamirRustamov/AF-Multimer-cluster.

Figure 1.

Comparison of different sampling methods in predicting the conformational landscapes of protein-protein interactions. (A–C) AF2 and AF2-Multimer predictions of CXCR4 in apo (A), agonist (B) and antagonist-bound (C) states using full-MSA yield only inactive receptor conformations. (D–F) The ColabFold implementation of random MSA subsampling can enhance the diversity of conformational states predicted by the AlphaFold2 model; however, it is highly sensitive to the max-msa parameter; (G, H) the percentage of predictions being predicted in active or inactive states using max-msa 64:128 (G) and 32:64 (H). (I) Workflow for predicting conformational landscape of protein–protein interactions using AlphaFold and unpaired sequence clustering. Initially, the MSA of the metamorphic protein is sampled or clustered using various techniques and predicted by AlphaFold2 models. Then, an unpaired MSA, comprising the full ligand MSA and the sampled/clustered MSA of the metamorphic protein, is constructed and provided as input to the AlphaFold model.

Figure 2.

AF-Multimer combined with MSA clustering predicts conformation selectivity for CXCR4 agonist (CXCL12) and antagonist (vMIP-II). (A) TM5–TM6 RMSD of AF-Multimer predictions for all clusters to active or inactive conformations, colored by ipTM in agonists (upper left) and antagonists-bound states (upper right); distribution of top 10% predictions (ranked by ipTM) for agonist (lower left) and antagonist (lower right) bound CXCR4. (B) TM5–TM6 RMSD of AF-Multimer predictions for all clusters to active or inactive conformations, colored by plDDTs in agonists (upper left) and antagonists-bound CXCR4 (upper right); distribution of top 10% (ranked by plDDT) predictions for agonist (lower left) and antagonist (lower right) bound CXCR4. (C) Results of 300 ns MD simulations showing the conformational space exploration by agonist (left) and antagonist (right) bound CXCR4 in a POPC membrane (starting position illustrated by a red star, final position by green). (D, E) Overall structure of top 5 (ranked by plDDT) predictions for agonist-bound (D) and antagonist-bound (E) CXCR4 with a zoom on TM5–TM6 conformational changes during receptor activation [reference active structure (8U4P) shown in E, reference inactive structure (3OE6) shown in D].

Figure 3.

AFcluster predicts the conformational selectivity for RAMP2 and Gs binding to GCGR. (A, B) plDDTs and distribution of top 10% predictions (ranked by plDDT) for apo (left), RAMP2-bound (center), and Gs-bound (right) GCGR. (C, D) Top 5 predictions for RAMP2-bound (C) and Gs-bound (D) GCGR, colored by chain (left) and plDDT (right) with zoom-in of TM5-TM6 region compared to reference active structure and inactive structures. (E) Overall plDDTs of GCGR in apo, RAMP2-bound, and Gs-bound states.

Figure 4.

AFcluster predicts fold switching for human lympholactin N terminus in the dimeric state. (A) Reference monomer and dimeric experimental structures for human lympholactin. (B) Conformational landscape for lympholactin in monomeric (left) and dimeric (right) forms, colored by plDDT. (C) ipTMs for AF-Multimer predictions of lympholactin dimer MSA clusters. (D) Distribution of the top 10% (ranked by plDDT) of AF-Multimer predictions for monomer (left) and dimer (right) lympholactin structures. (E) Distribution of the top 10% (ranked by ipTM) of AF-Multimer predictions for the dimer. F) Subunits orientation in dimeric lympholactin in the experimental structure, compared to predicted results close to monomeric (center) and dimeric states (right).

Figure 5.

AF-Multimer predicts the oligomeric open states for SH3 domain and thermonuclease. (A) Mice SH3 domain obtains close conformation in monomer state and open state during the dimerization; (B) RMSD for AF2 predictions of all clusters to open and closed conformations (colored by plDDT); (C, D) conformational landscape of AF-Multimer predictions of all clusters in dimer state colored by ipTM (C) and plDDT (D); (E) thermonuclease obtains close conformation in monomer state and open state during the dimerization; (F) RMSD for AF2 predictions of all clusters to open and closed conformations (colored by plDDT); (G, H) conformational landscape of AF-Multimer predictions of all clusters in dimer state colored by ipTM (C) and plDDT (D).