Predicting receptor-ligand pairing preferences in plant-microbe interfaces via molecular dynamics and machine learning

Abstract

Microbiome assembly, structure, and dynamics significantly influence plant health. Secreted microbial signaling molecules initiate and mediate symbiosis by binding to structurally compatible plant receptors. For example, lipo-chitooligosaccharides (LCOs), produced by nitrogen-fixing rhizobial bacteria and various fungi, are recognized by plant lysin motif receptor-like kinases (LysM-RLKs), which activate the common symbiotic pathway. Accurately predicting these molecular interactions could reveal complementary signatures underlying the initial stages of endosymbiosis. Despite the breakthrough in protein-ligand structure prediction with deep learning-based tools, such as AlphaFold3, the large size and highly flexible nature of signaling compounds like LCOs present major challenges for detailed structural characterization and binding-affinity prediction. Typical structure-/physics-based methods of ligand virtual screening are designed for small, drug-like molecules, often rely on high-resolution, experimentally determined structures of the protein receptors, and rarely achieve sufficient sampling to obtain converged thermodynamic quantities with large ligands. In this study, we developed a hybrid molecular dynamics/machine learning (MD/ML) approach capable of predicting binding affinity rankings with high accuracy in systems involving large, flexible ligands, despite limited experimental structural information. Using coarse initial structural models, the predictions using the MD/ML workflow achieved strong alignment with experimental trends, particularly in the top-affinity tier for four legume LysM-RLKs (LYR3) binding to LCOs and a chitooligosaccharide. Furthermore, the MD-based conformation selection protocol provided critical structural insights into substrate specificity and binding mechanisms. This study demonstrates a powerful method to screen for challenging cognate ligand-receptors and advance our understanding of the molecular basis of microbial colonization in plants.

Keywords: Lipo-chitooligosaccharides; LysM; Machine learning; Plant-microbe interactions; Protein-ligand binding affinity prediction; Signaling molecules.

This manuscript has been authored by UT-Battelle, LLC, under contract DE-AC05–00OR22725 with the US Department of Energy (DOE). The US government retains and the publisher, by accepting the article for publication, acknowledges that the US government retains a nonexclusive, paid-up, irrevocable, worldwide license to publish or reproduce the published form of this manuscript, or allow others to do so, for US government purposes. DOE will provide public access to these results of federally sponsored research in accordance with the DOE Public Access Plan (http://energy.gov/downloads/doe-public-access-plan).

Fig. 1

Sampling and conformation selection protocol. Overall molecular dynamics-based protocol used in MD/ML for conformation selection of ligand-bound protein complexes, starting with minimum structural information about the interaction.

Fig. 2

Spearman rank (SR) correlations between experimental and predicted binding affinities for LysM:LCO/CO complex structures obtained with MD/ML for four benchmark systems. SR values obtained using in-house trained ML models and previously published scoring functions (constituent scores; gray shade) using either multiple structures selected from the MD-based conformation selection protocol (orange bars) or using a single structure with the lowest molecular-mechanics noncovalent energy (blue bars) are displayed. Ligands and proteins from Malkov et al# (2016) were selected as the experimental reference data. Table 2 shows the source data used to generate this plot, i.e., the predicted and corresponding experimental values of binding affinity. In-house trained models: SVM=support-vector machines; RF=random forest; GBT=gradient-boosted trees; XGB=XGBoost. Constituent scores: XS=X-Score; DSX; DV=DeltaVina; RF-3 =RF-Score-3; Cy=Cyscore. Consensus SR values obtained from the in-house ML model and MM-derived rankings are also displayed.

Fig. 3

Non-covalent interaction energies between benchmark proteins and different conformers of LCO/CO ligands. The molecular mechanics (MM) interaction energies of MD/ML selected protein-ligand conformers were calculated using the CHARMM36m force field in GROMACS. The x-axis corresponds to the experimentally determined binding strength rank (1 being the strongest binder).

Fig. 4

Models of LysM proteins bound to LCO-V(C18:1_Δ11, S). The selected conformers with the lowest molecular mechanics non-covalent energy of interaction between protein and ligand are shown for LjLYR3, PvLYR3, GmLYR3–11, and PsLYR3. Amino acids that are not conserved among the four proteins and amino acids that are discussed in the text are depicted as licorice and labeled. Only the LysM ectodomain of these proteins is shown, with the LysM3 subdomain highlighted in green (cartoon representation). LCO-V(C18:1_Δ11, S) is represented with light orange carbons in licorice (oxygen, nitrogen, and sulfur atoms are colored red, blue, and yellow, respectively).

Fig. 5

Comparative analysis of the strongest to the weakest binder LCOs/CO with LjLYR3. The selected conformers with the lowest MM non-covalent energy of interaction with the protein were used as reference models in this analysis. Amino acids that are not conserved among the four benchmark proteins and amino acids that are discussed in the text are depicted as licorice and labeled. Only the LysM3 subdomain is shown (green cartoon representation). The LCOs/CO ligands are represented with light orange carbons in licorice (oxygen, nitrogen, and sulfur atoms are colored red, blue, and yellow, respectively). Published experimentally determined dissociation constants of each complex is shown in parentheses.