Deep Learning for Predicting Receptor-Binding Domain Dynamics in Emerging Zoonotic Coronaviruses

Introduction

The emergence of zoonotic coronaviruses from wildlife reservoirs, particularly bats and pangolins, represents a persistent threat to animal and public health. The receptor-binding domain (RBD) of the coronavirus spike glycoprotein is the primary determinant of host tropism and cross-species transmission potential. Understanding the structural dynamics and binding affinities of RBDs from diverse animal coronaviruses is essential for predicting spillover risk. Deep learning models, including AlphaFold2, RoseTTAFold, and graph neural networks (GNNs), have revolutionized the prediction of protein structures and their conformational ensembles. This article provides a technical examination of how these computational methods are applied to forecast RBD dynamics, binding affinities, and mutational escape in emerging zoonotic coronaviruses of veterinary significance.

Biological and Biophysical Context of RBD-Receptor Interactions

Coronavirus entry into host cells is mediated by the spike glycoprotein, a class I fusion protein that exists as a trimer on the virion surface. The RBD, located within the S1 subunit, undergoes conformational transitions between a closed (receptor-inaccessible) state and an open (receptor-accessible) state. This structural rearrangement is essential for binding to the host receptor, most commonly angiotensin-converting enzyme 2 (ACE2) in mammals. The RBD-ACE2 interface involves a network of hydrogen bonds, hydrophobic contacts, and salt bridges that collectively determine binding affinity and species specificity.

In bat sarbecoviruses and pangolin-related coronaviruses, the RBD sequence exhibits substantial variability, particularly in the receptor-binding motif (RBM). This variability directly influences the ability of the virus to utilize ACE2 orthologs from different mammalian species. The binding affinity between an RBD and a given ACE2 ortholog is a critical parameter for estimating zoonotic potential. Deep learning models are increasingly used to predict these affinities from sequence and structural data.

Deep Learning Architectures for Protein Structure Prediction

AlphaFold2 and RoseTTAFold

AlphaFold2 employs an end-to-end deep learning architecture that integrates multiple sequence alignments (MSAs) with pairwise residue distance predictions. The model uses an Evoformer module to process MSA and pair representations, followed by a structure module that generates atomic coordinates via iterative refinement. For RBD prediction, AlphaFold2 produces high-confidence models of the RBD in both open and closed conformations, depending on the input template and MSA depth.

RoseTTAFold uses a three-track architecture that simultaneously processes sequence, distance, and coordinate information. This architecture enables the model to predict protein structures with accuracy comparable to AlphaFold2, particularly for multi-domain proteins such as the spike trimer. RoseTTAFold is especially useful for modeling RBD dynamics because it can generate multiple conformational states from a single sequence input.

Graph Neural Networks for Binding Affinity Prediction

Graph neural networks represent protein structures as graphs where nodes correspond to residues and edges represent spatial or sequential relationships. For RBD-ACE2 binding prediction, GNNs encode the three-dimensional coordinates of the interface and learn to predict binding free energies (ΔΔG) from structural features. These models incorporate physicochemical properties such as electrostatic potential, hydrophobicity, and van der Waals contacts.

GNN-based predictors are trained on large datasets of experimentally determined binding affinities and structural complexes. They can generalize to novel RBD sequences by learning the geometric and chemical rules that govern protein-protein interactions. This capability is critical for evaluating the zoonotic potential of newly discovered coronaviruses from wildlife surveillance.

Molecular Dynamics Simulations and Deep Learning Integration

Molecular dynamics (MD) simulations provide atomic-level resolution of RBD conformational dynamics. Classical MD simulations using all-atom force fields (e.g., CHARMM, AMBER) can capture the opening and closing motions of the RBD on microsecond timescales. However, the computational cost of long-timescale simulations limits their application to large panels of RBD variants.

Deep learning methods address this limitation through several approaches. First, deep generative models, including variational autoencoders and normalizing flows, learn the conformational landscape of the RBD from short MD trajectories and then generate physically plausible conformations without explicit simulation. Second, machine learning force fields (e.g., DeepMD, SchNet) approximate the potential energy surface with neural networks, enabling MD simulations at reduced computational cost while maintaining quantum mechanical accuracy.

Markov state models (MSMs) constructed from MD trajectories can be combined with deep learning to predict the kinetics of RBD conformational transitions. MSMs partition the conformational space into discrete states and estimate transition probabilities between them. Deep learning can enhance MSM construction by learning low-dimensional embeddings of the conformational space that preserve the essential dynamics.

Mutational Scanning and Deep Learning

Deep mutational scanning (DMS) experiments systematically measure the effects of single amino acid substitutions on RBD function, typically using yeast or mammalian display systems coupled with high-throughput sequencing. The resulting fitness landscapes provide rich training data for deep learning models.

Convolutional neural networks (CNNs) and transformer-based models can predict the effects of mutations on RBD-ACE2 binding affinity from sequence alone. These models learn the epistatic interactions between residues, which are critical for accurate prediction because the effect of a mutation often depends on the background sequence context.

For zoonotic risk assessment, deep learning models trained on DMS data from known coronaviruses can be applied to predict the binding consequences of mutations observed in wildlife-derived sequences. This approach enables the identification of RBD variants with enhanced affinity for livestock or companion animal ACE2 orthologs.

Workflow for Predicting RBD Dynamics and Zoonotic Risk

The following Mermaid diagram illustrates a typical computational workflow integrating deep learning, MD simulations, and mutational scanning for predicting RBD dynamics and zoonotic spillover risk.

flowchart TD
    A[Sequence Surveillance: GISAID, NCBI], > B[MSA Construction and Phylogenetic Analysis]
    B, > C[Deep Learning Structure Prediction: AlphaFold2, RoseTTAFold]
    C, > D[RBD Conformational Ensemble Generation]
    D, > E[Graph Neural Network Binding Affinity Prediction]
    D, > F[Molecular Dynamics Simulations: All-Atom and Coarse-Grained]
    F, > G[Markov State Model Construction]
    G, > H[Kinetic Characterization of RBD Opening]
    E, > I[Binding Affinity Matrix: RBD vs. Host ACE2 Orthologs]
    H, > I
    I, > J[Deep Mutational Scanning Prediction]
    J, > K[Mutation Fitness Landscape]
    K, > L[Identification of High-Risk Variants]
    L, > M[Zoonotic Spillover Risk Assessment]
    M, > N[Experimental Validation: Pseudovirus Entry Assays]

Data Sources and Surveillance Databases

The Global Initiative on Sharing All Influenza Data (GISAID) and the National Center for Biotechnology Information (NCBI) GenBank database are primary repositories for coronavirus genomic sequences. GISAID provides curated sequence data with metadata on host species, geographic origin, and collection date. For zoonotic coronavirus surveillance, sequences from bat, pangolin, civet, and other wildlife species are particularly valuable.

Structural data for RBD-ACE2 complexes are deposited in the Protein Data Bank (PDB). These experimental structures serve as templates for homology modeling and as validation benchmarks for deep learning predictions. Cryo-electron microscopy (cryo-EM) structures of spike trimers in different conformational states provide additional constraints for modeling RBD dynamics.

Cross-Links to Related Articles

This article is part of a broader series on computational virology and zoonotic risk prediction. Readers are directed to the following related articles for additional depth on specific topics:

• Machine Learning-Driven Prediction of Receptor-Binding Dynamics in Emerging Zoonotic Coronaviruses
• Predicting Zoonotic Spillover: Computational Modeling of Receptor-Binding Dynamics in Emerging Bat Coronaviruses
• Computational Insights into Host Receptor Binding Dynamics of Emerging Zoonotic Coronaviruses Using Molecular Dynamics Simulations
• Computational Docking and Binding Affinity Prediction for Emerging Zoonotic Coronaviruses
• Deep Learning-Driven Prediction of Viral Receptor-Binding Domain Mutations
• Molecular Modeling of Zoonotic Spillover: Predicting Receptor-Binding Dynamics in Bat-Borne Coronaviruses
• Computational Prediction of Host Tropism and Receptor Binding Dynamics in Emerging Zoonotic Coronaviruses
• Understanding the Structural Dynamics of Zoonotic Spillover
• Computational Prediction of Cross-Species Receptor Binding Dynamics in Emerging Zoonotic Coronaviruses
• Computational Docking and Deep Learning for Predicting Cross-Species Coronavirus Receptor Binding Dynamics
• Deep Learning-Driven Prediction of Viral Receptor-Binding Domain Evolution and Escape Mutations
• Deep Mutational Scanning and Machine Learning Predictions of SARS-CoV-2 Spike Protein Receptor Binding Domain Escape Mutants
• Computational Prediction of Receptor-Binding Domain Mutations in Emerging SARS-CoV-2 Variants
• Deep Mutational Scanning and Computational Protein Design for Predicting Zoonotic Spillover Risk
• Deep Learning for Predicting Viral Host-Range Transitions and Zoonotic Potential
• Structural Prediction and Binding Dynamics of Zoonotic Spillover
• Structural and Evolutionary Dynamics of Zoonotic Viral Glycoproteins
• Computational Prediction of Spike Protein Mutations and ACE2 Binding Dynamics in Emerging Coronaviruses
• Deep Learning-Driven Prediction of Envelope Protein Dynamics in Zoonotic Bat Coronavirus Entry
• Predicting Spike Protein Evolution in Emerging Coronaviruses Using Structural Modeling and Machine Learning
• Spike Protein Dynamics and Host Receptor Binding: Computational Simulations
• Integrating Cryo-EM and Molecular Dynamics Simulations to Elucidate Glycan Shield Dynamics
• Molecular Dynamics Simulations of Bat Coronavirus Spike Protein-Receptor Interactions
• Computational Prediction of Viral Entry Dynamics
• In Silico Profiling of Viral Receptor-Binding Domain Evolutionary Trajectories
• Structure-Guided Antiviral Design
• Biological Foundation Models for Predicting Host Tropism in Emerging Zoonotic Viruses
• Computational Prediction of Cross-Species Receptor Binding: Bat Coronavirus Spike Protein Evolution
• Spike Protein Mutational Landscapes and ACE2 Binding Affinity Prediction Using Machine Learning
• AlphaFold and Beyond: Deep Learning for Protein Structure Prediction in Veterinary Virology
• In Silico Prediction of Viral Glycoprotein Dynamics
• Structural Prediction of Viral Envelope Glycoproteins Using AlphaFold2
• Spike Protein Dynamics and Host Range Prediction Using Molecular Dynamics and Machine Learning
• Biological Foundation Models for Predicting Host Tropism of Zoonotic Viruses
• Structural and Functional Annotation of Novel Bat Coronaviruses using AlphaFold2 and Molecular Docking
• Machine Learning for Variant Effect Prediction on Protein Stability
• Machine Learning-Guided Design of Pan-Coronavirus Spike Protein Inhibitors
• Protein-Protein Interface Design and Binding Energy Prediction
• Molecular Dynamics Simulations of Membrane-Bound Viral Glycoproteins
• Structural and Evolutionary Dynamics of Coronavirus Spike Protein
• AlphaFold and Beyond: Predicting Viral Protein Structures for Antiviral Target Discovery
• Zoonotic Spillover Pathways and Receptor Binding Evolution in Bat Reservoirs
• Molecular Dynamics Simulations of Viral Spike Glycoproteins
• Molecular Dynamics Simulations of Bat Coronavirus Spike Protein-Receptor Interactions: Implications for Zoonotic Risk Assessment

Challenges and Limitations

Deep learning models for protein structure prediction have known limitations when applied to viral glycoproteins. The high glycosylation density of the coronavirus spike protein can obscure the RBD surface and affect conformational sampling. Most deep learning models do not explicitly account for glycan shielding, which can lead to overestimation of binding affinities.

Another challenge is the limited availability of experimentally determined RBD-ACE2 complex structures for diverse animal coronaviruses. Deep learning models trained primarily on human ACE2 complexes may not generalize well to livestock or companion animal ACE2 orthologs. Transfer learning and few-shot learning approaches are being developed to address this data scarcity.

The conformational dynamics of the RBD are influenced by the quaternary structure of the spike trimer. Isolated RBD constructs may not fully recapitulate the steric constraints present in the full spike. Multiscale modeling approaches that integrate coarse-grained representations of the spike trimer with all-atom descriptions of the RBD are needed for accurate predictions.

Future Directions

The integration of deep learning with experimental structural biology techniques such as cryo-EM and hydrogen-deuterium exchange mass spectrometry will improve the accuracy of RBD dynamics predictions. End-to-end differentiable models that directly predict binding affinities from genomic sequences without intermediate structure prediction steps are an active area of research.

Foundation models trained on large corpora of protein sequences and structures, such as ESM-2 and ProtGPT2, offer the potential to predict RBD dynamics and binding properties directly from sequence. These models capture evolutionary and physicochemical constraints that are not explicitly encoded in structure-based approaches.

For veterinary applications, the development of species-specific ACE2 binding predictors is a priority. Deep learning models trained on panels of ACE2 orthologs from livestock species (e.g., swine, bovine, poultry) and companion animals (e.g., canine, feline) will enable rapid risk assessment when novel coronaviruses are detected in wildlife.

Conclusion

Deep learning has become an indispensable tool for predicting RBD dynamics and binding affinities in emerging zoonotic coronaviruses. The combination of structure prediction networks, graph neural networks, and molecular dynamics simulations provides a comprehensive computational framework for assessing spillover risk. Continued advances in model architecture, training data diversity, and integration with experimental validation will further enhance the predictive power of these methods. For veterinary virologists and computational biologists, these tools offer a pathway from sequence surveillance to actionable risk assessment for zoonotic coronavirus emergence.

References

1. Jumper J, Evans R, Pritzel A, et al. Highly accurate protein structure prediction with AlphaFold. Nature. 2021;596(7873):583-589.
2. Baek M, DiMaio F, Anishchenko I, et al. Accurate prediction of protein structures and interactions using a three-track neural network. Science. 2021;373(6557):871-876.
3. Senior AW, Evans R, Jumper J, et al. Improved protein structure prediction using potentials from deep learning. Nature. 2020;577(7792):706-710.
4. Torrisi M, Pollastri G, Le Q. Deep learning methods in protein structure prediction. Computational and Structural Biotechnology Journal. 2020;18:1301-1310.
5. AlQuraishi M. End-to-end differentiable learning of protein structure. Cell Systems. 2019;8(4):292-301.
6. Xu J, McPartlon M, Li J. Improved protein structure prediction by deep learning irrespective of co-evolution information. Nature Machine Intelligence. 2021;3(7):601-609.
7. Greener JG, Kandathil SM, Moffat L, Jones DT. A guide to machine learning for biologists. Nature Reviews Molecular Cell Biology. 2022;23(1):40-55.
8. Pakhrin SC, Shrestha B, Adhikari B, Kc DB. Deep learning-based advances in protein structure prediction. International Journal of Molecular Sciences. 2021;22(11):5553.
9. Wu Z, Johnston KE, Arnold FH, Yang KK. Protein sequence design with deep learning. Current Opinion in Chemical Biology. 2021;65:18-27.
10. Strokach A, Becerra D, Corbi-Verge C, Perez-Riba A, Kim PM. Fast and flexible protein design using deep graph neural networks. Cell Systems. 2020;11(4):402-411.
11. Jing B, Eismann S, Suriana P, Townshend RJL, Dror RO. Learning from protein structure with geometric vector perceptrons. International Conference on Learning Representations. 2021.
12. Ingraham J, Garg V, Barzilay R, Jaakkola T. Generative models for graph-based protein design. Advances in Neural Information Processing Systems. 2019;32.
13. Anand N, Eguchi R, Mathews II, et al. Protein sequence design with a learned potential. Nature Communications. 2022;13(1):746.
14. Hsu C, Verkuil R, Liu J, et al. Learning inverse folding from millions of predicted structures. bioRxiv. 2022.
15. Dauparas J, Anishchenko I, Bennett N, et al. Robust deep learning based protein sequence design using ProteinMPNN. Science. 2022;378(6615):49-56.
16. Minkyung B, Baker D. Accurate prediction of protein structures and interactions using a deep learning network. Science. 2021;373(6557):871-876.
17. Evans R, O'Neill M, Pritzel A, et al. Protein complex prediction with AlphaFold-Multimer. bioRxiv. 2021.
18. Bryant P, Pozzati G, Elofsson A. Improved prediction of protein-protein interactions using AlphaFold2. Nature Communications. 2022;13(1):1265.
19. Yin R, Feng BY, Varshney A, Pierce BG. Benchmarking AlphaFold for protein complex modeling. Biophysical Journal. 2022;121(3):74a.
20. Burke DF, Bryant P, Barrio-Hernandez I, et al. Towards a structurally resolved human protein interaction network. Nature Structural and Molecular Biology. 2023;30(2):216-225.
21. Gao M, Nakajima An D, Parks JM, Skolnick J. AF2Complex predicts direct physical interactions in multimeric proteins with deep learning. Nature Communications. 2022;13(1):1744.
22. Humphreys IR, Pei J, Baek M, et al. Computed structures of core eukaryotic protein complexes. Science. 2021;374(6573):eabm4805.
23. Zhang C, Freddolino PL, Zhang Y. COFACTOR: improved protein function prediction by combining structure, sequence and protein-protein interaction information. Nucleic Acids Research. 2017;45(W1):W291-W299.
24. Yang J, Roy A, Zhang Y. Protein-ligand binding site recognition using complementary binding-specific substructure comparison and sequence profile alignment. Bioinformatics. 2013;29(20):2588-2595.
25. Roy A, Kucukural A, Zhang Y. I-TASSER: a unified platform for automated protein structure and function prediction. Nature Protocols. 2010;5(4):725-738.
26. Zhang Y. I-TASSER server for protein 3D structure prediction. BMC Bioinformatics. 2008;9:40.
27. Zhou H, Zhou Y. Distance-based protein folding powered by deep learning. Proceedings of the National Academy of Sciences. 2020;117(48):30297-30303.
28. Zheng W, Li Y, Zhang C, et al. Deep-learning contact-map guided protein structure prediction in CASP13. Proteins: Structure, Function, and Bioinformatics. 2019;87(12):1149-1164.
29. Wang S, Sun S, Li Z, Zhang R, Xu J. Accurate de novo prediction of protein contact map by ultra-deep learning model. PLoS Computational Biology. 2017;13(1):e1005324.
30. Jones DT, Kandathil SM. High precision in protein contact prediction using fully convolutional neural networks and minimal sequence features. Bioinformatics. 2018;34(19):3308-3315.
31. Adhikari B, Hou J, Cheng J. DNCON2: improved protein contact prediction using two-level deep convolutional neural networks. Bioinformatics. 2018;34(9):1466-1472.
32. Michel M, Menendez Hurtado D, Elofsson A. PconsC4: fast, accurate and hassle-free contact predictions. Bioinformatics. 2019;35(15):2677-2679.
33. Hanson J, Paliwal K, Litfin T, Yang Y, Zhou Y. Improving prediction of protein secondary structure, backbone angles, solvent accessibility and contact numbers by using predicted contact maps. Bioinformatics. 2019;35(21):4223-4230.
34. Klausen MS, Jespersen MC, Nielsen H, et al. NetSurfP-2.0: improved prediction of protein structural features by integrated deep learning. Proteins: Structure, Function, and Bioinformatics. 2019;87(6):520-527.
35. Heffernan R, Paliwal K, Lyons J, et al. Improving prediction of secondary structure, local backbone angles, and solvent accessible surface area of proteins by iterative deep learning. Scientific Reports. 2015;5:11476.
36. Fang C, Shang Y, Xu D. MUFOLD-SS: new deep inception-inside-inception networks for protein secondary structure prediction. Proteins: Structure, Function, and Bioinformatics. 2018;86(5):592-598.
37. Busia A, Jaitly N. Next-step conditioned deep convolutional neural networks improve protein secondary structure prediction. arXiv. 2017.
38. Wang S, Peng J, Ma J, Xu J. Protein secondary structure prediction using deep convolutional neural fields. Scientific Reports. 2016;6:18962.
39. Zhou J, Troyanskaya OG. Deep supervised and convolutional generative stochastic network for protein secondary structure prediction. International Conference on Machine Learning. 2014.
40. Spencer M, Eickholt J, Cheng J. A deep learning network approach to ab initio protein secondary structure prediction. IEEE/ACM Transactions on Computational Biology and Bioinformatics. 2015;12(1):103-112.
41. Li Z, Yu Y. Protein secondary structure prediction using cascaded convolutional and recurrent neural networks. International Joint Conference on Artificial Intelligence. 2016.
42. Sonderby SK, Winther O. Protein secondary structure prediction with long short term memory networks. arXiv. 2014.
43. Chen D, Tian X, Zhou B, Gao J. Profold: protein secondary structure prediction using a hybrid deep learning method. IEEE Access. 2018;6:52768-52776.
44. Mirabello C, Pollastri G. Porter, PaleAle 4.0: high-accuracy prediction of protein secondary structure and relative solvent accessibility. Bioinformatics. 2013;29(16):2056-2058.
45. Magnan CN, Baldi P. SSpro/ACCpro 5: almost perfect prediction of protein secondary structure and relative solvent accessibility using profiles, machine learning and structural similarity. Bioinformatics. 2014;30(18):2592-2597.
46. Faraggi E, Zhang T, Yang Y, Kurgan L, Zhou Y. SPINE X: improving protein secondary structure prediction by multist