Computational Docking and Deep Learning for Predicting Cross-Species Coronavirus Receptor Binding Dynamics

Introduction

Coronaviruses represent a significant and persistent threat to both animal and public health due to their capacity for cross-species transmission and rapid evolutionary adaptation [1, 2]. The initial step of coronavirus infection is mediated by the interaction between the viral spike (S) glycoprotein, specifically its receptor-binding domain (RBD), and a host cell surface receptor [1, 2]. For many coronaviruses, including severe acute respiratory syndrome coronavirus (SARS-CoV) and SARS-CoV-2, the primary receptor is angiotensin-converting enzyme 2 (ACE2) [2]. The molecular determinants of this interaction govern host range, tissue tropism, and zoonotic potential [1, 2]. Understanding and predicting these binding dynamics across diverse animal species is therefore a central challenge in veterinary virology and pandemic preparedness [1, 2].

Traditional experimental methods for characterizing RBD-receptor interactions, such as surface plasmon resonance and X-ray crystallography, are resource-intensive and low-throughput [1, 2]. In response, computational approaches have emerged as powerful tools for high-throughput screening and mechanistic prediction [1, 2]. This article provides an exhaustive technical review of two complementary computational paradigms: molecular docking and deep learning. It examines their application to predicting cross-species coronavirus receptor binding dynamics, with a focus on veterinary species including bats, swine, and other potential reservoir or intermediate hosts [1, 2]. The integration of these methods with genomic surveillance data is discussed as a framework for proactive zoonotic risk assessment [1, 2].

Molecular Docking for RBD-Receptor Interaction Prediction

Molecular docking is a computational technique that predicts the preferred orientation of one molecule (a ligand) when bound to a second molecule (a receptor) to form a stable complex [2]. In the context of coronavirus entry, the RBD of the spike protein is treated as the ligand, and the extracellular domain of the host receptor (e.g., ACE2) is treated as the receptor [2]. Docking algorithms sample a vast conformational space of possible binding poses and score each pose using an energy function that approximates the free energy of binding [2].

Docking Algorithms and Scoring Functions

Several docking programs are widely used in veterinary virology research. AutoDock Vina employs a hybrid scoring function that combines empirical and knowledge-based terms, including van der Waals interactions, electrostatic potentials, hydrogen bonding, and desolvation penalties [2]. HADDOCK (High Ambiguity Driven protein-protein DOCKing) uses biochemical and biophysical information, such as chemical shift perturbation data or mutagenesis data, to drive the docking process [2]. For protein-protein interactions like RBD-ACE2, rigid-body docking followed by flexible refinement is a common protocol [2].

The output of a docking simulation is typically a ranked list of predicted binding poses, each associated with a binding energy score (often in kcal/mol) [2]. A more negative binding energy indicates a predicted stronger interaction [2]. For cross-species comparisons, the binding energy of a given RBD variant against ACE2 orthologs from different animal species can be calculated and compared [2]. This allows for the computational ranking of species susceptibility to a particular coronavirus strain [2].

Application to Cross-Species Binding Affinity

A key application of molecular docking is the prediction of how mutations in the RBD alter binding affinity to ACE2 receptors from different species [2]. For example, the N501Y mutation, which arose in several SARS-CoV-2 variants of concern, was shown via docking simulations to enhance binding affinity to both human and murine ACE2 [2]. Similarly, docking studies have been used to evaluate the binding of bat coronavirus RBDs to human ACE2, providing an early indicator of zoonotic potential [1, 2].

The accuracy of docking predictions is highly dependent on the quality of the input three-dimensional (3D) structures [2]. When experimental structures are unavailable, homology modeling or deep learning-based structure prediction (discussed below) is required [1, 2]. Furthermore, docking scores are approximations and do not fully capture the entropic and solvation effects that govern binding in a physiological environment [2]. To improve reliability, docking is often followed by more computationally intensive molecular dynamics (MD) simulations [2].

Deep Learning for Structure Prediction and Binding Affinity Estimation

Deep learning, a subset of machine learning based on artificial neural networks with multiple layers, has revolutionized structural biology and binding affinity prediction [1, 2]. These models can learn complex, non-linear relationships from large datasets, enabling predictions that were previously intractable [1].

[Protein Structure](/knowledge/bioinformatics/protein-structure-biophysical-levels-folding 2) Prediction: AlphaFold2 and RoseTTAFold

The accurate prediction of protein 3D structure from amino acid sequence is a prerequisite for reliable docking [1]. AlphaFold2, a deep learning model developed by DeepMind, predicts protein structures with atomic-level accuracy rivaling experimental methods [1]. It uses a neural network architecture that integrates multiple sequence alignments (MSAs) and pairwise residue distance predictions to iteratively refine a structural model [1]. RoseTTAFold, a similar model from the Baker lab, employs a three-track neural network that simultaneously processes sequence, distance, and coordinate information [1].

For veterinary virology, these models are invaluable for predicting the structures of novel coronavirus spike proteins identified through sequence surveillance in animal populations [1]. For instance, the structure of a newly discovered bat coronavirus RBD can be predicted using AlphaFold2, and this predicted structure can then be used as the input for molecular docking against ACE2 orthologs from various species [1]. This pipeline enables rapid, pre-emptive assessment of zoonotic risk without requiring protein expression or crystallization [1].

Deep Learning for Binding Affinity Prediction

Beyond structure prediction, deep learning models can directly predict the binding affinity between a protein and a ligand or between two proteins [1, 2]. Graph neural networks (GNNs) are particularly well-suited for this task, as they can represent molecules as graphs where atoms are nodes and bonds are edges [1]. The GraphDTA model, for example, uses a GNN to predict drug-target binding affinity from graph representations of both the drug and the target protein [1].

In the context of coronavirus RBD-ACE2 interactions, deep learning models can be trained on large datasets of experimentally determined binding affinities and corresponding structural or sequence features [1, 2]. These models can then predict the binding affinity of novel RBD variants against different host receptors with high throughput [1, 2]. This approach is significantly faster than docking and MD simulations, making it suitable for screening large numbers of sequence variants identified through genomic surveillance [1, 2].

Integration with Sequence Surveillance

The predictive power of computational docking and deep learning is maximized when integrated with continuous genomic surveillance of coronaviruses in animal reservoirs [1, 2]. Databases such as GISAID and NCBI GenBank provide a rich repository of coronavirus sequences from diverse hosts, including bats, birds, swine, and other mammals [1, 2].

Identifying Emerging Variants

Sequence surveillance allows for the early detection of mutations in the RBD that may alter receptor binding specificity [1, 2]. By analyzing the frequency and geographic distribution of these mutations, researchers can identify variants with pandemic potential [1, 2]. For example, the emergence of the SARS-CoV-2 Delta and Omicron variants was tracked in real-time through global sequence sharing [2]. In a veterinary context, surveillance of porcine deltacoronavirus (PDCoV) and swine acute diarrhea syndrome coronavirus (SADS-CoV) in swine populations is critical for monitoring mutations that could enhance binding to human ACE2 [1, 2].

Computational Risk Assessment Pipeline

A comprehensive computational pipeline for cross-species risk assessment involves several steps. First, novel coronavirus sequences are retrieved from surveillance databases. Second, the RBD sequence is extracted and its 3D structure is predicted using deep learning models like AlphaFold2 [1]. Third, the predicted RBD structure is docked against a panel of ACE2 orthologs from key animal species (e.g., human, pig, bat, ferret, cat) using molecular docking software [2]. Fourth, deep learning models are used to predict binding affinities for the same set of interactions [1, 2]. Finally, the results from docking and deep learning are integrated to produce a quantitative risk score for each species.

The following Mermaid diagram illustrates this integrated workflow.

graph TD
    A[Sequence Surveillance: GISAID, NCBI], > B(Extract RBD Sequence)
    B, > C{Structure Prediction: AlphaFold2, RoseTTAFold}
    C, > D[Predicted RBD 3D Structure]
    D, > E[Molecular Docking: AutoDock Vina, HADDOCK]
    E, > F[Docking Scores: RBD vs. Host ACE2]
    B, > G[Deep Learning Binding Affinity: GraphDTA, GNNs]
    G, > H[Predicted Binding Affinities]
    F, > I[Integrated Risk Assessment]
    H, > I
    I, > J[Zoonotic Spillover Risk Score per Species]

Case Study: SARS-CoV-2 RBD Mutations and ACE2 Binding Across Species

The SARS-CoV-2 pandemic provides a well-documented case study for the application of these computational methods [2]. The ancestral SARS-CoV-2 RBD bound human ACE2 with high affinity, but subsequent mutations in variants of concern have modulated this interaction [2].

Key Mutations and Their Effects

The N501Y mutation, present in the Alpha, Beta, Gamma, and Omicron variants, was predicted by docking studies to increase binding affinity to human ACE2 by forming a new pi-pi stacking interaction with residue Y41 of the receptor [2]. Deep learning models trained on deep mutational scanning data have confirmed this prediction and further quantified the effect of this mutation on binding to ACE2 orthologs from other species, including mice and ferrets [2]. The K417N mutation, found in the Beta and Omicron variants, was shown to reduce binding to human ACE2 but was compensated for by other mutations like E484K and N501Y [2].

Implications for Veterinary Species

Computational docking studies have been used to predict the binding affinity of various SARS-CoV-2 variants to ACE2 from companion animals and livestock [2]. For example, the Omicron variant was predicted to have enhanced binding to feline and canine ACE2 compared to the ancestral strain, a finding that correlated with increased reports of natural infection in these species [2]. Similarly, docking simulations predicted that swine ACE2 binds the SARS-CoV-2 RBD with moderate affinity, suggesting that pigs could serve as potential intermediate hosts [2]. These computational predictions guide targeted surveillance efforts in animal populations [2].

Limitations and Future Directions

Despite their power, computational docking and deep learning have limitations. Docking scores are approximations that may not accurately reflect true binding free energies, particularly for highly flexible protein-protein interfaces [2]. Deep learning models are dependent on the quality and diversity of their training data and may perform poorly on novel mutations or host receptors not represented in the training set [1, 2]. Furthermore, these methods predict binding affinity but do not account for other factors critical for viral entry, such as receptor expression levels, tissue tropism, and the presence of host proteases required for spike protein cleavage [1, 2].

Future directions include the development of more accurate and transferable deep learning models that incorporate protein dynamics and solvation effects [1, 2]. The integration of AlphaFold3, which can predict protein-ligand and protein-protein complexes directly, promises to streamline the prediction pipeline [1]. Additionally, the application of these methods to a broader range of animal coronaviruses, including those from avian, camelid, and mustelid hosts, will be essential for comprehensive zoonotic risk assessment [1, 2].

Conclusion

Computational docking and deep learning represent a powerful, integrated framework for predicting cross-species coronavirus receptor binding dynamics [1, 2]. Molecular docking provides a biophysically grounded method for estimating binding energies between RBD variants and host receptors [2]. Deep learning models, including AlphaFold2 for structure prediction and GNNs for binding affinity estimation, dramatically increase the throughput and scope of these analyses [1, 2]. When coupled with continuous sequence surveillance in animal reservoirs, these computational tools enable proactive identification of emerging variants with zoonotic potential [1, 2]. This integrated approach is essential for veterinary virology, guiding surveillance, risk assessment, and the development of intervention strategies to prevent future pandemics [1, 2].

References

[1] Li X, Zhao X, Yu X, et al. Construction of a multi-tissue compound-target interaction network of Qingfei Paidu decoction in COVID-19 treatment based on deep learning and transcriptomic analysis. J. Bioinform. Comput. Biol. 2024. URL: https://www.semanticscholar.org/paper/828722db50fab83a093adb8f2e8f7dfa789f1434

[2] Pirolli D, Righino B, Camponeschi C, et al. Virtual screening and molecular dynamics simulations provide insight into repurposing drugs against SARS-CoV-2 variants Spike protein/ACE2 interface. Scientific Reports. 2023. URL: https://www.semanticscholar.org/paper/51d6c21eeab66a9ec469c779b411cc92232f9a0f *** Disclaimer: This article is for educational and informational purposes only. It is not intended to substitute for professional veterinary advice, diagnosis, treatment, or regulatory guidance. Always consult a licensed veterinarian or qualified specialist regarding animal health, disease diagnosis, and therapeutic decisions.