Spike Protein Mutational Landscapes and ACE2 Binding Affinity Prediction Using Machine Learning

Abstract

The continuous evolution of viral spike glycoproteins, particularly within the receptor-binding domain (RBD), necessitates robust computational frameworks for predicting how amino acid substitutions alter binding affinity to the host angiotensin-converting enzyme 2 (ACE2) receptor. This review examines the integration of molecular dynamics simulations, Rosetta Flex ddG, and deep learning architectures such as AlphaFold2 and ESM-1v for scoring mutational effects on spike-ACE2 interfaces. The discussion emphasizes the biophysical principles underlying these methods, their application to emerging variants, and their relevance to zoonotic risk assessment and veterinary vaccine design. A case study on the Omicron lineage illustrates how computational pipelines can resolve the interplay between immune evasion and receptor engagement.

1. Introduction

Viral spike proteins mediate host cell entry by binding to specific receptors, with ACE2 serving as the primary entry portal for several coronaviruses [1, 2]. The RBD of the spike glycoprotein undergoes continuous mutational pressure from host immune responses, leading to the emergence of variants with altered receptor binding characteristics [3, 4]. Understanding the mutational landscape of the spike protein is critical for predicting host range, transmissibility, and vaccine efficacy in both human and animal populations [5, 6].

Computational approaches have become indispensable for systematically evaluating the effects of individual mutations and their combinations on binding affinity [7, 8]. These methods range from physics-based free energy perturbation to statistical potential scoring and deep learning models trained on large-scale mutational scanning data [9, 10]. The ability to predict binding affinity changes from sequence or structure alone enables rapid assessment of emerging variants before experimental characterization [11, 12].

This review provides a technical overview of the principal computational strategies for spike protein mutational landscape analysis and ACE2 binding affinity prediction. The focus is on the underlying biophysical mechanisms, algorithmic implementations, and their integration into predictive workflows.

2. Biophysical Basis of Spike-ACE2 Recognition

The spike-ACE2 interface is characterized by a network of hydrogen bonds, salt bridges, and hydrophobic contacts distributed across the RBD surface [13, 14]. The binding interface involves approximately 15 to 20 residues on each partner, with key contact residues including K417, N439, Y453, L455, F456, A475, F486, N487, Y489, Q493, G496, Q498, T500, N501, and Y505 on the spike RBD [15, 16]. These residues form a complementary electrostatic and shape-based interface with the N-terminal helix of ACE2 [17].

Mutations at these positions can alter binding affinity through several mechanisms: direct modification of intermolecular contacts, changes in side chain conformation, alterations in backbone dynamics, and long-range allosteric effects transmitted through the protein structure [18, 19]. The D614G substitution, for example, reshapes allosteric networks and opening mechanisms of the spike trimer without directly contacting ACE2 [6]. Similarly, mutations in the Omicron lineage exhibit epistatic interactions where the effect of one mutation depends on the presence of others [3].

Glycan shielding further modulates accessibility of the RBD to both ACE2 and antibodies [4]. The N-linked glycans at positions N165 and N234 stabilize the RBD in an open conformation, while glycans at N343 and N709 can occlude antibody epitopes [20]. Mutations that alter glycosylation motifs or glycan processing can therefore indirectly affect receptor binding by changing the conformational equilibrium of the spike [21].

3. Computational Methods for Mutational Effect Prediction

3.1 Molecular Dynamics Simulations

Molecular dynamics (MD) simulations provide atomistic resolution of spike-ACE2 interactions by solving Newtonian equations of motion for all atoms in the system [22, 23]. All-atom MD simulations typically employ explicit solvent models and physiological ionic strength to capture the electrostatic environment of the interface [24]. Simulation timescales ranging from hundreds of nanoseconds to several microseconds are required to observe conformational transitions relevant to binding [25].

Binding free energy calculations from MD trajectories commonly use the Molecular Mechanics Poisson-Boltzmann Surface Area (MM-PBSA) or Molecular Mechanics Generalized Born Surface Area (MM-GBSA) methods [22]. These approaches compute the free energy difference between bound and unbound states by averaging over conformational ensembles. The MM-PBSA method has been applied to quantify the impact of RBD mutations on ACE2 binding, with computed values correlating well with experimental surface plasmon resonance measurements [26].

Coarse-grained MD models reduce computational cost by representing groups of atoms as single interaction sites, enabling simulations of larger systems over longer timescales [27]. The Martini force field, for example, has been parameterized for protein-protein interactions and applied to study spike conformational dynamics [28]. These models are particularly useful for exploring the effects of mutations on spike trimer opening and RBD accessibility [29].

3.2 Rosetta Flex ddG

Rosetta Flex ddG is a computational protocol for predicting changes in binding free energy upon mutation [30]. The method combines backbone flexibility with side chain repacking to model the structural response to amino acid substitutions. The protocol involves three stages: a low-resolution pre-relaxation, a high-resolution refinement with side chain optimization, and a final energy evaluation using the Rosetta all-atom energy function [31].

The Rosetta energy function includes terms for van der Waals interactions, hydrogen bonding, solvation, electrostatic interactions, and backbone torsional preferences [32]. Flex ddG has been benchmarked on large datasets of protein-protein interface mutations and achieves a root-mean-square error of approximately 1.0 to 1.5 kcal/mol for binding free energy changes [33]. Application to spike RBD mutations has successfully recapitulated the enhanced ACE2 binding observed for N501Y and the reduced binding for K417N [34].

One limitation of Rosetta Flex ddG is its sensitivity to the input structure conformation. Multiple starting structures from MD simulations or crystal structures of different conformational states can improve prediction accuracy [35]. The method also does not explicitly account for changes in protonation states or the presence of glycans, which may be important for certain mutations [4].

3.3 Deep Learning Models

Deep learning approaches have emerged as powerful tools for predicting mutational effects from sequence and structure [10, 32]. Protein language models such as ESM-1v and ESM-2 are trained on millions of natural protein sequences using masked language modeling objectives [32]. These models learn evolutionary constraints that reflect functional and structural requirements, enabling zero-shot prediction of mutation effects without task-specific training data.

ESM-1v computes a log-likelihood ratio for each mutation by comparing the probability of the mutant sequence to the wild-type sequence under the model [32]. The resulting scores correlate with experimental measurements of binding affinity and fitness for viral proteins. Application to SARS-CoV-2 spike mutations has demonstrated that ESM-1v scores predict escape from neutralizing antibodies and changes in ACE2 binding [10].

AlphaFold2, originally developed for protein structure prediction, has been adapted for mutational effect prediction through structure-based scoring [19]. The AlphaFold2 confidence metrics, including the predicted local distance difference test (pLDDT) and predicted aligned error (PAE), can identify regions of structural perturbation upon mutation [19]. Recent implementations use AlphaFold2 to generate structural ensembles for mutant proteins and compute binding scores using the predicted interface distances [28].

Graph neural networks that operate on protein structures represent another class of deep learning models for binding affinity prediction [23]. These models encode the three-dimensional arrangement of atoms and residues as graphs, with nodes representing atoms or residues and edges representing spatial proximity. Message passing between nodes captures the local chemical environment and long-range interactions that determine binding energetics [23].

4. Workflow for Mutational Landscape Analysis

The following Mermaid diagram illustrates a typical computational workflow for predicting spike protein mutational effects on ACE2 binding affinity.

flowchart TD
    A[Viral Genome Sequencing], > B[Spike Protein Sequence Alignment]
    B, > C[Identification of Mutations]
    C, > D[Structural Modeling]
    D, > E[MD Simulations]
    D, > F[Rosetta Flex ddG]
    D, > G[Deep Learning Scoring]
    E, > H[MM-PBSA Binding Free Energy]
    F, > H
    G, > H
    H, > I[Consensus Binding Affinity Prediction]
    I, > J[Zoonotic Risk Assessment]
    I, > K[Vaccine Design Prioritization]
    I, > L[Experimental Validation]

The workflow begins with viral genome sequencing and identification of spike protein mutations relative to a reference sequence [2, 12]. Structural modeling using homology modeling or AlphaFold2 generates three-dimensional coordinates for the mutant spike RBD [19]. Parallel computational pipelines then apply MD simulations, Rosetta Flex ddG, and deep learning models to score each mutation [10, 30, 32]. The resulting predictions are integrated into a consensus score that accounts for the strengths and limitations of each method [23]. This consensus score informs downstream applications including zoonotic risk assessment and vaccine design prioritization [7, 21].

5. Case Study: Omicron Lineage Mutational Landscape

The Omicron variant of SARS-CoV-2 and its subsequent sublineages, including BA.1, BA.2, BA.5, and more recent variants such as KP.3.1.1 and XEC, have accumulated a large number of RBD mutations relative to the ancestral virus [1, 34]. The BA.1 lineage alone contains 15 RBD mutations, many of which are located at the ACE2 binding interface [3]. Deep mutational scanning studies have systematically mapped the effects of these mutations on ACE2 binding and antibody escape [1, 24].

Computational analyses using Rosetta Flex ddG predicted that the Omicron RBD mutations K417N, E484A, and Q493R collectively reduce ACE2 binding affinity, while N501Y and Q498R enhance binding [34]. The net effect is a modest increase in overall binding affinity, consistent with experimental measurements [16]. MD simulations revealed that the Omicron mutations alter the conformational dynamics of the RBD, increasing the population of the open state and enhancing ACE2 accessibility [16, 28].

Deep learning models including ESM-1v and AlphaFold2-based scoring have been applied to predict the fitness of Omicron subvariants [10, 32]. These models successfully identified the emergence of immune escape mutations such as F456L and L455F in the KP.3 lineage [1, 9]. The combination of computational predictions with experimental deep mutational scanning data provides a comprehensive view of the mutational landscape and its implications for continued viral evolution [24].

The Omicron case study demonstrates the importance of integrating multiple computational methods to capture the complex epistatic interactions that govern spike protein function [3]. Single-mutant predictions often fail to recapitulate the effects of combined mutations due to non-additive energetic contributions [3, 5]. Machine learning models that incorporate epistasis, such as those based on pairwise interaction terms or graph neural networks, are essential for accurate prediction of variant fitness [23, 32].

6. Implications for Zoonotic Risk Assessment and Veterinary Vaccine Design

The computational frameworks described above have direct applications in veterinary virology for assessing the zoonotic potential of animal coronaviruses [21, 26]. Bat coronaviruses, for example, harbor spike proteins with varying affinities for ACE2 receptors of different mammalian species [26]. Predicting which animal species are at risk of infection requires structural modeling of cross-species ACE2 binding and identification of mutations that expand host range [21].

Machine learning models trained on large-scale mutational scanning data can be applied to predict the effects of mutations in animal coronavirus spike proteins on binding to human ACE2 [10, 32]. These predictions inform surveillance efforts by identifying viral strains with elevated pandemic potential [7]. The same computational pipelines can be used to design veterinary vaccines that target conserved epitopes less prone to mutational escape [5, 7].

Vaccine design for animal populations benefits from the ability to predict which spike mutations are likely to emerge under immune pressure [7, 13]. Structure-based deep learning approaches can identify epitopes that are both immunogenic and mutationally constrained, providing targets for broadly protective vaccines [5, 14]. The integration of computational predictions with experimental validation in animal models accelerates the development of effective veterinary vaccines [18].

7. Limitations and Future Directions

Current computational methods for predicting mutational effects on binding affinity have several limitations. Force field-based methods such as MM-PBSA and Rosetta Flex ddG are sensitive to the starting structure and may not capture large conformational changes induced by mutations [30, 33]. Deep learning models require large training datasets and may generalize poorly to mutations that are underrepresented in the training data [10, 32]. The incorporation of glycan effects and post-translational modifications remains challenging for most methods [4].

Future developments will likely focus on integrating multiple data sources, including deep mutational scanning, cryo-electron microscopy structures, and evolutionary sequence information, into unified predictive models [24, 28]. The use of protein language models pre-trained on diverse viral sequences may improve generalization to novel variants [32]. Advances in coarse-grained MD simulations will enable routine screening of large numbers of mutations on full-length spike trimers [27].

8. Conclusion

The prediction of spike protein mutational effects on ACE2 binding affinity requires a multi-faceted computational approach that combines physics-based simulations, statistical potentials, and deep learning models. Each method contributes unique information about the biophysical consequences of amino acid substitutions, and their integration provides robust predictions for emerging viral variants. The application of these methods to veterinary virology enables proactive assessment of zoonotic risk and rational design of vaccines for animal populations. Continued development of computational tools and their validation against experimental data will further enhance our ability to anticipate viral evolution and mitigate its impact on animal and public health.

References

[1] Shao C, Yang L, Xiao C et al. Deep mutational scanning reveals the antibody escape and infectivity landscape of SARS-CoV-2 Omicron JN.1 and XEC receptor-binding domains. Emerg Microbes Infect. 2026. https://pubmed.ncbi.nlm.nih.gov/42324717/

[2] Gupta S, Chaudhary A, Bhatnagar S. SARS-CoV-2 Evolution and Its Implications for RT-PCR Diagnostic Performance. J Med Virol. 2026. https://pubmed.ncbi.nlm.nih.gov/42312714/

[3] Moulana A, Dupic T, Desai MM. Epistasis and background dependence in the evolution of Omicron variants of the SARS-CoV-2 spike protein. Mol Biol Evol. 2026. https://pubmed.ncbi.nlm.nih.gov/42142336/

[4] Kumar A, Yadav AJ, Tripathi T et al. Glycan shielding and epitope reorganization drive sotrovimab resistance in SARS-CoV-2 Omicron variants. Arch Biochem Biophys. 2026. https://pubmed.ncbi.nlm.nih.gov/42128042/

[5] Alshahrani M, Gatlin W, Ludwick M et al. Mechanisms of Binding and Immune Escape Resistance for Broadly Neutralizing Antibodies Targeting Distinct Conserved SARS-CoV-2 Spike Epitopes: A Hierarchical Approach Integrating Mutational Profiling and Energy Landscape Analysis. Int J Mol Sci. 2026. https://pubmed.ncbi.nlm.nih.gov/42123600/

[6] Kearns FL, Bogetti AT, Calvó-Tusell C et al. D614G reshapes allosteric networks and opening mechanisms of SARS-CoV-2 spikes. Proc Natl Acad Sci U S A. 2026. https://pubmed.ncbi.nlm.nih.gov/42101997/

[7] Uzer F, Erendor F, Sanlioglu S. SARS-CoV-2 Variants and Immune Evasion: Mapping the Future of Vaccine Design. Rev Med Virol. 2026. https://pubmed.ncbi.nlm.nih.gov/42047168/

[8] Alshahrani M, Gatlin W, Ludwick M et al. Frustration Landscapes of Broadly Neutralizing SARS-CoV-2 Spike Antibodies Targeting Conserved Epitopes Reveal Energetic Logic of Escape-Proof and Escape-Prone Mechanisms. bioRxiv. 2026. https://pubmed.ncbi.nlm.nih.gov/41959354/

[9] Kim YJ, Jeong SJ, Lee HR. Functional characterization of spike RBD mutations in SARS-CoV-2 Omicron-derived subvariants KP.3.1.1, LP.8.1, and NB.1.8.1. J Microbiol. 2026. https://pubmed.ncbi.nlm.nih.gov/41937394/

[10] Alamri MA, Alawam AS. Machine learning-guided rational engineering of ACE2-derived peptides for broad-spectrum neutralization of SARS-CoV-2 variants. Saudi Pharm J. 2026. https://pubmed.ncbi.nlm.nih.gov/41915253/

[11] Nguyen HL, Thai NQ, Tran L et al. Exploring the diverse binding ability of SARS-CoV-2 variant RBDs to different antibody classes: a computational study. RSC Adv. 2026. https://pubmed.ncbi.nlm.nih.gov/41853206/

[12] Fiaz N, Basheer A, Zahoor I et al. Distinct mutational landscapes and phylogenomic insights of the SARS-CoV-2 delta variant in Pakistan and India: Regional evolution, novel mutations, and epidemiological implications. PLoS One. 2026. https://pubmed.ncbi.nlm.nih.gov/41758893/

[13] Heng F, Magaret CA, Rouphael NG et al. The neutralizing antibody titer correlate of COVID-19 risk in the COVID-19 variant immunologic landscape (COVAIL) trial was not modified by SARS-CoV-2 amino acid sequence distances. Vaccine. 2026. https://pubmed.ncbi.nlm.nih.gov/41698311/

[14] Alshahrani M, Parikh V, Foley B et al. Dissecting binding and immune evasion mechanisms for ultrapotent Class I and Class 4/1 neutralizing antibodies of SARS-CoV-2 spike protein using a multi-pronged computational approach: neutral frustration architecture of binding interfaces and immune escape hotspots drives adaptive evolution. Phys Chem Chem Phys. 2026. https://pubmed.ncbi.nlm.nih.gov/41623222/

[15] Xu W, Guo T, Su H. Evolutionary aspect of spike glycoprotein's conformational dynamics. Phys Chem Chem Phys. 2026. https://pubmed.ncbi.nlm.nih.gov/41589433/

[16] Dutta M, Voth GA. Mutation and ACE2-induced allosteric network rewiring in Delta and Omicron SARS-CoV-2 spike proteins. Biophys J. 2026. https://pubmed.ncbi.nlm.nih.gov/41536063/

[17] Nan L, Mao T, Chan CWF et al. Droplet-based single-cell pairing for high-throughput interaction mapping of antigen-receptor combinations. Sci Adv. 2025. https://pubmed.ncbi.nlm.nih.gov/41385644/

[18] Sun H, Jiang Y, Lan M et al. Engineering a multivalent antibody nanoparticle to overcome SARS-CoV-2 Omicron immune evasion. PLoS Pathog. 2025. https://pubmed.ncbi.nlm.nih.gov/41359663/

[19] Di Salvatore V, Maleki A, Mohajer B et al. Exploring SARS-CoV-2 spike protein mutations through genetic algorithm-driven structural modeling. Bioinform Adv. 2025. https://pubmed.ncbi.nlm.nih.gov/41268478/

[20] Mani N, Suresh R, Chakraborty S. Cleaved versus Uncleaved: How furin cleavage reshapes the conformational landscape of SARS-CoV-2 spike. Protein Sci. 2025. https://pubmed.ncbi.nlm.nih.gov/41236412/

[21] Izhari MA, Alghamdi F, Alodeani EA et al. Evolutionary Insight into Fatal Human Coronaviruses (hCoVs) with a Focus on Circulating SARS-CoV-2 Variants Under Monitoring (VUMs). Biomedicines. 2025. https://pubmed.ncbi.nlm.nih.gov/41153733/

[22] Ahuja P, Rao L, Dsouza N. Targeting SARS-CoV-2 variants with Bromelain: an in-silico study of Spike-ACE2 inhibition using docking, molecular dynamic simulation, and MM-PBSA. In Silico Pharmacol. 2025. https://pubmed.ncbi.nlm.nih.gov/41113175/

[23] Mehra R, Thakur S. The structure-based approaches to computing viral fitness. Adv Protein Chem Struct Biol. 2025. https://pubmed.ncbi.nlm.nih.gov/40973410/

[24] Call MJ, Call ME, Wu X. Insights from deep mutational scanning in the context of an emerging pathogen. Biochem Soc Trans. 2025. https://pubmed.ncbi.nlm.nih.gov/40905952/

[25] Alvarez N, Gonzalez-Jimenez I, Rasheed R et al. Genetic and Immunological Profiling of Recent SARS-CoV-2 Omicron Subvariants: Insights into Immune Evasion and Infectivity in Monoinfections and Coinfections. Viruses. 2025. https://pubmed.ncbi.nlm.nih.gov/40733536/

[26] Ferreira E Silva Y, Fokoue HH, Batista PR. Exploring the Intrinsic Structural Plasticity and Conformational Dynamics of Human Beta Coronavirus Spike Glycoproteins. J Chem Inf Model. 2025. https://pubmed.ncbi.nlm.nih.gov/40673918/

[27] Alshahrani M, Parikh V, Foley B et al. Deciphering the Mechanistic Continuum of Broadly Neutralizing Class 4 Antibodies Targeting Conserved Cryptic Epitopes of the SARS-CoV-2 Spike Protein : Operating at the Intersection of Binding, Allostery and Immune Escape. bioRxiv. 2025. https://pubmed.ncbi.nlm.nih.gov/40672205/

[28] Feng Z, Huang J, Baboo S et al. Structural and functional insights into the evolution of SARS-CoV-2 KP.3.1.1 spike protein. Cell Rep. 2025. https://pubmed.ncbi.nlm.nih.gov/40618371/

[29] Alshahrani M, Parikh V, Foley B et al. Conformational Landscaping and Dynamic Mutational Profiling of Binding Interactions and Immune Escape for Broadly Neutralizing Class I Antibodies with SARS-CoV-2 Spike Protein: Distributed Binding Hotspot Networks Underlie Mechanism of Viral Resistance Against Existing Variants. bioRxiv. 2025. https://pubmed.ncbi.nlm.nih.gov/40510568/

[30] Dreyer FA, Schneider C, Kovaltsuk A et al. Computational design of therapeutic antibodies with improved developability: efficient traversal of binder landscapes and rescue of escape mutations. MAbs. 2025. https://pubmed.ncbi.nlm.nih.gov/40458889/

[31] Sharma A, Maurya S, Kumar S et al. An integrated multiscale computational framework deciphers SARS-CoV-2 resistance to sotrovimab. Biophys J. 2026. https://pubmed.ncbi.nlm.nih.gov/40394898/

[32] Ito J, Strange A, Liu W et al. A protein language model for exploring viral fitness landscapes. Nat Commun. 2025. https://pubmed.ncbi.nlm.nih.gov/40360496/

[33] Álvarez-Herrera M, Ruiz-Rodriguez P, Navarro-Domínguez B et al. Genome data artifacts and functional studies of deletion repair in the BA.1 SARS-CoV-2 spike protein. Virus Evol. 2025. https://pubmed.ncbi.nlm.nih.gov/40308784/

[34] Chakraborty C, Bhattacharya M, Abdelhameed AS. Recent SARS-CoV-2 evolution trajectories indicate the emergence of Omicron's several subvariants and the current rise of KP.3.1.1 and XEC. Virology. 2025. https://pubmed.ncbi.nlm.nih.gov/40187091/

[35] Mani N, Suresh R, Chakraborty S. Cleaved vs. Uncleaved: How Furin Cleavage Reshapes the Conformational Landscape of SARS-CoV-2 Spike. bioRxiv. 2025. https://pubmed.ncbi.nlm.nih.gov/40161653/ *** 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.