Computational Design of Viral Glycoprotein Binders for Neutralization

Introduction

The development of neutralizing binders that target viral glycoproteins represents a central strategy in antiviral intervention. Glycoproteins such as influenza hemagglutinin (HA) and coronavirus spike (S) mediate host cell attachment and membrane fusion, making them primary targets for antibody-based neutralization [1, 2]. In veterinary virology, analogous targets exist for pathogens including avian influenza virus, porcine respiratory coronavirus, and equine arteritis virus. Traditional binder discovery relies on animal immunization or phage display, but computational design now enables de novo generation of high-affinity binders that can be optimized for stability, specificity, and manufacturability [3]. This review examines the biophysical principles and algorithmic frameworks for computational design of viral glycoprotein binders, with emphasis on structural alignment, contact interface energy optimization, and the integration of protein language models.

Glycoprotein Architecture and Neutralization Interfaces

Viral glycoproteins are trimeric or homodimeric assemblies that display receptor-binding domains (RBDs) and fusion machinery on their surface. For class I fusion proteins such as coronavirus S and influenza HA, the RBD undergoes conformational rearrangements upon receptor engagement, exposing conserved epitopes that are accessible to neutralizing antibodies [1, 2]. The spatial arrangement of glycan shields further modulates epitope accessibility, creating a steric challenge for binder design. Computational approaches must account for these dynamic features by incorporating multiple conformational states derived from cryo-electron microscopy and X-ray crystallography [3].

Structural alignment of target epitopes across viral strains is critical for designing cross-reactive binders. Sequence conservation analysis identifies residues that are both solvent-exposed and essential for receptor binding, such as the receptor-binding motif (RBM) in coronavirus spike [2]. In influenza HA, the stem region is more conserved than the globular head, providing a rationale for targeting stem epitopes with computationally designed binders [1]. These structurally defined interfaces are then used as templates for binder docking and affinity maturation.

Computational Frameworks for Binder Design

Rosetta-based Design

The Rosetta macromolecular modeling suite provides a comprehensive platform for designing protein-protein interactions. The core algorithm performs Monte Carlo sampling of backbone and side-chain conformations to optimize the binding energy of a designed binder against a fixed target surface [1, 3]. Key steps include: (i) docking of a scaffold protein onto the target epitope using the RosettaDock module, (ii) hotspot residue prediction using computational alanine scanning, (iii) iterative design of interface residues to maximize shape complementarity and electrostatic matching, and (iv) computational saturation mutagenesis to improve affinity [3].

Energy functions in Rosetta combine physics-based terms (Lennard-Jones potential, Coulombic electrostatics, implicit solvation) with knowledge-based terms (Ramachandran preferences, Dunbrack rotamer libraries). The interface energy (ΔΔG_binding) is calculated as the difference between the complex energy and the sum of isolated partner energies. A threshold of −15 Rosetta energy units (REU) or lower is often used to identify designs with high predicted affinity [1, 3]. For glycoprotein targets, additional carbohydrate modeling is required because glycans can contribute favorably or unfavorably to binding depending on their orientation.

Deep Learning Approaches

Recent advances integrate generative models into binder design. RFdiffusion, a diffusion-based neural network originally trained on protein structures, can generate novel backbone conformations that adopt the desired binding mode [1]. The process begins with a noise-corrupted representation of the target interface, which is iteratively denoised to produce a binder backbone that fits into the target surface. This is followed by sequence design using ProteinMPNN or similar inverse folding models. The combination of RFdiffusion and Rosetta score refinement has yielded high-affinity nanobodies against the SARS-CoV-2 spike protein in silico [1].

Large language models (LLMs) pretrained on antibody repertoires can generate complementarity-determining region (CDR) sequences with high structural plausibility [2]. For instance, the generative LLM reported by He et al. produced CDR-H3 loops targeting the SARS-CoV-2 RBD, with binding affinity confirmed via surface plasmon resonance [2]. These models learn the sequence-structure mapping from millions of natural and synthetic antibody sequences, enabling the design of CDRs that are absent from any known antibody library. The integration of such LLM-generated sequences with Rosetta-based interface optimization represents a hybrid approach that combines data-driven diversity with physics-based validation.

Protein Engineering as a Test Case

Zahradník and Schreiber [3] systematically explored the design of protein-protein interactions using the SARS-CoV-2 spike-ACE2 interaction as a test case. They demonstrated that computational saturation mutagenesis followed by deep mutational scanning can isolate high-affinity ACE2 variants that act as competitive inhibitors of viral entry. Their workflow employed Rosetta for initial binding energy calculations and then validated designs using yeast surface display and flow cytometry [3]. This paradigm directly translates to veterinary targets: for example, designing a soluble form of the feline coronavirus receptor (fAPN) that binds to the viral spike and blocks infection.

Workflow for Binder Design

The following diagram summarizes the computational pipeline for designing glycoprotein binders.

graph TD
    A[Target Glycoprotein Structure], > B[Epitope Selection & Conformational Ensemble]
    B, > C[Scaffold Selection or De Novo Generation]
    C, > D[RFdiffusion / RosettaDock Docking]
    D, > E[Interface Residue Optimization]
    E, > F[Energy Minimization & ΔΔG Evaluation]
    F, > G[Sequence Design via ProteinMPNN or LLM]
    G, > H[Rosetta Score Refinement]
    H, > I[Experimental Validation]
    I, > J[Affinity Maturation Iteration]
    J, > C

Table 1 lists key computational tools and their roles.

Structural Analysis of Binder-Glycoprotein Complexes

Visualizing the designed binder-glycoprotein complex in a three-dimensional viewer is essential for evaluating contact interfaces. Key parameters to assess include:

• Buried surface area (BSA): A total BSA of 800–1200 Ų for the interface is typical for high-affinity binders. Lower values indicate insufficient contact.
• Shape complementarity (Sc): As calculated by the Sc statistic in Rosetta, values above 0.6 suggest good geometric fit.
• Polar contacts: The number of hydrogen bonds and salt bridges per 100 Ų of BSA should exceed 2.0 for stable binding.
• Van der Waals clashes: Clashes with energy >1.0 REU should be eliminated by side-chain repacking.

For glycoprotein targets, the presence of N-linked glycans near the epitope requires explicit modeling. The Rosetta carbohydrate library allows incorporation of glycan conformations, and the interface energy can be adjusted to penalize designs that clash with glycans. In the Virtual Lab study [1], automated AI agents managed these complexities by iteratively refining designs based on structural feedback.

Examples from Recent Literature

The Virtual Lab of AI agents designed new SARS-CoV-2 nanobodies by integrating RFdiffusion for backbone generation and Rosetta for sequence design [1]. The system autonomously selected epitope targets on the spike RBD, generated binder candidates, and filtered them using docking energy and conformational stability. This fully automated pipeline produced nanobodies with nanomolar affinity, validated by biolayer interferometry [1].

He et al. [2] employed a pretrained generative LLM to design CDR-H3 sequences targeting the SARS-CoV-2 spike RBD. By conditioning the model on structural constraints from the target epitope, they generated CDR sequences that achieved equilibrium dissociation constants (Kd) in the low nanomolar range. The study confirmed that computational CDR design can produce antibodies with neutralization activity comparable to those derived from immunization [2].

Zahradník and Schreiber [3] used iterative computational saturation mutagenesis to evolve the human ACE2 receptor into a high-affinity spike binder. Starting from the wild-type ACE2 structure, they simulated all single mutations at the interface and experimentally screened the top candidates. The final variant exhibited picomolar affinity and effectively blocked pseudovirus entry into Vero cells [3]. This methodology is directly applicable to veterinary receptor targets such as avian ANPEP (aminopeptidase N) for infectious bronchitis virus or porcine ACE2 for coronaviruses.

Veterinary Relevance and Translational Considerations

Computational binder design offers a rapid response platform for emerging veterinary pathogens. For avian influenza H5N1 or H9N2, the conserved stem epitope of HA is a viable target for cross-neutralizing binders that could be administered as passive immunotherapeutics in poultry. For porcine reproductive and respiratory syndrome virus (PRRSV), the GP5-M protein complex represents a glycoprotein target that could be approached with de novo designed single-domain antibodies (nanobodies). The biophysical principles remain identical: structural alignment of the target epitope across isolates, Rosetta-based interface optimization, and validation via binding assays [3].

A critical advantage of computational design is the elimination of animal immunization, aligning with ethical standards and reducing development time. The designed binders can be expressed in bacterial, yeast, or plant systems for economical production in veterinary settings. Additionally, the compact size of nanobodies (~15 kDa) allows better tissue penetration and lower production costs compared to conventional antibodies [1].

Conclusion

Computational design of viral glycoprotein binders has matured from a theoretical exercise to a practical tool for generating neutralizing agents. The integration of Rosetta energetic scoring, RFdiffusion backbone generation, and protein language models enables the creation of binders with affinities rivaling those from biological immune systems. For veterinary virology, these methods hold promise for controlling pathogens such as avian influenza, coronaviruses, and arteriviruses through passive immunization or as part of diagnostic capture agents. Continued refinement of energy functions and inclusion of glycoprotein dynamics will further enhance design success rates.

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.

References

[1] Swanson K, Wu W, Bulaong NL, et al. The Virtual Lab of AI agents designs new SARS-CoV-2 nanobodies. Nature. 2025. URL: https://pubmed.ncbi.nlm.nih.gov/40730228/

[2] He H, He B, Guan L, et al. De novo generation of SARS-CoV-2 antibody CDRH3 with a pre-trained generative large language model. Nat Commun. 2024. URL: https://pubmed.ncbi.nlm.nih.gov/39127753/

[3] Zahradník J, Schreiber G. Protein Engineering in the Design of Protein-Protein Interactions: SARS-CoV-2 Inhibitors as a Test Case. Biochemistry. 2021. URL: https://pubmed.ncbi.nlm.nih.gov/34196543/