Optimizing the breadth of SARS-CoV-2-neutralizing antibodies in vivo and in silico

ABSTRACT

Since the emergence of SARS-CoV-2, the ongoing arms race between mutating viruses and human antibodies has revealed several novel strategies by which antibodies adapt to viral escape. While SARS-CoV-2 viruses exhibit high variability in epitopes targeted by neutralizing antibodies, certain epitopes remain conserved owing to their essential roles on viral fitness. Antibodies can acquire broadly neutralizing activity by targeting these vulnerable sites through affinity-based somatic evolution of immunoglobulin genes. Notably, the specificity encoded in antibody germline genes also plays a fundamental role in acquiring the breadth. In-depth genetic and structural analyses of the antibody repertoires have uncovered multiple strategies for adapting to evolving targets. The integration of large-scale antibody datasets with computational approaches increases the feasibility and efficiency of designing broadly neutralizing antibody therapeutics from ancestral antibody clones with limited initial efficacy. In this review, we discuss strategies to optimize antibody breadth for the development of broadly neutralizing antibody therapeutics and vaccine antigens.

Introduction

SARS-CoV-2 virus caused a global health crisis, resulting in over 770 million COVID-19 cases and more than 7 million deaths worldwide, according to the WHO as of May 11, 2025.¹ The COVID-19 pandemic has underscored the critical need for effective vaccines and therapeutics to tackle the rapid evolution of the virus. SARS-CoV-2 infection triggers multiple layers of immune response, comprised of both innate and adaptive immunity.² Among the elicited immunity, neutralizing antibodies targeting the viral spike glycoprotein are crucial in preventing infection and viral transmission, and serving as an established immune correlate of protection against symptomatic disease.^3,4

The spike protein of SARS-CoV-2 is a trimeric glycoprotein that facilitates viral entry into host cells via interactions with angiotensin-converting enzyme 2 (ACE2).⁵ It exists in a dynamic equilibrium between prefusion and postfusion states, with conformational changes influencing the accessibility of neutralizing epitopes. The receptor-binding domain (RBD), which contains major epitopes for neutralizing antibodies, is highly susceptible to mutations that result in amino acid substitutions and deletions. Notably, SARS-CoV-2 has evolved multiple variants with mutations in the spike protein, particularly within the RBD, evading from neutralizing antibodies and posing challenges to the efficacy of both antibody therapeutics and vaccines.⁶

The molecular and structural basis of SARS-CoV-2 antibody responses provides key insights for the rational design of therapeutics and vaccines, a strategy that is potentially extended to other rapidly mutating viruses. Neutralizing antibodies against RBD epitopes often block the ACE2 binding directly, while the antibodies against non-RBD spike epitopes can also show neutralizing activity through other mechanisms. The evolutionary trajectory of the virus has led to the periodic emergence of antibody evading variants, increasing the demand for identifying and developing broadly neutralizing antibodies that retain potency against emerging variants.

Over the last decade, there have been rapid advances in high-throughput sequencing technologies and computational algorithms. Computer-aided antibody design has a long history.^7–9 However, data on the pairing of antibody light and heavy chains from B-cell repertoires has only become accessible in the past decade,^10,11 and the use of such big data for various medical and engineering applications has only just begun.^12,13 Along with repertoire data, SARS-CoV-2 represents the first instance in which an overwhelming amount of experimental information on a single antigen has become available, opening new avenues for antibody science.

Here, we provide an overview of the key molecular features of SARS-CoV-2 antibody responses, with a focus on the genetic and structural determinants of antibody breadth and potency. We discuss the evolutional interplay revealed by SARS-CoV-2 and its neutralizing antibody and highlight computational approaches that integrate structural modeling, machine learning, and antibody engineering to optimize therapeutic and vaccine strategies (Figure 1).

Figure 1.

Antibody generation in vivo and in silico. Antibody repertoires can be generated either in silico or in vivo, although the former approach is still in its early stages. In the fight against SARS-CoV-2, the ideal antibodies have broadly neutralizing activities capable of targeting multiple viral variants.

Antibody responses against emerging SARS-CoV-2 variants with escape mutations

Structural studies identified the RBD as a major epitope class.¹⁴ Serological analyses and deep mutational scanning experiments further elucidated immunodominant sites on the RBD, confirming that RBD antibodies account for the majority of neutralizing activity in sera.^15,16 On the other hand, only a single supersite on the N-terminal domain (NTD) dominates NTD-directed neutralizing responses,¹⁷ while the S2 subunit includes conserved epitopes targeted by broader but generally less potent neutralizing antibodies. RBD-targeting antibodies can block the interaction between spike protein and ACE2, making them potent neutralizers. Most SARS-CoV-2 vaccines are designed to induce spike antibodies, particularly RBD antibodies, to prevent infection and diseases.

Since 2020, numerous structures of RBD antibodies in complex with spike proteins or RBDs have been deposited into the Protein Data Bank (PDB), allowing the classification of antibodies into several groups. Barnes et al. categorized RBD antibodies into four classes based on their binding epitopes.¹⁴ Class 1 and 2 antibodies bind directly to the ACE2 receptor-binding site, with class 1 recognizing only open conformations and class 2 recognizing both open and closed conformations of the trimeric spike protein. Class 3 antibodies target conserved epitopes outside the ACE2 binding site, and class 4 antibodies bind cryptic epitopes accessible only in the open conformation.

Concordant with other RNA viruses, SARS-CoV-2 viruses accumulate nucleotide mutations and changes its viral behavior over time. Some mutations in the spike protein increase infectivity and/or immune evasion, leading to the emergence of multiple variants. Early in the pandemic, the Alpha variant (N501Y) spread rapidly and dominated globally due to increased infectivity. Subsequently, the Beta and Gamma variants with triple amino-acid substitutions in spike were sequentially found in different countries. These variants possess K417N/T, E484K, and N501Y mutations in the RBD and escaped from some spike antibodies in addition to enhanced ACE2 binding. The emergence of Omicron variants introduced more mutations in a relatively short time (~few months). BA.1 accumulated 15 RBD mutations (e.g., K417N, E484A, and Q493R), with profound evasion from neutralizing antibodies induced by pre-Omicron strains. BA.2 shared most BA.1 mutations but introduced unique substitutions (T376A, D405N, R408S), maintaining a similar level of antibody escape. Subsequent variants, such as BA.4/5 (L452R, F486V), BQ.1.1 (R346T, K444T, L452R, N460K), and XBB lineages (R346T, L368I, V445P, G446S, N460K, F486S/P, F490S), added mutations to further evade antibodies.

In late July 2023, BA.2.86 variant emerged with an astonishing number of spike mutations – more than 30 spike mutations relative to its ancestral BA.2, including 11 amino-acid substitutions and one amino acid deletion in the RBD.^18,19 This variant raised concern for immune evasion, but BA.2.86 displayed distinct properties, such as remarkably high ACE2 binding affinity with only moderate immune escape relative to the dominant XBB-derived variants. While BA.2.86 RBD introduced mutations (e.g., K356T for creating a new glycan site) and reverted a prior mutation (R493Q), BA.2.86 did not outcompete XBB-derived variants, likely due to its residual sensitivity to many neutralizing antibodies.^20,21 Under continued immune pressure, the BA.2.86 sub-lineage JN.1 acquired a critical RBD mutation (L455S), which lies in the ACE2-binding site, and rapidly became dominant.²² This mutation allowed JN.1 to significantly evade neutralizing antibodies targeting that region compared to BA.2.86. Although L455S slightly reduced ACE2 binding affinity, the trade-off in immune evasion appears to provide a selective advantage. JN.1 showed a growth advantage over both BA.2.86 and other competing variants in the populations with diverse immune histories.

Some RBD mutations simultaneously influence both ACE2 binding and immune escape. A notable example is the “FLip” mutation (L455F and F456L), involving adjacent amino acids in the ACE2 binding site. This mutation affects ACE2 binding and is observed in major Omicron subvariants (e.g., GK.1.1 and HK.3). Structural analysis of XBB.1.5 RBD with an experimentally introduced FLip mutation revealed reorganization of the RBD-ACE2 interaction, contributing to enhanced ACE2 binding.²³

Antibody escape mutations can be rationalized through the structural and physicochemical changes in the spike protein. Studies using cryo-electron microscopy (cryo-EM) and X-ray crystallography have revealed several mechanisms by which Omicron mutations evade antibody binding (Figure 2a): (1) Reducing geometric complementarity. Many Omicron RBD mutations physically clash with antibody binding. For example, G446S inserts a side chain that would collide with the CDRs of many class 3 antibodies.^25,26 A cryo-EM study also showed that N440K sterically disrupts the contact with a class 3 antibody (C135)²⁷ whereas F486P observed in XBB.1.5 appears to sterically prevent binding by a class 1 antibody (NIV-10).²⁸ (2) Reducing electrostatic complementarity. Several mutations in the Omicron RBD (G339D, K356T, K417N, N440K, T478K, N481K, E484A, Q493R, and Q498R) alter electrostatic properties. For instance, K417N eliminates a salt bridge between the residue at position 417 on the RBD and negatively charged residues in class 1 antibodies. This single change was sufficient for Beta and Omicron BA.1 to completely escape from a group of class 1 antibodies.²⁵ Likewise, E484A removes a charged residue that many class 2 antibodies rely on for binding, thus reducing their affinity.²⁵ Structural analysis of BA.2.86 revealed substantial electrostatic changes on the RBD surface relative to BA.2,²⁹ which likely perturb antibody-binding interfaces. (3) Reducing hydropathic complementarity. Hydrophobicity is a primary driving force for protein folding and binding. Antibody–antigen interfaces exhibit a mixture of hydrophobic and hydrophilic properties. A study on B-cell repertoires in rabies-vaccinated individuals suggested that hydrophilic epitopes tend to induce antibodies with hydrophilic paratopes.³⁰ In the context of SARS-CoV-2, the paratopes in neutralizing antibodies that retained the binding activity up to Omicron BA.5 tended to exhibit higher hydrophobicity than those that lost the binding to Omicron BA.5 mutations.²⁸ This findings suggest that hydrophobic paratopes contribute to the neutralizing breadth up to BA.5 variants by increasing the binding toward the hydrophobic RBD epitopes, which are observed from ancestral to BA.5 strains, corresponding to class 1 and class 4 epitopes (Figure 2b). This hydrophobic nature of the RBD surface was suggested to correlate with the frequency of up and down states of RBDs in the trimeric spike protein.³¹

Figure 2.

Strategies for the immune escape of viral proteins. (a) Physicochemical changes observed in SARS-CoV-2 RBD, including reductions in geometric, electrostatic, and hydropathic complementarity. (b) Evolution of surface hydrophobicity in SARS-CoV-2 RBD. The ACE2-binding site is centered in each panel. Hydrophobicity was calculated by the spatial aggregation propensity metric.²⁴

The structural and functional rationale for IGHV3-53/66-based neutralization

After infection or vaccination, certain antibody gene rearrangements recur across individuals, forming so-called convergent or public clones. In the antibody response against SARS-CoV-2, IGHV3-53, and IGHV3-66 have been identified as dominant public germlines (IGHV3-66 differs from IGHV3-53 by a single conservative mutation, H-V12I).^32,33 These antibodies target the ACE2 binding site on the RBD and have shown their potent neutralizing activity against several variants. Importantly, IGHV3-53/66 antibodies were identified in multiple convalescent individuals³⁴; an analysis of 1,593 anti-RBD monoclonal antibodies from 32 studies found IGHV3-53/66 to be the most frequently used IGHV genes among potent neutralizers.³⁵ Even after the emergence of Omicron variants, this trend persists. An analysis of 141 BA.1 neutralizing antibodies, isolated from seven individuals who had received two or three doses of the mRNA vaccine followed by BA.1 breakthrough infection, showed that 18.4% (26/141) used the IGHV3-53/66 germline genes.³⁶ Similarly, another study of 545 neutralizing antibodies, isolated from individuals following BA.1 breakthrough infection, identified 28 antibodies capable of neutralizing BA.1 with IC₅₀ < 100 ng/mL, of which 32.1% (9/28) used the IGHV3-53/66.³⁷ A similar trend was observed in individuals who had received three doses of the ancestral mRNA vaccine, followed by repeated BA.5 exposure via breakthrough infection and ancestral/BA.5 bivalent booster.³⁸ However, the prevalence of IGHV3-53/66 varies based on immune history. Jian et al.³⁹ conducted a study in convalescent individuals in China who had received an inactivated vaccine instead of mRNA vaccine. Consistent with other studies, antibodies reactive to the ancestral strain showed significant IGHV3-53/66 usage in vaccinated individuals with BA.5 breakthrough infection. In contrast, IGHV3-53/66 was rarely observed in the individuals without vaccination, where the initial antigen exposure occurred through Omicron infection. Given their prevalence and efficacy, IGHV3-53/66 antibodies have been intensively studied for their structural features, breadth of neutralization, and therapeutic potential.

As seen in other neutralizing antibodies, the neutralization breadth of IGHV3-53/66 antibodies can be limited by spike mutations. These antibodies predominantly bind the ACE2 receptor-binding site on the RBD, thereby susceptible to mutations around ACE2 binding site. Notably, single RBD mutations such as K417N or E484K found in Beta and Gamma variants dramatically reduced neutralizing potency of many IGHV3-53/66 antibodies.³⁵ The Omicron BA.1 lineage escaped over 85% of tested monoclonal antibodies, including IGHV3-53/66 antibodies.²⁵ Thus, while IGHV3-53/66 antibodies are highly potent against early strains, their breadth against divergent variants was limited.

On the other hand, affinity maturation can broaden the neutralization spectrum of IGHV3-53/66 antibodies. With additional exposure by booster vaccination or breakthrough infection, B cells utilizing IGHV3-53/66 genes accumulate somatic hypermutations that enable recognition of mutated RBDs. For instance, memory B cells induced by the ancestral mRNA vaccine further evolved their antibodies, including those from IGHV3-53/66, through somatic hypermutations, allowing redirection of the antibody specificity against the Omicron variants after two Omicron exposure.^38, In separate study, the most potent antibodies from Omicron-convalescent individuals utilized IGHV3-53/66 and neutralized most of the tested variants, including BA.4/5, BQ.1, and XBB lineages.⁴⁰ The most potent one exhibited extraordinary breadth against ancestral, Delta, BA.1, BA.2, BA.3, BA.4/5. BA.2.75. BF.7, BQ.1, XBB, XBB.1, XBB.1.5, XBB.1.16, EG.5, EG.5.1, FL.1.5, FL.1.5.1, and HK.3. These broadly neutralizing IGHV3-53/66 antibodies carried unusually higher levels of somatic hypermutation, suggesting the importance of somatic hypermutations in achieving the breadth. An additional example is obtained from an IGHV3-53-derived antibody, D6,⁴¹ originated from an infection of the ancestral strain. This antibody potently neutralizes Omicron subvariants BA.1, BA.2, BA.4/5, BF.7 and retains activity against XBB, albeit with some reduced potency. Of note, the D6 antibody carried a higher number of somatic hypermutations compared to a highly similar antibody showing a narrower breadth.

IGHV3-53/66 antibodies have several distinct features^34,42: (1) They often possess a shorter CDR-H3 (≤10, Kabat/Chothia definitions), which potentially results in a geometrically flat paratope. (2) Germline-encoded CDR-H1 and CDR-H2 motifs (NY and SGGS, respectively) play important roles in RBD recognition. It has been suggested that IGHV3-53/66 antibodies exhibit low levels of somatic hypermutations due to their germline-encoded motifs for RBD recognition. Although IGHV3-53/66 antibodies that cross-react with many variants have many somatic hypermutations,^40,41,43 only part of these mutations is likely to significantly contribute to the RBD binding. Therefore, identifying these beneficial mutations is an important issue to be addressed for improving the reactivity of IGHV3-53/66 antibodies.

Since the emergence of the JN.1 variant, many of IGHV3-53/66 antibodies were evaded through the L455S mutation.⁴⁴ In addition, IGHV3-53/66 antibodies were significantly evaded by more recent Omicron variants, such as KP.2 and KP.3 due to the specific mutations of L455S, F456L, and Q493E.³⁹ Only limited numbers of IGHV3-53/66 antibodies remain effective, and the most potent IGHV3-53/66 antibody identified so far is BD55-1205,⁴⁵ which was isolated from an individual exposed to the ancestral strain and exhibits neutralizing potency against ancestral, BA.1, BA.2, BA.5, BQ.1.1, XBB.1.5, HK.3.1, JN.1, KP.2, and KP.3. It possesses a nine-residue short CDR-H3, a typical feature of IGHV3-53/66 antibodies, and many physical contacts with the RBD backbone, including backbone-mediated hydrogen bonds at the antibody–antigen interface.

Computer-guided mining of B-cell repertoires to identify antigen-specific antibodies

The COVID-19 pandemic led to rapid advances in computational immunology and antibody engineering. Since 2020, academia and industry have leveraged in silico techniques to identify potent SARS-CoV-2 neutralizing antibodies from B-cell repertoires and to design optimized antibodies targeting the viral spike protein. These approaches have accelerated antibody discovery and increased the resistance against viral variants, which complement traditional lab-based antibody generation (Figure 3).

Figure 3.

Computer-guided B-cell repertoire mining and antibody design. Computational approaches can identify potent antibodies from deep sequenced B-cell repertoires and also refine existing antibodies. Structure-based design that starts from an antibody-antigen complex is ideal, but even an antibody sequence alone can be optimized with sequence-based machine learning methods.

High-throughput sequencing of B-cell receptor (BCR) genes has enabled the profiling of millions of antibody sequences, providing a fingerprint of immune history and activity.^13,46,47 Over the past decade, various computational techniques, such as clustering algorithms, statistical modeling, and machine learning (ML), have been developed to decode these repertoires.^48,49 Modern repertoire mining leverages computational approaches to detect meaningful patterns in antibody sequence data that would be impossible to identify by eye.

Two related goals have emerged for computational analysis: (1) identifying antibodies specific for pathogens (e.g., SARS-CoV-2) and (2) detecting immune signatures indicative of disease or vaccination status. While these objectives differ in focus, they share essential computational approaches (e.g., clustering, statistical modeling, machine learning). These computational approaches leverage a shared conceptual framework: detecting meaningful patterns hidden in the immense diversity of antibody sequences. Clustering algorithms can identify antibody convergence across multiple individuals responding to SARS-CoV-2 infection or vaccination, guiding the discovery of public neutralizing antibodies. Simultaneously, the same methods uncover subtle repertoire-level signatures indicative of disease status or healthy condition. For instance, deep learning methods trained on repertoire sequences have successfully identified SARS-CoV-2-specific antibody responses,⁵⁰ as well as general immune signatures diagnostic of autoimmune diseases.⁵¹ Given our central focus on generating broadly neutralizing SARS-CoV-2 antibodies, subsequent sections will illustrate computational approaches primarily toward antibody discovery, highlighting representative studies employing machine learning and structural modeling to discover or enhance SARS-CoV-2 neutralizing antibodies (Figure 3).

Early in the pandemic, large antibody databases such as CoV-AbDab⁵² and Observed Antibody Space (OAS)⁵³ compiled thousands of coronavirus-binding antibodies. Structural information of these anti-SARS-CoV-2 antibodies, with or without the antigenic proteins, has also been accumulated in the PDB.⁵⁴ Although these public databases sometimes contain erroneous information,⁵⁵ these data provide a valuable starting point for data-driven models in computer-aided antibody design and identification. When sufficient learning data with appropriate labels (e.g., disease status and information on cross-reactivity) are available, supervised learning can classify repertoires by disease or predict antigen specificity. As SARS-CoV-2 is the first example ever where much experimental information became publicly available, there are a variety of computational problem settings in the literature. Early efforts in “specificity prediction” focused on predicting epitopes rather than predicting which variants can be neutralized by antibodies. B-cell epitope prediction is highly complex and remains an unsolved problem. Any amino acid residues can, in principle, be immunogenic if they are located at the surface of the antigenic proteins.⁵⁶ Predicting B-cell epitopes on SARS-CoV-2 RBD, however, may be a straightforward problem to solve because the epitopes on the RBD appear to be limited,¹⁴ the epitopes highly depend on their germline genes, and there are a number of reference structures in the PDB. Computational structural modeling methods (SPACE⁵⁷ and its successor SPACE2⁵⁸ were developed to cluster antibodies that bind to the same epitope, even when antibody sequences differ significantly. These methods outperform sequence-based predictions by incorporating predicted antibody structure information. They identify functional convergence among antibodies from diverse lineages and species, demonstrating that structural data enhance predictions of epitope specificity beyond sequence-only methods.

A few studies attempted to classify epitopes more broadly. A systematic survey of 7,997 SARS-CoV-2 antibodies was able to identify recurring molecular features (public clonotypes).⁵⁰ They showed distinct convergent sequence features across RBD, NTD, and S2 domains and provided proof-of-concept for specificity prediction using deep learning models. These insights emphasize the significance of convergent sequences and somatic hypermutation patterns in predicting antibody specificity. Similarly, Saputri et al. proposed clustering antibodies based on similarity in their CDRs to predict antigen specificities, enabling the identification of epitope-specific antibodies.⁵⁹ The epitope prediction problem was formulated as to whether antibodies bind to RBD or non-RBD epitopes within the spike proteins. Their method was validated on COVID-19 BCR data and vaccination-induced responses, suggesting that CDR-based clustering can robustly annotate unknown antibodies with antigen specificity (i.e., anti-RBD antibodies vs. others). Language model-based embeddings, such as BCR-specific embeddings and protein transformer models like ESM2⁶⁰ and ProtT5,⁶¹ effectively encode BCR specificity.⁶² These embeddings significantly improve specificity prediction tasks, particularly when considering paired-chain data (heavy and light chains). The study indicates that embeddings derived from machine learning language models hold promise for improved accuracy in predicting BCR specificity and can greatly enhance antibody discovery and analysis pipelines.⁶² These computational methods that predict epitopes would be useful to derive starting antibodies that can be further computationally designed to overcome immune escapes.

A generative adversarial network (GAN) is a type of neural network model for generative artificial intelligence. A GAN-driven synthetic antibody library was developed to discover therapeutic antibody candidates.⁶³ Combined with phage panning experiments, this platform rapidly identifies potent antibodies toward SARS-CoV-2 RBD.⁶⁴ The selected antibodies were experimentally validated, and the most potent antibody, derived from IGHV3-30, neutralized ancestral, D614G, Alpha, and Delta as well as Omicron BA.5 variants. In another study, a random forest model was trained on experimental nanobody screening data against pre-Omicron variants to identify effective antibodies against newer Omicron variants.⁶⁵ The selected antibodies were experimentally validated, and two potent antibodies were identified. A designed antibody is effective against Omicron BA.1, BA.5, Beta, Gamma, and Mu, but not Delta and Lambda, while the other neutralized Omicron BA.1, BA.5. In both studies, the newer variants later than BA.5 were not tested. While Loomis et al.⁶⁴ and McIlroy et al.⁶⁵ provided comprehensive variant-specific antibody neutralization data, particularly useful for discussing cross-reactivity and broad neutralization potential, the early works on the “specificity prediction” problems did not explicitly address specificity for SARS-CoV-2 variants.^50,57–59 Considering the apparently limited epitope space of SARS-CoV-2 RBD and the rapid evolution of the virus, predicting antibody escapes is a far more difficult problem that needs to be immediately addressed.

When labeled data are unavailable or insufficient for supervised training, bioinformatic, unsupervised techniques can still cluster similar antibody sequences that expand after infection. These approaches exploit the observation as mentioned earlier that unrelated individuals often generate public clonotypes in response to the same antigen, as seen in IGHV3-53/66 antibodies for SARS-CoV-2. A striking example is the work of Abbate et al., who attempted to identify groups of convergent antibodies in COVID-19 patient repertoires using sequence similarity with a graph clustering method.⁶⁶ The convergent sequences were statistically enriched and likely specific to SARS-CoV-2. Although no experimental validations were conducted on isolated sequences, some of the identified antibody sequences were indeed present in the COVID-19 database and confirmed as binders to SARS-CoV-2 epitopes.

Overall, ML and structural modeling techniques have enabled the mining of immense B-cell repertoires to identify neutralizing antibodies in silico, which can then be synthesized and tested experimentally. This dramatically reduces the search space before experimental validation.

Computer-guided structure-based antibody design for increasing the resistance to viral escapes

Once candidate antibody sequences are identified, structural modeling can be used to predict the 3D structure of the antibodies and their binding modes to antigen structures. While simulation-based methods for antibody structure prediction take a lot of time even with supercomputers,⁶⁷ limiting repertoire-scale analysis of antibody structures, recent advancements in ML-based antibody structure prediction, such as AbodyBuilder,⁶⁸ Repertoire Builder,⁶⁹ and Ig-Fold,¹²⁴ have significantly improved speed while steadily enhancing accuracy, although there is still much room for improvement. The emergence of AlphaFold has had a huge impact across various fields,⁷⁰ and this is also true for antibody engineering. Several groups independently showed that AlphaFold can be adapted to predict antibody-antigen complexes.^71,72 Gao and Skolnick developed AF2Complex,⁷³ a deep learning framework that inputs antibody and antigen sequences and predicts if and how they interact. In tests on ~1,000 antibodies, this method correctly identified 32% of the true binders (recall of 32%) to the SARS-CoV-2 spike at a precision (the percentage of predicted positive results that are truly positive) of 82%.⁷⁴

Molecular dynamics (MD) simulations complement static structural modeling by revealing the dynamic behavior of antibody–antigen interactions at the atomic level. MD allows us to refine docked complexes and assess stability, binding free energy, and the neutralization mechanism. In the context of SARS-CoV-2, MD simulations became a useful tool to evaluate how antibody-spike complexes tolerate mutations. Cao et al. used MD simulations to understand the structural dynamics and interactions between fully glycosylated SARS-CoV-2 spike trimers and various neutralizing antibodies.⁷⁵ Their computational analysis suggested critical residues for spike–antibody interactions, providing insights into how specific mutations that occurred in SARS-CoV-2 variants might influence antibody-binding efficacy. They also revealed how glycans on the spike protein can either facilitate or prevent antibody binding, highlighting the dynamic roles of glycans in spike–antibody interactions.⁷⁵

MD simulations have also helped understand how antibodies neutralize the virus. A study by Fredericks et al. combined deep BCR sequencing and MD simulations to explore why certain patient-derived antibodies were protective.⁷⁶ They sequenced B cells from COVID-19 patients and found survivors had an expanded clone of an anti-RBD antibody classified as class 3. MD simulations of the antibody bound to the spike RBD showed that, even though the antibody does not directly block the ACE2 receptor site, it locks the spike in conformations that prevent ACE2 attachment. In other words, the antibody neutralized the virus by an allosteric mechanism. This mechanistic insight obtained by simulating multiple antibody-spike conformational states highlights a role of MD simulations in elucidating neutralization strategies. More broadly, MD simulations and free energy calculations have been instrumental for in silico affinity maturation, enabling the prediction of antibody mutations that strengthen antigen binding.^77–79 By analyzing the atomic interactions at the antibody–spike interface, computational approaches can suggest CDR modifications to improve affinity or breadth, which can then be experimentally validated.

While vaccinations and infections produce neutralizing antibodies, the immune pressure by these antibodies can drive virus evolution. Therefore, expanding the neutralizing potency of existing antibodies through rational design is highly desirable (Figure 3). As antibodies evolve in the body, they can also be evolved through computational design calculations.^8,80 Early attempts at computational design for SARS-CoV-2 were to redesign anti‐SARS‐CoV antibodies based on molecular simulations and crystal structures available at that time, which resulted in limited success.^81,82

More recently, as antibodies from SARS-CoV-2 convalescent individuals and vaccine recipients became available, studies focused on redesigning existing anti‐SARS‐CoV-2 antibodies. These antibodies often exhibited limited potency against newly emerging SARS‐CoV‐2 variants. Thus, efforts have focused on designing antibodies to enhance neutralizing potency against emerging variants (Table 1).^28,83–88 A key challenge is that design calculations for a single antigen variant (i.e., single-state design) can compromise binding to the original antigen while acquiring binding affinity for the target variant. There are two computational approaches to address this issue. The first approach is the multi-state design (MSD),⁸⁹ where two or more antigen “states” are considered during design calculations. The goal of MSD is to identify an antibody sequence that adopts stable structures across multiple states. MSD minimizes the computed energy for each of these states simultaneously, ensuring that the design is energetically favorable regardless of the antigen variant encountered. MSD approaches are often employed to incorporate backbone flexibilities^90,91 and to design bispecific proteins.^92,93 In the context of SARS-CoV-2 neutralizing antibodies, these “states” can include complex forms with one or more variants, respectively. There are several software for MSD calculations,^90,94–97 but, to our knowledge, studies explicitly employing MSD have not yet been reported concerning SARS-CoV-2. The second approach is to screen designed antibodies with various antigen variants after the initial design with a target antigen. In the screening step, computational matrices derived from antibody-antigen complexes in the PDB are essential.⁹⁸ For example, a neutralizing antibody (IGHV3-7, class 1 epitope) obtained from a SARS-CoV-2 convalescent individual showed neutralizing potency against Omicron subvariants such as BA.5 and BQ.1.1 but was less effective against the XBB subvariants.²⁸ To address this, the FastDesign algorithm in Rosetta protein modeling suite⁹⁹ was employed to computationally design approximately 2,000 variants. These design variants were evaluated based on physicochemical and geometrical metrics, including interaction energy (<−35 kcal/mol), shape complementarity (Sc > 0.65), and the number of hydrogen bonds at the interface (>4). A computationally optimized antibody showed enhanced neutralizing potencies against multiple variants (Ancestral, Alpha, Beta, Delta, BA.1, BA.2, BA.5, BA.2.75, BA.2.75.2, BQ.1.1, XBB, XBB.1.5).²⁸ In another study, a structure-based design approach with RosettaAntibodyDesign¹⁰⁰ was applied to broad sarbecovirus-neutralizing antibody S2H97 (IGHV5–51, the epitope class cannot be classified).⁸⁴ The designed antibodies were analyzed with interaction energy and the number of hydrogen bonds at the interface (though the specific values of each criterion are not described) and experimentally validated through SPR and pseudovirus assays, which resulted in an antibody that exhibited significantly improved neutralization breadth and potency across SARS-CoV-1, SARS-CoV-2 (Alpha, Beta, Gamma, Delta, BA.5, BA.2.3.20, BQ.1.1, XBB), and related bat coronaviruses (RaTG13 and WIV1).

Table 1.

Commuter-guided structure-based antibody design of anti-SARS-CoV-2 RBD antibodies.

Initial structures	Method	Resolution	VH genes	Epitopea	Specificity of the most potent antibodyb
NIV-10 (8HES)	X-ray crystallography	2.20 Å	IGHV3-7	1	Ancestral, Alpha, Beta, Delta, BA.1, BA.2, BA.5, BA.2.75, BA.2.752, BQ.1.1, XBB, XBB.1.528
UT28K (7X7O)	X-ray crystallography	3.75 Å	IGHV1–58	1	Ancestral, BA.183
S2H97 (7M7W)	X-ray crystallography	2.65 Å	IGHV5–51	NAc	SARS-CoV, Alpha, Beta, Gamma, Delta, BA.5, BA.2.3.20, BQ.1.1, XBB, bat coronaviruses (RaTG13 and WIV1)84
P36-5D2 (7FAF/7FAE)d	Electron microscopy	3.69 Å/3.65 Å	IGHV1–3	3	D614G, Alpha, Beta, Gamma, Delta, Kappa, Epsilon, Eta85
COV2–2130 (7L7E)	X-ray crystallography	3.00 Å	IGHV3-15	3	D614G, Delta, BA.1, BA.1.1, BA.2, BA.2.12.1, BA.4, BA.5, BA.5.586
8G3 (9JS4)	Electron microscopy	3.80 Å	IGHV3-53	1	Alpha, Beta, Gamma, Delta, Kappa, BA.1, Omicron BA.2.75, BQ.1, BQ.1.1, XBB, CH.1.1, JN.187
Etesevimab (7C01)	X-ray crystallography	2.88 Å	IGHV3-66	1	BA.1, BA.2, BA.2.12.1, BA.2.75, BA.4/5, BF.7, BA.4.6, BQ.1, BQ.1.1, XBB, XBB.1.5, EG.5, BA.2.86, KP.288

1. ^(a)Epitope classes defined by Barnes et al.¹⁴
2. ^(b)Only variants that were explicitly validated in the literature are shown.
3. ^(c)S2H97 antibody cannot be classified into the epitope classes of Barnes et al.¹⁴
4. ^(d)PDB codes used in the design calculations were not explicitly stated in the literature.⁸⁵

Recently, machine learning has gained significant attention in computational protein design.¹⁰¹ Approaches like transformer-based language models, which are trained on vast antibody sequence databases, can sample novel CDR sequences that exhibit desirable properties. Following this trend, based on structural and binding experimental data from the SKEMPI 2.0 database,¹⁰² a geometric neural network model was developed to predict how amino acid changes in the CDRs can impact binding affinity and neutralization against different SARS‐CoV‐2 variants.⁸⁵ By combining this predictive model with experimental validations, Shan et al. were able to optimize an antibody sequence (IGHV1–3, class 3 epitope), identifying mutations that enhance its ability to fight various SARS‐CoV‐2 variants (D614G, Alpha, Beta, Gamma, Delta, Kappa, Epsilon, Eta). Likewise, Liao et al. employed the Relax application in Rosetta⁹⁹ and online resources of pre-trained machine learning models, GeoBiologics and mmCSM-PPI,¹⁰³ to restore neutralizing potency against the JN.1 variant, starting from an existing neutralizing antibody (IGHV3-53, class 1 epitope).⁸⁷ They identified 11 single mutations that showed improved neutralizing activities toward the JN.1 variant. Among these computationally identified mutations, a manual combination of four mutations resulted in a 1,500-fold increase in neutralizing potency compared to the wild-type antibody (Alpha, Beta, Gamma, Delta, Kappa, BA.1, BA.2.75, BQ.1, BQ.1.1, XBB, CH.1.1, JN.1).

Affinity maturation through design calculations sometimes sacrifices other physicochemical and biological properties. Desautels et al. employed various computational design approaches, including Rosetta, FoldX,¹⁰⁴ MD simulations, to re-optimize a clinical antibody, COV2-2130 (IGHV3-15, class 3 epitope).⁸⁶ Thermal stability and humanness of the designed antibodies were also evaluated through free energy perturbation¹⁰⁵ and a deep learning language model, AbBERT,¹⁰⁶ respectively. COV2-2130 was a potent anti-RBD antibody that failed against Omicron BA.1/BA.1.1 due to the viral mutations. The design calculations for binding were performed against Delta, BA.1 and BA.1.1 RBDs, respectively, which was followed by the stability and humanness assessments. In total, 376 designed antibodies were experimentally validated. In vitro, a redesigned antibody possessing four mutations successfully restored neutralization against Omicron BA.1 and BA.1.1, while also enhancing potency against the other variants tested (D614G, Delta, BA.1, BA.1, BA.2, BA.212.1, BA.4, BA.5, BA.5.5). Importantly, it conferred protection in mouse models challenged with SARS-CoV-2 (D614G, BA.1.1, BA.5). The thermal-shift assay also suggested that the most potent designed antibody exhibited a comparable melting temperature to the parental wild-type antibody. Deep mutational scanning experiments further confirmed the redesigned antibody had a broadened resistance profile. The cryo-EM structure of the most potent designed antibody with the BA.2 RBD further helped explain the observed restoration of potency to early Omicron variants.

While protein side chains can be sampled quickly using a rotamer library,¹⁰⁷ most of the current methods for computational protein design assume a relatively rigid protein backbone. This is, in many cases, valid because backbone motions are indeed small when comparing backbone conformations between wild-type proteins and the corresponding mutated proteins observed in high-resolution crystal structures.¹⁰⁸ However, these subtle motions may not be enough to mitigate steric hindrance that could emerge at the antibody–antigen interfaces upon SARS-CoV-2 evolution, and more drastic conformational changes of protein backbones may be necessary. In this context, through backbone remodeling of CDR-H3 by Rosetta,¹⁰⁹ Su et al. successfully restored the broad neutralization activity of an antibody, etesevimab (IGHV3-66, class 1 epitope), against multiple challenging Omicron subvariants, notably KP.2.⁸⁸ One of the designed variants exhibited significantly enhanced neutralization against all the variants tested (BA.1, BA.2, BA.2.12.1, BA.2.75, BA.4/5, BF.7, BA.4.6, BQ.1, BQ.1.1, XBB, XBB.1.5, EG.5, BA.2.86, KP.2).

Conclusion and perspectives

High-throughput sequencing technologies, together with computational methods, prompted the identification of potent antibodies from large-scale B-cell repertoires, which can be further modified for various applications. With rapid advances in artificial intelligence technologies, it may become possible to generate synthetic antibody repertoires purely in silico. There are several such efforts in the literature, but most of them rely only on computational evaluations and still lack experimental validation.^110,111

The studies on computer-aided antibody design in this review primarily focused on existing antibodies identified in human B-cell repertoires. As noted in the previous section, the neutralization breadth of cilgavimab (COV2-2130)⁸⁶ and etesevimab⁸⁸ has been successfully enhanced through computational optimization. Similar strategies could be applied to broaden the activity of therapeutic antibodies such as bamlanivimab, casirivimab, imdevimab, sipavibart, sotrovimab, and tixagevimab, all of which received emergency use authorization or full approval before being escaped by emerging variants.

Another approach in the context of computational antibody design is the de novo generation of entirely new antibody sequences. Since antibodies are highly conserved in both sequence and structure, the primary focus on de novo antibody design is to generate sequences of CDRs, especially CDR-H3, that specifically bind to target antigens. In this context, He et al. employed large language models (LLMs) to create new antibody sequences. They focused on generating diverse CDR-H3 sequences optimized for binding the SARS-CoV-2 RBD, followed by structural modeling and docking to select design candidates.¹¹² Experimental validation confirmed that some in silico-generated antibodies neutralized the variants tested (ancestral, Alpha, Delta, XBB). Other de novo design studies on antibodies have reached similar milestones, although extensive wet-lab work is still required for validation and improvement.^113,114 At present, the main goal of de novo design is to generate novel binders for specific targets rather than to achieve broad neutralization. Higher-accuracy de novo binder design is expected soon, but attaining broad potency will likely demand breakthroughs in both algorithmic innovation and our mechanistic understanding of broadly neutralizing antibodies. Sequence-based maturation of in silico-generated antibodies, without structural information, could also be leveraged as the volume of available antibody sequence data continues to grow.^115,116

Beyond LLMs, another trend in computational protein design is the use of diffusion models for de novo protein design.^117–121 In a study by Baker et al., single-domain antibodies were computationally generated against various antigens, including the SARS-CoV-2 RBD.¹²² Since the overall structures are conserved among antibodies, the framework regions of single-domain antibodies were derived from a previously published humanized VHH framework,¹²³ and the associated CDRs were designed de novo. However, despite their elegant success in the de novo design of single-domain antibodies, the de novo design of antibodies possessing both heavy and light chains remains challenging. It is also noteworthy that, while their computational success is impressive, they experimentally evaluated 9,000 design constructs per target antigen using a high-throughput yeast surface display system. Therefore, high-throughput experimental validation remains a key challenge in B-cell repertoire mining and computer-aided antibody design.

Biographies

Dr. Daisuke Kuroda was a senior scientist at the Research Center for Drug and Vaccine Development at the National Institute of Infectious Diseases in Japan and, since 2025, has served as an associate professor at Nihon University. His background in bioinformatics expanded to include protein engineering and structural biology, integrating experimental approaches with computational techniques. His current research focuses on computer-aided antibody design and immune repertoire analyses, advancing our understanding of antibodies and enabling next-generation therapeutics. He has published over 70 papers on antibody design, protein interactions, and drug discovery.

Dr. Saya Moriyama is laboratory chief of the Laboratory of Vaccine Evaluation, Research Center for Vaccine Development at NIID, JIHS, and a guest professor at Osaka University. She has worked on basic immunology and infectious immunology for nearly two decades and published more than 50 peer-reviewed papers. Her research focuses on broadly neutralizing antibodies against viruses, and vaccine development and evaluation.

Dr. Yoshimasa Takahashi is director of the Research Center for Vaccine Development at NIID, JIHS, and guest professor at seven universities. His research focuses on dissecting the humoral, cellular, and innate immune responses that underly the efficacy and adverse reactions of active/passive immunization. His team has identified broadly protective antibodies against influenza, SARS-CoV-2, and other viruses. Several vaccines and antibody therapeutics are under development using structure- and computer-aided approaches. He has over 150 publications and serves as scientific advisory committees for national and international organization.

Funding Statement

This work was supported by the Japan Agency for Medical Research and Development [JP23wm0325047 and JP25wm0325075 to D.K., JP22fk0108556 to S.M., and JP25gm1810004, JP243fa627009, JP243fa627005, and JP243fa727002 to Y.T.] and by grant from Japan Ministry of Health, Labour and Welfare [23HA2021 to Y.T.].

Disclosure statement

No potential conflict of interest was reported by the author(s).

Ethical approval

This work does not involve human participants and it was not necessary to obtain ethical approval from the ethics committee.