The search strategy utilized targeted strings designed to capture the intersection of machine learning and structural virology, including: AI-directed protein design, machine learning for antibody optimization, generative modeling for SARS-CoV-2 antigens, and de novo protein design COVID-19.

Studies were primarily included based on their contribution to the methodological shift toward data-driven paradigms or their direct impact on the accelerated development of vaccines and antibody therapeutics published between 2020 and late 2025. Of the 45 total references cited in this review, 17 are utilized in the Introduction to establish the broader scientific and historical context. The remaining 28 references were specifically selected and analyzed to describe the technical evolution of AI/ML-based strategies in the field of computational protein design. To maintain a focused scope, articles focusing exclusively on clinical outcomes or epidemiological trends without a substantial computational design or structural modeling component were excluded.

The quality and technical rigor of the selected literature were assessed based on three primary pillars: Peer-review status, clarity of the machine learning architectures employed, and experimental or biophysical validation to ground the reported computational predictions. The insights gathered from this literature selection are synthesized to discuss how these advancements inform the development of truly pandemic-ready biologicals, as schematically summarized in Figure 1.

Sequence-based modeling constitutes the foundation of many AI-driven protein design approaches. By learning statistical patterns from large collections of natural protein sequences, these models capture evolutionary constraints associated with folding, stability, and function. Recent deep learning architectures have enabled sequence models to represent long-range dependencies, contextual relationships, and epistatic interactions that were previously difficult to capture using classical bioinformatics methods [6, 10, 20, 21]. During the COVID-19 pandemic, sequence-based analyses were widely applied to rapidly screen antibody repertoires, identify conserved viral regions, and prioritize targets less susceptible to immune escape [1, 2, 5, 12, 22]. These approaches were particularly valuable in the earliest phases of the outbreak, when structural information was scarce but viral genomic sequences were rapidly accumulating. In this context, sequence-based models enabled rapid hypothesis generation and target prioritization at a time when experimental resources were limited.

Sequence-based approaches have also proven useful for assessing mutational tolerance and identifying regions of functional constraint across viral proteins. Such analyses informed the selection of epitopes likely to remain stable under immune pressure, thereby supporting the design of more durable vaccines and antibody therapeutics [11, 23, 24]. However, these predictions are inherently probabilistic and depend strongly on the diversity and representativeness of training data.

Despite their utility, sequence-only models have inherent limitations. They may fail to fully capture three-dimensional interactions, conformational dynamics, post-translational modifications, and context-dependent effects that strongly influence binding and stability [4, 7, 10, 15, 16, 25]. Consequently, sequence-based predictions are most effective when used as an initial filtering step, guiding subsequent structure-aware analyses rather than serving as standalone decision tools. Recognizing these strengths and limitations is essential for their effective translational application.

Structure-based modeling leverages three-dimensional information to predict protein folding, binding interactions, and conformational changes that underlie biological function. The availability of high-quality experimental structures, combined with major advances in deep learning–based structure prediction, dramatically expanded the applicability of structure-guided design during the COVID-19 pandemic [4, 7, 11, 19, 21]. Structural models of the SARS-CoV-2 spike protein and its complexes with neutralizing antibodies enabled rational identification of binding hotspots, informed affinity maturation strategies, and supported the design of stabilized antigens [3, 6, 13, 26]. Beyond static representations, structure-based approaches facilitated comparative analysis of antibody binding modes and epitope accessibility, providing mechanistic explanations for differences in neutralization potency. Importantly, structure-based modeling also contributed to understanding why certain antibodies retained activity against emerging variants while others rapidly lost efficacy. By visualizing how specific mutations altered local geometry or electrostatic interactions, these models provided actionable insights for therapeutic redesign and cocktail formulation [18, 22].

Nevertheless, significant uncertainties persist, particularly flexible regions, transient conformations, glycosylated surfaces, and multimeric assemblies. These limitations underscore the need for cautious interpretation of structural predictions and for integrating multiple modeling approaches with experimental validation. In translational settings, structure-based predictions must therefore be viewed as probabilistic guides rather than definitive answers.

Beyond screening existing biological sequences, generative AI models enable de novo protein design by learning to sample novel structures and sequences optimized for predefined objectives. These approaches allow for the simultaneous consideration of multi-objective properties, including binding affinity, stability, solubility, and manufacturability [10, 16, 25, 27]. During the COVID-19 response, generative methods were applied to design mini-proteins and optimized antibody scaffolds, demonstrating their potential to expand the functional space of therapeutics beyond evolutionary precedents [13, 15, 26].

A defining feature of the COVID-19 pandemic was the rapid emergence of viral variants with altered transmissibility and immune evasion properties. AI-driven models integrating sequence evolution, structural constraints, and mutational scanning data were widely used to assess how specific substitutions might impact antibody binding and vaccine efficacy. These approaches supported real-time evaluation of emerging variants and informed therapeutic updates [1, 2, 5, 26, 29]. Immune escape prediction represents one of the most challenging and consequential applications of AI-directed protein design. While short-term mutational effects can often be captured with reasonable accuracy, forecasting longer-term evolutionary trajectories under immune pressure is considerably more complex. Viral evolution is shaped by trade-offs between transmissibility, immune evasion, replication efficiency, and fitness costs, factors that are difficult to model comprehensively [2, 6, 16, 25, 26].

The COVID-19 experience highlighted the need for cautious interpretation of immune escape predictions and for continuous model updating as new data become available. Rather than serving as definitive forecasts, these models are best viewed as tools for scenario exploration and risk prioritization, informing preparedness strategies rather than dictating fixed design decisions.

Experience from COVID-19 highlighted that computational approaches have moved beyond supportive roles to become essential tools in pharmaceutical development. While experimental validation remains indispensable, AI-enabled modeling provides a structured framework for narrowing candidate spaces, integrating diverse datasets, and accelerating the discovery and optimization of both vaccines and antibody therapeutics under conditions requiring speed, adaptability, and scalability [2, 3, 11, 12, 26]. Thus, the success of AI driven computational approaches depends heavily on experimental validation, manufacturability considerations, and regulatory pathways, highlighting the importance of integration across the drug development pipeline.

While these applications illustrate the functional scope of AI-directed computational protein design, their performance is ultimately determined by the underlying machine learning paradigms used to represent, learn, and optimize protein sequence–structure–function relationships. Understanding these methodological foundations is essential for interpreting both the strengths and limitations of AI-enabled design pipelines.

Protein language models (PLMs) apply self-supervised learning to large-scale sequence datasets, treating amino acid sequences as contextual token streams to learn latent representations that encode evolutionary constraints and functional signatures without requiring explicit structural input [15, 25]. Tools such as ESM-2 and ProtT5 exemplify this approach, leveraging transformer-based architectures to predict mutation effects and functional annotations across hundreds of millions of sequences [16, 27]. This scalability enables rapid downstream applications, such as the variant fitness and immune escape predictions performed by EVE/EVEscape, making PLMs particularly valuable for early-stage discovery and tracking rapidly evolving pathogens [26].

Structure-driven approaches explicitly incorporate three-dimensional information, encoding spatial relationships such as backbone geometry and intermolecular interfaces using graph neural networks (GNNs) or geometric deep learning [7, 8]. AlphaFold3 and RoseTTAFold have established the benchmark for biomolecular complex prediction, enabling structure-guided therapeutic design with high fidelity [7, 8]. These structural insights are extended by specialized docking and scoring tools like DiffDock and GNINA, which utilize diffusion models and convolutional neural networks to predict protein-ligand binding poses [30, 31]. By enforcing physical and spatial constraints, these models achieve high predictive accuracy for tasks dominated by structural determinants, including antibody-antigen interface modeling as seen in platforms like DeepAb and IgFold [11, 32].

Building on these predictive foundations, generative AI frameworks now integrate heterogeneous data sources—including sequence, structure, and experimental mutational data—within unified architectures. Denoising diffusion probabilistic models, such as RFdiffusion and Chroma, allow for the de novo generation of functional protein backbones and binders tailored to specific viral epitopes [13, 28]. These are often coupled with sequence-design tools like ProteinMPNN or autoregressive models like ProGen2 to ensure fold-stability and biological function [15, 25].

Furthermore, the maturation of immune-informatics resources, such as the IEDB Analysis Resource and NetMHCpan, facilitates the identification of high-affinity epitopes for vaccine antigen design [23, 24]. While these multimodal models improve robustness, experience from the COVID-19 pandemic highlighted the need to validate computational predictions under real-world experimental constraints. Table 1 summarizes the key features, architecture, and contributions of these representative tools. It should be noted that these listings are not intended as a definitive catalogue, as numerous additional methods and resources exist beyond those included to address the complexities of protein engineering.

A pivotal shift in the pandemic response was the move from stabilizing viral proteins to the de novo design of synthetic binders. Researchers utilized Rosetta-based and early diffusion models to design small (approx. 55-residue) proteins targeting the SARS-CoV-2 Spike Protein Receptor Binding Domain (RBD) [33]. These designed miniproteins, such as LCB1, demonstrated picomolar neutralization potency, exceeding that of many monoclonal antibodies. Crucially, experimental validation through cryo-EM confirmed that the binders engaged the target exactly as predicted in the computational model [33, 34]. Their superior thermal stability and ability to be produced in E. coli provided a scalable alternative to traditional biologics.

AI played a critical role in the development of multivalent antigens, which provide broader protection than monomeric proteins. The design of the GBP510 (Ruvaxivid) vaccine utilized computational tools to create a self-assembling protein nanoparticle (icosahedral symmetry) decorated with 60 copies of the SARS-CoV-2 RBD [35, 36]. Clinical trial data showed that this AI-scaffolded design elicited significantly higher neutralizing antibody titters compared to the monomeric RBD or mRNA-based benchmarks in certain populations. This provides concrete evidence that computational geometry can directly enhance the human immune response.

As variants like Omicron emerged, AI was deployed to future-proof antibody therapeutics by identifying conserved epitopes. Studies employing Deep Mutational Scanning (DMS) data combined with Graph Neural Networks (GNNs) allowed for the rapid optimization of antibodies like Sotrovimab derivatives. By predicting escape trajectories, researchers were able to engineer antibodies that maintained high-affinity binding across multiple variants of concern [37–39]. Experimental assays confirmed that these AI-optimized candidates retained neutralization against the BA.2 and BA.5 subvariants when many traditional antibodies failed [40, 41].

Importantly, these case studies reveal that the success of AI-enabled approaches depends not only on algorithms but also on high-quality data availability and experimental throughput. Regions with limited sequencing or experimental capacity were less able to benefit from these advances, highlighting equity and infrastructure considerations that must be addressed for future pandemic preparedness.

A significant challenge in modern protein design is the escalating computational cost associated with training and deploying state-of-the-art models. The high-fidelity diffusion models and large-scale molecular dynamics (MD) simulations require massive GPU clusters, creating a compute divide that favors well-funded institutions. Further, the energy requirements for these iterative design cycles raise concerns regarding the environmental sustainability of always-on pandemic surveillance systems.

As AI models are increasingly trained on vast repositories of viral sequences and human immune repertoires, data privacy and biosecurity have emerged as paramount concerns. The use of patient-derived monoclonal antibody data requires rigorous anonymization and secure handling to prevent the exposure of sensitive clinical information. Further, there is an ongoing debate regarding the dual-use potential of generative AI, where the same tools used to design vaccines could theoretically be repurposed to enhance viral fitness or escape, necessitating robust international governance and red-teaming protocols.

The primary operational challenge remains the disparity between the speed of AI-generated designs and the throughput of wet-lab validation. Even with high-affinity predictions, the design-build-test cycle is often slowed by traditional synthesis and purification timelines, highlighting the need for more integrated, automated bio-foundries to achieve true pandemic readiness.

Addressing these integrated challenges, from high computational overhead to the ethical management of biological data, remains a prerequisite for the maturation of the field. Only by harmonizing rapid algorithmic innovation with sustainable infrastructure and robust biosecurity protocols can AI-directed design transition from an experimental urgency into a dependable pillar of public health. Ultimately, navigating these multifaceted limitations will define the trajectory of the next-generation of pandemic-ready therapeutics.