The review, which synthesizes literature published between 2020 and late 2025, emphasizes that AI-driven design does not replace experimental science; rather, it reframes it. By learning complex sequence–structure–function relationships from large-scale biological data, these models prioritize hypotheses, narrow the candidate space, and enable rational iteration under severe time constraints. During COVID-19, AI pipelines contributed to antibody discovery, antigen stabilization, variant impact assessment, and scaffold design, compressing early discovery timelines and enabling rapid iteration in response to viral evolution. The diversity of successful approaches, spanning sequence-based learning, structure-aware modeling, and generative optimization, highlighted both the flexibility of AI methods and the importance of aligning computational strategy with the specific biological question at hand.

Methodologically, the field rests on three complementary paradigms. Sequence-based protein language models, exemplified by tools such as ESM-2 and ProtT5, apply self-supervised learning to hundreds of millions of natural protein sequences, treating amino acid strings as contextual token streams to learn latent representations that encode evolutionary constraints on folding, stability, and function. During the pandemic, these models were widely applied to rapidly screen antibody repertoires, identify conserved viral regions, and prioritize targets less susceptible to immune escape, tasks that were particularly valuable in the earliest phases of the outbreak when structural information was scarce but viral genomic sequences were accumulating rapidly. Yet sequence-only models have inherent limitations: they may fail to capture three-dimensional interactions, conformational dynamics, post-translational modifications, and context-dependent effects that strongly influence binding and stability. Consequently, they work best as an initial filtering step guiding subsequent structure-aware analyses rather than as standalone decision tools.

Structure-informed machine learning explicitly incorporates three-dimensional information, encoding spatial relationships such as backbone geometry and intermolecular interfaces using graph neural networks or geometric deep learning. AlphaFold3 and RoseTTAFold have established the benchmark for biomolecular complex prediction, enabling structure-guided therapeutic design with high fidelity, while specialized docking tools like DiffDock and GNINA predict protein–ligand binding poses using diffusion models and convolutional neural networks. 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. 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. Significant uncertainties persist, however, particularly for flexible regions, transient conformations, glycosylated surfaces, and multimeric assemblies, underscoring that structural predictions must be viewed as probabilistic guides rather than definitive answers.

The most striking technical shift captured by the review is the transition from purely statistical sequence modeling to structure-first generative paradigms driven by denoising diffusion probabilistic models and flow-matching architectures. These frameworks, such as RFdiffusion and Chroma, treat protein design as the reversal of a physical corruption process: a model is trained to denoise a random cloud of points into a biologically plausible protein fold, explicitly constructing three-dimensional coordinates throughout the generative sampling process rather than treating proteins as one-dimensional token streams. This allows the AI to navigate the complex equivariant space of protein geometry, maintaining global symmetry and fold topology with atomic-level intent. A critical component is the in silico validation of hallucinated structures before they reach the wet lab: generated backbones are subjected to a self-consistency evaluation in which an independent, high-fidelity structure predictor such as AlphaFold2 or ESMFold forward-folds the AI-designed sequence, and a design is considered feasible only if the predicted structure matches the original generative backbone with high precision, typically below two angstroms in root-mean-square deviation. This iterative workflow, combining generative sampling with structural consistency oracles and biophysical filtering, is often coupled with sequence-design tools like ProteinMPNN or autoregressive models like ProGen2 to ensure fold stability and biological function.

The review highlights three case studies that moved computational workflows from theory to experimental and clinical reality. First, researchers used Rosetta-based and early diffusion models to design small proteins of roughly 55 residues targeting the SARS-CoV-2 spike receptor binding domain. These designed miniproteins, such as LCB1, demonstrated picomolar neutralization potency, exceeding that of many monoclonal antibodies, and cryo-EM validation confirmed that the binders engaged the target exactly as predicted in the computational model. Their superior thermal stability and ability to be produced in E. coli provided a scalable alternative to traditional biologics. Second, AI played a critical role in the development of multivalent antigens: the GBP510 (Ruvaxivid) vaccine used computational tools to create a self-assembling protein nanoparticle with icosahedral symmetry decorated with 60 copies of the SARS-CoV-2 receptor binding domain, and phase 3 clinical trial data showed that this AI-scaffolded design elicited significantly higher neutralizing antibody titers compared to monomeric RBD or mRNA-based benchmarks in certain populations, providing concrete evidence that computational geometry can directly enhance the human immune response. Third, as variants like Omicron emerged, AI was deployed to future-proof antibody therapeutics by identifying conserved epitopes: studies employing deep mutational scanning data combined with graph neural networks allowed rapid optimization of antibodies like sotrovimab derivatives, and experimental assays confirmed that AI-optimized candidates retained neutralization against the BA.2 and BA.5 subvariants when many traditional antibodies failed. Related work has shown that computational methods can restore the potency of clinical antibodies against Omicron and preemptively optimize them for broad neutralization and robustness against viral escape.

Immune escape prediction represents one of the most challenging and consequential applications of AI-directed protein design. Models integrating sequence evolution, structural constraints, and mutational scanning data, such as EVE/EVEscape, were widely used to assess how specific substitutions might impact antibody binding and vaccine efficacy, supporting real-time evaluation of emerging variants and informing therapeutic updates. Yet the review is candid about the limits: while short-term mutational effects can often be captured with reasonable accuracy, forecasting longer-term evolutionary trajectories under immune pressure is considerably more complex, because viral evolution is shaped by trade-offs between transmissibility, immune evasion, replication efficiency, and fitness costs that are difficult to model comprehensively. 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, and they require continuous updating as new data become available.

The pandemic also exposed significant translational and regulatory friction. Regulatory acceptance requires transparent, reproducible, and validated computational models, along with clear documentation of how predictions are generated and used; variability in model predictions during COVID-19 underscored the need for standardized benchmarking, reporting practices, and version control. Current regulatory frameworks were not designed with AI-driven discovery in mind, and while the pandemic prompted limited flexibility in review processes, long-term adoption of AI-designed biologics will require clearer guidance on model validation, lifecycle management, update policies, and post-approval monitoring. Hybrid workflows combining in silico screening with targeted wet-lab validation proved essential for successful translation, reducing attrition and improving confidence in AI-generated candidates while enabling early assessment of manufacturability, stability, and developability, factors that are critical for clinical success but often overlooked in purely computational studies. The review also flags a compute divide: high-fidelity diffusion models and large-scale molecular dynamics simulations require massive GPU clusters, favoring well-funded institutions, and the energy requirements of iterative design cycles raise environmental sustainability concerns for always-on pandemic surveillance systems.

Emerging concerns around data privacy and biosecurity add further complexity. As AI models are increasingly trained on vast repositories of viral sequences and human immune repertoires, the use of patient-derived monoclonal antibody data requires rigorous anonymization and secure handling to prevent exposure of sensitive clinical information. 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 immune escape, necessitating robust international governance and red-teaming protocols. Meanwhile, 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. Regions with limited sequencing or experimental capacity were less able to benefit from these advances during COVID-19, highlighting equity and infrastructure considerations that must be addressed for future preparedness.

Looking ahead, the review argues that preparing for future pandemics, including hypothetical Disease X scenarios, requires a shift from reactive therapeutic development to anticipatory design. Pre-emptive development of broad-spectrum vaccines and antibodies targeting conserved viral elements represents a central goal, offering the potential for baseline protection against entire viral families, while curated libraries of ready-to-deploy biologics could accelerate response timelines by enabling rapid selection and adaptation of pre-validated candidates. Integration of One Health data spanning human, animal, and environmental surveillance will be critical for identifying zoonotic risks before spillover occurs, and coupled with global data-sharing platforms enabling real-time model updating, these approaches could support coordinated and proactive responses to emerging threats. The authors conclude that the maturation of AI-enabled protein design depends on its integration as a supportive component of global pandemic preparedness, complementing established therapeutic development frameworks rather than replacing them. Only by harmonizing rapid algorithmic innovation with sustainable infrastructure, rigorous experimental validation, and transparent regulatory oversight can AI-directed design transition from an experimental urgency into a dependable pillar of public health, ensuring that the global scientific community is better equipped for future infectious disease challenges.

Subject of Research: Artificial intelligence directed computational protein design: lessons from COVID-19 for pandemic-ready vaccines and antibody therapeutics

Article Title: Artificial intelligence directed computational protein design: lessons from COVID-19 for pandemic-ready vaccines and antibody therapeutics

Article References: Artificial intelligence directed computational protein design: lessons from COVID-19 for pandemic-ready vaccines and antibody therapeutics. (n.d.). https://doi.org/10.3389/jpps.2026.16146

Image Credits: AI Generated

DOI: 10.3389/jpps.2026.16146

Keywords: artificial intelligence, protein design, COVID-19, pandemic preparedness, vaccines, antibody therapeutics, Artificial, intelligence, directed, computational, protein, design