mRNA, RNA, and Vaccine Design
RNA and vaccine design combine sequence, structure, immunology, delivery, manufacturing, and population biology. AI helps with parts of that stack, not the whole stack at once. AlphaFold 3 supports antigen and complex modelling; Evo, Nucleotide Transformer, and Orthrus support RNA sequence and representation work; AlphaGenome supports regulatory-sequence analysis; T-cell epitope predictors support immunogenicity triage. Program success still depends on delivery, dosing, safety, and clinical evidence.
Use this chapter to:
- Support antigen selection, RNA construct design, immune-response prediction, delivery constraints, and vaccine-program evidence.
- Antigen design, RNA stability, codon choice, innate immune activation, LNP delivery, manufacturing, and immunogenicity are separate problems.
Prerequisites: Protein Structure Prediction for the AF3 lineage; Nucleic Acid and Genome Models for RNA and genome modelling.
Summary: Support antigen selection, RNA construct design, immune-response prediction, delivery constraints, and vaccine-program evidence. AI helps with structure, sequence, and epitope hypotheses, but vaccine success still depends on immune biology, delivery, and clinical evidence.
Key point: Antigen design, RNA stability, codon choice, innate immune activation, LNP delivery, manufacturing, and immunogenicity are separate problems. Open question: whether sequence and structure predictions translate into delivery, immune response, durability, and protection.
Bottom line: Vaccine AI connects genomics, protein structure, immunology, RNA biology, delivery science, manufacturing, clinical trials, and public health.
What is this field trying to solve? Support antigen selection, RNA construct design, immune-response prediction, delivery constraints, and vaccine-program evidence.
What is the core idea? Antigen design, RNA stability, codon choice, innate immune activation, LNP delivery, manufacturing, and immunogenicity are separate problems.
What is the current state of the field? AI helps with structure, sequence, and epitope hypotheses, but vaccine success still depends on immune biology, delivery, and clinical evidence.
What do we know, and what remains open? Known reference points include AlphaFold 3, Evo, Evo 2, Nucleotide Transformer, Orthrus, NetMHC tools, IEDB, viral sequence databases, RNA design tools, and vaccine trial evidence. What remains open is whether sequence and structure predictions translate into delivery, immune response, durability, and protection.
Why does this matter? Vaccine AI connects genomics, protein structure, immunology, RNA biology, delivery science, manufacturing, clinical trials, and public health.
Introduction
Genome and protein models matter for vaccine and RNA design because they support antigen analysis, protein design, variant interpretation, and sequence optimisation. AlphaFold 3 addresses biomolecular interaction prediction (Abramson et al., 2024). Evo and AlphaGenome illustrate the move toward long-context sequence models (Nguyen et al., 2024; Avsec et al., 2026). Orthrus adds a peer-reviewed RNA foundation model trained for mature RNA representation and property prediction (Fradkin et al., 2026). Industrial deployments (Moderna ChatGPT Enterprise, OpenAI April 2024) demonstrate broad AI adoption in vaccine programs.
mRNA vaccine design is not only antigen choice. The mRNA construct, nucleoside modifications, untranslated regions, cap, poly(A) tail, purification, and delivery system determine expression and immunogenicity. Karikó and colleagues showed that modified nucleosides can suppress innate immune recognition of RNA (Karikó et al., 2005), and Pardi and colleagues reviewed the delivery and formulation constraints that made mRNA vaccines a platform rather than a single design problem (Pardi et al., 2018).
mRNA construct anatomy
An mRNA construct is a designed system. The coding sequence determines antigen protein sequence, but the untranslated regions tune expression, the cap and poly(A) tail affect translation and stability, nucleoside chemistry affects innate immune sensing, codon choice and RNA structure affect expression, purification affects tolerability, and the delivery vehicle determines which cells receive the transcript. AI work that optimizes one component should be read as component-level evidence.
LinearDesign is a strong peer-reviewed example of algorithmic mRNA design tied to measured vaccine-relevant outcomes. Zhang and colleagues used a lattice parsing approach to optimize mRNA sequence design and reported improvements in stability and immunogenicity in the tested settings (Zhang et al., 2023). The correct reading is not “AI designs vaccines.” It is that sequence-level optimization can improve a construct when the objective and assay are specified.
Epitope prediction and immune escape
T-cell epitope prediction helps rank candidate peptides for HLA presentation, but population HLA diversity, antigen processing, immune history, and disease context limit transfer. NetMHCpan and NetMHCIIpan are mature triage tools, not direct measures of vaccine protection (Reynisson et al., 2020).
For rapidly evolving viruses, language models can help read evolutionary constraints and escape pathways. Hie and colleagues showed that language models trained on viral sequence variation capture patterns relevant to viral escape (Hie et al., 2021). This evidence supports surveillance and antigen-prioritization workflows. It does not replace neutralization assays, cellular immunity assays, animal studies, or effectiveness studies.
Delivery is the hard constraint
Delivery determines whether an RNA design reaches the right tissue, persists long enough, avoids unacceptable inflammation, and produces antigen at a useful level. Lipid nanoparticle formulation, dose, route, repeat dosing, storage, and population tolerability can dominate the biology of a candidate sequence. That is why RNA-design claims should name the delivery system and the measured expression and immune-response data, not only the sequence objective.
What is demonstrated?
Demonstrated capability includes protein antigen modelling, epitope-informed design support, and sequence-to-function modelling for selected genomic and molecular outputs. AlphaFold 3 demonstrated interaction modelling relevant to antigen-antibody and protein-nucleic acid questions (Abramson et al., 2024). Evo demonstrated sequence modelling across DNA, RNA, and proteins (Nguyen et al., 2024), and Orthrus demonstrated RNA-specific representation learning in Nature Methods (Fradkin et al., 2026). Verified industrial deployments (Moderna and OpenAI) show that vaccine companies have integrated general-purpose AI into operational workflows.
Sequence-design evidence is strongest when it is tied to measured expression or antigen presentation. Massively parallel 5’ UTR design showed that regulatory sequence can tune translation (Sample et al., 2019). Work on mRNA structure, stability, and translation shows why optimization is multi-objective rather than a simple codon-frequency problem (Leppek et al., 2022). For cellular immunity, NetMHCpan and NetMHCIIpan remain mature tools for antigen-presentation triage, with HLA coverage and training-data limits that must be stated (Reynisson et al., 2020).
| Evidence Anchor | What It Supports | Practical Constraint |
|---|---|---|
| AlphaFold 3 | Biomolecular interaction structure prediction | Immune response is not settled by structure alone |
| Evo / Evo 2 / Nucleotide Transformer / Orthrus | Long-context biological sequence and RNA representation modelling | RNA therapeutic behaviour requires delivery and assay data |
| Modified nucleoside literature | Innate immune sensing and RNA tolerability design | Immunogenicity is context dependent |
| UTR and structure optimisation | Expression and stability tuning | Must be validated in the intended cell type and delivery system |
| NetMHCpan / NetMHCIIpan | MHC presentation triage | HLA diversity and antigen processing limit transfer |
| FDA drug AI materials | Regulatory attention to AI in drug and biologic development | Context of use drives evidentiary needs |
| Moderna ChatGPT Enterprise | Verified industrial deployment | Operational use; not direct evidence of vaccine efficacy |
What is theoretical?
Theoretical capability includes integrated vaccine design systems that jointly model antigen structure, immune escape, expression, delivery, and population-level strain coverage. Such systems require data across immunology, manufacturing, and clinical outcomes. Production-grade integration is partial; component pipelines exist at major vaccine developers.
What is beyond current capability?
Beyond current capabilities includes reliable vaccine design from sequence surveillance alone. Immunogenicity, durability, safety, delivery, and real-world effectiveness require experiments and trials. AI-only design of an approved vaccine, without the full preclinical and clinical cascade, remains beyond current capabilities.
What would make this more promising?
Vaccine and RNA design become more promising when the model output is tied to expression, delivery, immune response, and protection evidence.
| Claim | Evidence that raises or lowers confidence |
|---|---|
| “This antigen or epitope is promising” | Binding, antigen presentation, neutralization, cellular immunity, or immune-escape data support the intended immune response |
| “This RNA construct is improved” | Expression, stability, purity, innate immune sensing, and formulation data are measured in the relevant delivery system |
| “This escape prediction should guide design” | Surveillance, variant fitness, neutralization panels, HLA coverage, and antigenic distance are evaluated together |
| “This vaccine candidate is ready to advance” | Formulation, dose, animal data, safety, manufacturing, and clinical evidence plan are specified |
| “Operational AI improved the program” | The model changed a named design, manufacturing, trial, or regulatory decision with measurable quality or speed impact |
The claim should name whether the evidence supports construct design, antigen prioritization, formulation, manufacturing, immunogenicity, or clinical protection.
What should researchers, biotech teams, funders, and program leaders do with this?
- Keep antigen modelling separate from immune-response prediction.
- Validate expression, stability, and formulation before immunogenicity claims.
- Use neutralisation, cellular immunity, and safety assays as separate evidence layers.
- Avoid fixed timeline claims unless a specific program supplies verified dates.
- For RNA sequence work, cite the specific tool (Evo, Orthrus, RNA-FM, classical RNAfold) and its scope.
- Treat company AI-adoption announcements (Moderna ChatGPT Enterprise) as operational facts, not efficacy claims.