Traditional vaccine development starts by growing the pathogen. Reverse vaccinology starts with its genome. Instead of culturing an organism and testing attenuated or inactivated forms, you scan its proteome computationally, identify the fragments the immune system is most likely to recognise, and assemble them into a single designed construct before anything reaches a bench.
The approach was first used at scale against Neisseria meningitidis serogroup B, where conventional methods had failed for decades, and produced a licensed vaccine. It is now standard practice for difficult pathogens. This article walks through the pipeline and is honest about what each step can and cannot tell you.
Step 1: Choose the right proteins
A bacterial proteome may contain thousands of proteins. Most are useless as vaccine targets because the immune system never encounters them. The filter is straightforward:
- Subcellular localisation. Surface-exposed and secreted proteins are accessible to antibodies. Cytoplasmic proteins generally are not. Tools such as PSORTb and CELLO handle this classification.
- Antigenicity. Probabilistic tools such as VaxiJen score whether a protein is likely to provoke an immune response.
- Conservation. A target that varies between strains produces a vaccine that works against some isolates and not others. Align across all available genomes and keep what is conserved.
- Absence of human homologues. This is the safety-critical filter. A target resembling a human protein risks autoimmunity, so anything with meaningful similarity to the human proteome is discarded.
Step 2: Predict the epitopes
The immune system does not recognise whole proteins — it recognises short fragments. Three classes matter, and they work differently.
T-cell epitopes (MHC class I and II)
These are peptides presented on the cell surface by MHC molecules. Prediction here is genuinely good: tools such as NetMHCpan and the IEDB suite are trained on large experimental datasets and MHC class I binding prediction is reasonably reliable.
The complication is population coverage. HLA genes are among the most polymorphic in the human genome, and a peptide binding one allele may not bind another. A vaccine designed against alleles common in one population may perform poorly elsewhere. Serious designs check predicted coverage across the HLA distributions of the intended target populations.
B-cell epitopes
These are the regions antibodies bind. Prediction is markedly less reliable, because most real B-cell epitopes are conformational — assembled from residues distant in sequence but adjacent in the folded structure. Linear predictors miss these by construction. Structure-based tools such as ElliPro and DiscoTope perform better when you have a good structure, but this remains the weakest link in the pipeline and should be described as such.
Step 3: Assemble the construct
Selected epitopes are joined into a single multi-epitope construct. The details matter:
- Linkers separate epitopes so they are processed and presented correctly. GPGPG is common between helper T-cell epitopes, AAY between class I epitopes, and EAAAK to connect functional domains rigidly.
- An adjuvant is usually fused at the N-terminus — often a TLR agonist such as beta-defensin or a bacterial flagellin fragment — because a peptide construct alone is typically poorly immunogenic.
- A purification tag such as a histidine tag is added for downstream expression.
Step 4: Check the construct is actually usable
A designed sequence can be immunologically attractive and physically impractical. Standard checks:
- Physicochemical properties (ProtParam): molecular weight, theoretical pI, instability index, aliphatic index and GRAVY. A high instability index or strong hydrophobicity predicts expression and solubility problems.
- Allergenicity (AllerTOP, AlgPred): a vaccine that triggers allergy is unusable.
- Toxicity (ToxinPred) at the peptide level.
- Solubility on overexpression, since insoluble constructs land in inclusion bodies and complicate production.
Step 5: Model the structure and validate it
Because the construct is a novel sequence, no experimental structure exists. Secondary structure is predicted (PSIPRED, SOPMA), then a three-dimensional model is built — increasingly with AlphaFold, which handles designed sequences better than older threading methods, though confidence scores across linker regions deserve scepticism.
The model is refined (GalaxyRefine) and validated with a Ramachandran plot and quality scores such as ERRAT and ProSA-web before anything is built on top of it.
Step 6: Test the immune interaction
The construct is docked against the relevant immune receptor — commonly TLR4 for a flagellin-adjuvanted design, or specific MHC alleles for epitope presentation — using a protein–protein docking method such as HADDOCK or ClusPro.
Docking alone is not sufficient. The complex is then run through molecular dynamics simulation to test whether the predicted interface holds up in explicit solvent. RMSD, radius of gyration and hydrogen bond occupancy at the interface tell you whether the interaction is durable or an artefact of the docking scoring function. Many attractive docked complexes dissolve within the first few nanoseconds.
Immune simulation tools such as C-ImmSim are frequently used to model the predicted antibody and cytokine response over repeated doses. These are coarse models and should be read as illustrative rather than predictive.
Step 7: Prepare for expression
Finally the protein sequence is reverse-translated and codon-optimised for the expression host, usually E. coli K12. The Codon Adaptation Index and GC content are checked, restriction sites are removed or added as needed, and the sequence is cloned in silico into an expression vector such as pET-28a(+).
What this pipeline does and does not deliver
It delivers a well-reasoned, testable candidate quickly and cheaply, with the obviously bad options already eliminated. For pathogens that resist conventional approaches, it is often the most practical starting point available.
It does not deliver a vaccine. Every step rests on predictions with real error rates, and B-cell epitope prediction in particular is unreliable. Immune simulation output is not immunogenicity data. The genuine value is triage: narrowing an intractable search space to a small number of candidates worth expressing, purifying and testing in animals. Work presented as though the computational result were the endpoint is the main reason this field attracts scepticism.
