Epitope Conservancy Analysis: A Step-by-Step Guide
Epitope conservancy analysis measures the fraction of protein sequences in a set that contain your epitope at or above a sequence identity threshold you choose. You run it on the IEDB Epitope Conservancy Analysis tool once you have assembled a non-redundant, strain-balanced set of the antigen. The sequence set and the threshold decide the answer, not the tool.
Every other step in an immunoinformatics vaccine design workflow quietly assumes one genome. You picked a target, predicted B-cell epitopes and T-cell epitopes, screened them, and assembled a construct. Conservancy is the step that asks whether any of that survives contact with the next strain. It is also the step most often reported as a single number with no explanation of how that number was produced, which is why reviewers push back on it.
What does an epitope conservancy analysis actually measure?
The definition comes from the paper that introduced the tool, Bui and colleagues in BMC Bioinformatics in 2007 (PMID 17897458, open access at PMC2233646). Two terms do the work, and students routinely swap them:
“Identity refers to the extent to which two amino acid sequences are invariant, and is measured as the percentage of identical amino acids in the alignment of two sequences. Conservancy is defined as the fraction of protein sequences that contain the epitope considered at or above a specified level of identity.”
So identity is a per-sequence measurement and conservancy is a count across the set. An epitope that matches at 100% identity in 40 of 50 sequences has a conservancy of 80% at a 100% threshold. Change the threshold to 80% identity and the same epitope, in the same set, may qualify in 48 of those 50, a conservancy of 96%. Nothing about the biology moved. Only the question you asked moved.
One mechanical detail matters more than it looks. The tool performs a best local alignment of the epitope against each protein in the set, so you do not pre-align anything, and a hit that needs a gap is not a hit. The paper is direct about why:
“But, to be relevant in an immunological context, it is crucial that the entire epitope sequence is completely aligned with absolutely no gaps.”
That single sentence is the reason a generic alignment viewer is the wrong instrument here. If you want the mechanics of alignment itself, we cover those separately in what sequence alignment is and how it works. This guide is about the decisions wrapped around the tool.
Why does the sequence set decide the answer before the tool runs?
The tool takes two inputs, and one of them you build yourself. The authors flag the sequence set as the part that determines whether the output means anything:
“Based on our experience, to achieve the best results it is recommended that the protein sequence set utilized be constructed such that redundancies are eliminated and the representation of different substrains and serotypes is balanced.”
Here is how badly that can go wrong with real data. On 24 September 2026, NCBI Virus held 12,707 nucleotide records for Monkeypox virus (taxid 10244), and 12,387 of them had been released on or after 1 January 2022. More than 97% of everything available for that species entered the database during and after one outbreak. You can repeat both counts and get today’s figures:
curl -s 'https://api.ncbi.nlm.nih.gov/datasets/v2alpha/virus/taxon/10244/dataset_report?page_size=1'
curl -s 'https://api.ncbi.nlm.nih.gov/datasets/v2alpha/virus/taxon/10244/dataset_report?page_size=1&filter.released_since=2022-01-01'Read total_count in each response. A student who downloads everything and pastes it into Step 2 has not measured conservancy across Monkeypox virus. They have measured conservancy across the 2022 outbreak clade, with a thin historical tail, and their headline figure will collapse the moment a reviewer asks about older lineages. The number is not wrong. The label on it is.
A defensible set is built in four moves:
- Choose the antigen, then the sequences of that antigen. Pull protein sequences for the specific gene product you predicted epitopes on, not whole proteomes. NCBI Virus and NCBI Datasets both let you filter by taxon and gene before downloading.
- Drop partial records. A partial CDS that stops before your epitope region counts as a sequence without the epitope and silently pushes conservancy down. Filter on completeness, or check record lengths against the reference protein length before you paste.
- Remove exact and near-exact duplicates. Five hundred identical sequences from one sequencing effort are one observation repeated, and they will drown out a rare lineage that matters clinically. The tool offers a checkbox on the input form to strip duplicated sequences automatically, which handles the exact matches. It cannot do the next move for you.
- Balance the lineages you intend to claim. If your paper will say the construct covers three clades, the set needs all three represented, not one clade at 90%. Public phylogenies such as Nextstrain are a quick way to see which lineages exist before you decide what to include.
Write down what went into the set and in what proportion. That paragraph is what turns a conservancy percentage into a claim someone can check, and it takes two sentences in a methods section.
Linear or discontinuous: which calculation option should you pick?
Step 3 of the live form offers exactly two calculation options, checked on the tool page on 24 September 2026: “Epitope linear sequence conservancy” and “Epitope discontinuous sequence conservancy”, plus a sequence identity threshold field.
| Option | What it aligns | Use it when | Stated limit |
|---|---|---|---|
| Epitope linear sequence conservancy | The continuous epitope sequence, aligned whole against each protein, with no gaps allowed | Your epitopes came from a linear predictor: MHC class I and II binders, linear B-cell predictions, anything you will string into a construct | Says nothing about conformation. A conserved linear stretch can still sit in a differently folded region |
| Epitope discontinuous sequence conservancy | The pattern of residues at their positions, for epitopes defined as scattered residues on a folded surface | Your epitope came from a structure-based B-cell predictor and is defined as a residue set rather than a peptide | Pattern conservation is not cross-reactivity, in the authors’ own words |
The caveat on the second row is worth quoting in full, because it is the honest limit of the method and citing it makes your methods section stronger rather than weaker:
“The epitope conservancy analysis tool developed here can be used to assess the pattern conservation of discontinuous epitopes. Nevertheless, pattern-wise conserved discontinuous epitopes may not be cross-reactive due to the unknown influence of neighboring and inter-dispersed amino acids.”
The same paper points at the better instrument when you have one: “As a result, if antigen structures are available, it may be better to predict cross-reactivity based on the epitope’s 3D structural conservation.” If you already have a validated model from the structure prediction and validation step, say so and compare the surfaces rather than leaning on pattern conservancy alone.
What sequence identity threshold should you set?
There is no universal answer, and any page that gives you one without naming its source is guessing. The tool exposes the threshold as a field precisely because it is a per-project judgement: the paper defines conservancy in terms of “a specified level of identity” and does not prescribe the level.
What you can do is show the reader how much the threshold moves the result. Here is the arithmetic on a deliberately small, made-up set, so the numbers are illustrative and reproducible rather than a measurement from any real pathogen. Take a 9-mer epitope and a set of 10 sequences of the same antigen:
- In six sequences, all nine residues match. Identity 9/9, which is 100%.
- In three sequences, one position differs. Identity 8/9, which is 88.9%.
- In one sequence, three positions differ. Identity 6/9, which is 66.7%.
At a 100% identity threshold, conservancy is 6/10, which is 60%. At an 80% threshold, the three single-substitution sequences now qualify and conservancy is 9/10, which is 90%. At a 60% threshold it is 10/10. One epitope, one set, three answers. This is why a reported conservancy figure without its threshold is unreadable.
Two anchors help you choose rather than guess:
- Anchor on the claim you want to make. If the claim is that a single peptide will be recognised identically in every listed strain, the threshold is 100% and you should expect fewer epitopes to survive. If the claim is that the region is stable enough to be worth targeting, a lower threshold is defensible as long as you state it and justify it.
- Anchor on where the substitutions fall. A substitution at an MHC anchor position is not equivalent to one at a solvent-facing position, and a percentage cannot tell them apart. Go back to your MHC binding predictions and rerun the variant peptide rather than deciding from identity alone.
Our own group made this choice explicitly in a published multi-epitope design. In the monkeypox study (Viruses, 2022, PMC9693848, summarised on our research page), conservation was run across three strains, MPXV-WRAIR7-61, Sierra Leone and COP-58, and the paper states that “epitopes exhibiting 100% identity in sequences among the selected monkeypox strains were selected since they possessed less probability of any mutation”. That is a strict threshold, stated with its reason and with the strain list attached. Copy the habit, not the number.
Want the guided, hands-on version?
Our live Molecular Modeling & MD Simulations cohort bootcamp takes you from zero to running real docking and MD workflows, with a portfolio project for your grad-school applications.
How do you run the analysis, step by step?
The IEDB Epitope Conservancy Analysis tool is a three-step form, and the run itself is trivial once the inputs are ready.
- Step 1, Epitope Sequence(s). Paste your epitopes, one per line or as FASTA. These are the peptides that survived your antigenicity, allergenicity and toxicity screening, not the raw predictor output. Running conservancy on 300 unscreened predictions wastes the step.
- Step 2, Protein Sequence(s). Paste the curated set you built above, in FASTA, with headers you can still read in the output. Keep strain names in the headers. When a single sequence is the reason your conservancy dropped, you want to know which one it was without going back to the download.
- Step 3, Calculation option(s). Pick linear or discontinuous, set the sequence identity threshold, and submit. The program reports, per epitope, the fraction of protein sequences matching above or below the identity level you gave, and also calculates the minimum and maximum matching identity level for each epitope. A position mapping of each epitope onto the matching protein fragments sits behind the “View details” column.
Read the minimum identity column, not only the headline fraction. Two epitopes can share the same conservancy figure and mean different things: if the lowest identity observed is one residue short of a full match, the epitope varies slightly everywhere, while a much lower minimum means the epitope is effectively absent from a subset of your strains. The original paper also documents a link to browse sequences at NCBI from within the form and a button to download the results to a file, which is the version you should keep for your methods section.
The tool is a Java web application, and the paper lists its requirements as Java 1.4 or higher and Tomcat 4.0 or higher for anyone who needs to run it locally, with no licence restrictions for non-academic users.
What do you do with an epitope that is immunogenic but poorly conserved?
This is the decision the percentage exists to inform, and there are four honest options rather than one rule:
- Drop it. The default when the epitope is one of several candidates with similar predicted binding and the others are conserved. Cheap, and it costs you nothing you cannot replace.
- Keep it and add its variant. If the epitope is strongly predicted and the variation is a single common substitution, include both peptides in the construct. That is a design decision with a cost in construct length, so weigh it against the linker and ordering constraints we cover in the construct design guide.
- Keep it and narrow the claim. If it is conserved across the clade you are actually targeting, say that in the abstract instead of implying species-wide coverage. A narrower claim that holds is worth more than a broad one that does not.
- Re-check whether the substitution matters functionally. Run the variant peptide through the binding predictor before you discard a good candidate over a conservative substitution outside the anchor positions.
How is conservancy different from population coverage and antigenicity screening?
These three steps all produce a percentage and students mix them up constantly, usually by reporting one and calling it another. They ask different questions about different populations.
| Analysis | The question it answers | What the input set is | What a failing result means |
|---|---|---|---|
| Epitope conservancy | Is this epitope present across the strains of the pathogen? | Protein sequences of the antigen from many strains | The epitope is variable, so the construct may miss some lineages |
| Population coverage | Do the HLA alleles that bind these epitopes occur in the human populations you are targeting? | Your epitope and allele pairs, against allele frequency data | The construct may not be presented in some human populations |
| Antigenicity, allergenicity and toxicity screening | Is this peptide suitable as an immunogen at all? | The individual peptide sequences | The peptide is dropped from the candidate list |
The pairing to hold onto: conservancy is about the pathogen population and population coverage is about the host population. A construct needs both, and reporting one as though it covered the other is the single most common misreading of this step.
What is the inverse test, and why does a match mean rejection?
Conservancy asks where your epitope appears and wants the answer to be “everywhere in the pathogen”. There is a second, opposite question, and almost nothing written for students covers it: does your epitope also appear in the host proteome? Here a match is a reason to reject the epitope, because similarity to a human protein raises the risk of an autoimmune cross-reaction.
The test is a protein-protein BLAST of your epitopes, or the assembled construct, against Homo sapiens, taxid 9606, at NCBI BLAST. Restrict the organism field to taxid 9606, use blastp, and read the alignments rather than only the E-values: a short peptide against a large proteome produces weak statistics by construction, so a low-scoring but exact stretch deserves your attention more than the E-value suggests.
This is the check our monkeypox paper reports for the finished construct, where the outcome was the one you want: the vaccine construct “is dissimilar to human proteins (taxid: 9606; Homo sapiens)” in a protein-protein BLAST analysis. Run it on the construct, and run it again after any codon-level or linker-level redesign that changes the peptide sequence.
What goes wrong, and how do you fix it?
| What you see | Cause | Fix |
|---|---|---|
| Every epitope returns 100% conservancy across hundreds of sequences | The set is near-identical, usually a bulk download dominated by one outbreak or one sequencing effort | Deduplicate, then rebalance across lineages. Re-run and report both the set composition and the result |
| Conservancy is far lower than expected and the minimum identity is very low | Partial or truncated records in the set that never contained the epitope region | Filter to complete sequences of the antigen, or check each record’s length against the reference protein |
| An epitope returns no match anywhere | Inputs swapped between Step 1 and Step 2, or the protein set is the wrong gene product | Confirm the epitope is in Step 1 and the proteins in Step 2, then confirm the set is the antigen the epitope was predicted on |
| A hit looks right in an aligner but the tool does not count it | The match requires a gap, and the method aligns the entire epitope with no gaps | Treat it as a non-match. If the gapped region matters biologically, that is a separate structural question |
| Discontinuous output is hard to interpret | The option reports pattern conservation, which is not the same as conformational cross-reactivity | Report it as pattern conservation, and use structural comparison where antigen structures exist |
| Results cannot be reproduced months later | The sequence set was never saved, and public databases grow | Archive the exact FASTA set with its download date, and state the date in the methods |
Where does this step sit in the workflow?
Conservancy belongs immediately after epitope prediction and screening, and before you commit to a construct. Running it late is expensive: every downstream step, from construct assembly to structure modelling to docking, has to be redone if a core epitope turns out to be strain-specific. It also feeds back into target selection: an antigen whose surface-exposed regions are all variable may be the wrong antigen, and that is better learned at the start than after the construct has been docked and simulated.
If you are building the underlying skills rather than working on a specific project, the computational biology skills roadmap sets out the order we recommend, and the immunoinformatics pillar guide walks the full nine-step workflow this spoke belongs to.
Frequently asked questions
What does conservancy mean in an epitope conservancy analysis?
It is “the fraction of protein sequences that contain the epitope considered at or above a specified level of identity”, in the definition given by Bui and colleagues in 2007. It is a count across your sequence set, while identity is the per-sequence percentage of matching residues.
Is there a standard sequence identity threshold I should use?
No. The tool exposes the threshold as a field because the choice depends on your claim and your sequence set, and the original paper prescribes no universal value. Pick one, state it beside every conservancy figure you report, and say why you chose it.
Do I need to align my sequences before running the tool?
No. The method performs a best local alignment of the epitope against each protein in your set. You do need the set itself to be curated, and a match that requires a gap does not count, because the whole epitope must align with no gaps to be immunologically meaningful.
Can I run conservancy on discontinuous epitopes?
Yes, using the discontinuous calculation option, but read the output as pattern conservation rather than proof of cross-reactivity. The authors note that pattern-wise conserved discontinuous epitopes may still not be cross-reactive, and recommend structure-based comparison where antigen structures are available.
Is conservancy the same thing as population coverage?
No. Conservancy is about the pathogen population, asking whether the epitope is present across strains. Population coverage is about the human population, asking whether the HLA alleles that present your epitopes are common in the groups you are targeting. A vaccine design needs both, reported separately.
Written by the StemSkills Lab team, who have spent more than ten years in sequence and structural bioinformatics, drug discovery and design, and multiscale molecular modeling, and whose published multi-epitope vaccine designs are listed on our research page.
Want the guided, hands-on version?
Our live Molecular Modeling & MD Simulations cohort bootcamp takes you from zero to running real docking and MD workflows, with a portfolio project for your grad-school applications.