What this project does, in one page
Dengue has four serotypes, and antibodies against the wrong part of the envelope (E) protein can make a later infection worse (antibody-dependent enhancement, ADE). We score every region of E for how likely it is to attract neutralizing antibodies, and we report every pre-registered test, including the ones that failed. The output is a shortlist of hypotheses for laboratory testing. Nothing here shows protection.
The pipeline E protein only
| Step | What happens | Where to look |
|---|---|---|
| 1. Sequences | NCBI dengue genomes, aligned across the four serotypes; conservation computed per residue. | Explorer |
| 2. Labels | Antibody-contact residues from E–antibody structures, grouped into epitope classes. | Tracks |
| 3. Model | Residue-level model scores each position of E (score S). | Tracks, structure colouring |
| 4. Tests | Held-out epitope classes, held-out antibody 4E11, a sealed set scored once. Criteria frozen first. | Evidence |
| 5. Ranking and design | 15-residue windows ranked, fusion loop penalised for ADE risk, HLA coverage for Bangladesh, mRNA designs. | Candidates, mRNA |
| Methods | Plain-words description of the model, data, tests, protocols and everything that failed. | Methods |
| 6. Lab testing | Not done. Needs neutralization assays (the follow-up proposed to ICDDR,B or BCSIR). | From computer to bench, Lab workbench |
Results at a glance read from the data file, not typed
| Question | Answer | Outcome |
|---|---|---|
| Loading computed data… | ||
For researchers how to use this, what is new, what is not
| Use case | How |
|---|---|
| Choose regions to test in the lab | Take the ranked 15-residue windows in the candidate table. Each lists the model score, conservation, ADE flag, supporting neutralization papers and Bangladesh class I HLA coverage. They are hypotheses for neutralization assays or mutagenesis. |
| Check a residue or region you already care about | Select a serotype and E in the explorer, then colour the 3D structure by conservation or by model score S and compare with the tracks. |
| Judge how far to trust a ranking | Read the evidence section: every pre-registered test, its criterion and whether it passed, including the failures. |
Explore a serotype
Pick a serotype in the top bar, then a protein. The genome map shows where each dengue protein sits on the virus genome. Colour shows how much each protein is the same across all four serotypes. We measured this by comparing thousands of public dengue genomes (5,463 passed quality control), not by assigning labels by hand.
Genome map NC_001474.2
Envelope (E)
VariableConservation, epitope and ADE tracks
One value per amino acid for the selected protein. For the envelope (E) protein the tracks also show the model score, a flag for the fusion loop (the stretch linked to antibody-dependent enhancement), and how many published neutralization studies mention each position. Every protein also shows peptides that tested positive in public immune assays. These are hypotheses from computation and published experiments, not evidence of protection.
Envelope (E) —
Ranked E candidate windows 15 aa · development-only
The five best-scoring, non-overlapping stretches of 15 amino acids for each serotype. Each is ranked by the model score, multiplied by how conserved the stretch is, and set to zero if it overlaps the fusion loop. "Class I cov." estimates what share of people in Dhaka carry an immune-system type (HLA) able to present the stretch to T cells; it is based on 141 people and is a minimum, so true coverage is likely higher. The sealed test set is used up, so this ranking cannot be re-tested, and nothing here has been tested in a lab.
| Serotype | Window | Sequence | Score S | Cons. K | Gold papers | Class I cov. | Binding panel alleles |
|---|
What was tested, what failed, and what the designs are
Every result below was tested against pass marks written down and locked before the test was run. Failures are shown next to successes. None of it is laboratory evidence of protection.
Validation of the E epitope model criteria fixed before scoring
| Test | Result | Pre-set criterion | Outcome |
|---|
Host transcriptomics: severity signature across cohorts gate not passed
| Train → test | AUROC [95% CI] | Gate |
|---|
HLA input: Bangladesh Dhaka Bangalee, n = 141 AFND, 2-digit groups
How common each immune-system type (HLA) is in a Dhaka Bangalee sample of 141 people. T cells see viral fragments only when an HLA molecule presents them, so a good vaccine stretch should be presentable by common local types. We used the HLA-A and HLA-B families only (the source data has no HLA-C, and no tool was run for the DRB1 family). Our predictor works at a finer level than this data, so each stretch gets a low and a high estimate. Types the predictor does not cover count as non-binding, so both estimates err on the low side. The candidate table above lists which predictor alleles bind each stretch.
mRNA design files design stage · not expressed or tested
Computer designs of mRNA that would make cells produce the chosen stretches. For each serotype there is a short construct (the top three stretches joined by short flexible linkers, 56 amino acids) and a full prM/E control, which stands in for the inserts used in licensed vaccines. Each is designed two ways: the most common human codon for every amino acid, or LinearDesign (a research-only tool that balances codon choice against RNA stability). GC is the share of G and C bases, CAI measures how closely codon use matches human genes, and MFE per nucleotide is predicted RNA folding energy (more negative means more stable). These are computer numbers only: a more stable RNA does not always make more protein. The designs have no signal peptide, UTRs or poly-A tail, so they are starting points, not finished vaccine sequences.
| Design | Method | aa | GC | CAI | MFE / nt | Sequence |
|---|
How a wet lab would use this, and how it would be proven
This page is a hypothesis generator. Nothing here has been tested in a laboratory, and a computer cannot show that a region protects people. This section says what a lab can take from the page today, what is not built yet, and the ladder of experiments that would turn a prediction into evidence.
What exists today, and what does not deployment status
| Available now | Not available |
|---|---|
|
|
The ladder of experiments each rung must pass before the next is worth doing
| Step | Experiment | Prediction being tested | What would count against the model | Needs |
|---|---|---|---|---|
| 1. Binding | Make recombinant E (ideally the E dimer, because many broad antibodies need two E proteins together) with point or alanine mutations at top-ranked positions and at matched low-ranked positions. Measure binding of known neutralizing antibodies and of convalescent serum by ELISA. | Mutating top-ranked positions removes more antibody binding than mutating matched low-ranked positions. | No difference between top and matched low positions. | Protein expression, ELISA. No live virus. |
| 2. Neutralization | Put the same mutations into reporter virus particles or virus, and run a plaque or focus reduction neutralization test (PRNT50 or FRNT50) with antibodies or sera. | Mutating top-ranked positions lowers neutralization titre more than the matched controls. | Binding changes but neutralization does not, or the reverse. | Virus work under the lab's biosafety rules; an in-house neutralization assay. |
| 3. Enhancement (ADE) | Test the same antibodies or sera in an enhancement assay with Fc-receptor-bearing cells (for example K562), and include fusion-loop and non-fusion-loop epitopes. | Antibodies to the regions we flag do or do not enhance infection. This also tests the fusion-loop rule, which is only a prior today. | Antibodies to the flagged loop neutralize without enhancing (the rule over-penalises), or antibodies to unflagged regions enhance (the rule misses risk). | Cell-based enhancement assay. |
| 4. Immunogenicity | Immunise mice with an mRNA construct built from the top windows, against a prM/E control construct. Measure neutralizing titres against all four serotypes and enhancement. | The construct raises neutralizing antibodies that cover more serotypes than the control without more enhancement. | Low or serotype-restricted titres, or more enhancement than the control. | mRNA synthesis, animal facility and ethics approval. |
| 5. Protection | Challenge studies in animals and, much later, human trials. | Protection from disease. | Not reachable from this project. | Beyond the scope of this work. |
How to run it so a negative result counts proposal, not yet agreed with any lab
- Write the pass mark first. Before any experiment, record how many positions will be tested, which are the matched controls, what size of difference counts as a hit, and how the data will be analysed. We would then freeze that note the same way the computer tests were frozen.
- Blind the readout where possible: whoever measures should not know which constructs are top-ranked and which are controls.
- Report everything, including constructs that did not express or antibodies that did not work.
- What success would mean. If the top-ranked positions matter more than matched controls in steps 1 and 2, the ranking carries information beyond exposure. That supports the method. It still would not show protection.
- What failure would mean. If they do not, the ranking should be dropped or reworked, and the page should say so.
Plan an experiment from the model, in four steps
This is where a lab turns the model's scores into a concrete list of mutants to make, writes down the pass mark before starting, and checks the results afterwards. It runs in your browser. Nothing you paste or type is uploaded anywhere.
Loading the model data…
1Choose the sequence which virus are you working with?
The model needs the protein sequence of the dengue envelope (E) protein, about 495 amino acids, written as letters (for example MRCIGISNRD…). If you sequenced the virus you will have DNA or RNA: translate it to protein first. If you only want to use one of the four standard reference viruses, pick it below.
What do the scores mean, and what can they not do?
The score (0 to 1) says how much a position looks like the positions antibodies were seen touching in 37 solved structures. It ranks positions; it is not a probability and has no cut-off. For your own sequence the tool finds the closest reference serotype and copies its exposure and neighbourhood features, so a single amino-acid change barely moves the score, and if you lost a sugar-attachment site the scores near it are unreliable. The tool warns you when that happens.
2Design the mutant panel which positions to mutate, and which controls to use
You will mutate each chosen position to alanine (to glycine if it is already alanine) and test whether antibody binding drops. The key idea is a matched pair: one high-scoring position and one low-scoring control that is just as exposed on the surface. Without the control, a drop in binding could simply mean "any exposed residue matters", which the model does not claim.
Exposure is how much of a position's surface can be touched from outside the virus particle (0 = buried, 1 = fully exposed). Buried positions are skipped, because mutating them can break the protein's folding.
Advanced settings
3Write the pass mark first, then record results do this before any experiment
Decide now what result would count as success, and sign it, so the answer cannot quietly change after you see the data. This is the same discipline the computer tests followed. You get two files: a pass-mark note to sign and keep, and a results sheet already filled with your constructs.
What to measure and how to fill the sheet
- Make the wild-type protein and every mutant in the panel (the same way, the same day if you can).
- Measure binding (or your chosen readout) of the same antibody or serum to each. Run the wild type on the same plate.
- In the sheet, write yes or no in expressed. If a mutant did not express, say so; do not leave it out silently.
- Enter signal_mutant and signal_wildtype (or fold_change_vs_wt directly). Fold change = mutant ÷ wild type: 1 means no change, 0.3 means 70% of binding was lost.
- If you can, have someone else label the tubes so the person measuring does not know which are test and which are controls.
4Check what you found paste the filled results sheet
Load the results sheet you filled in. The tool compares each high-scoring mutant with its own matched control, applies the pass mark from step 3, and says plainly whether the result is supported, not supported, or inconclusive. It never changes the pass mark for you.
How the model works and how we tested it
Written for biologists. You do not need a machine-learning background: each term is explained, and every number on this page comes from a result file in the project, not typed by hand.
The idea what the model is asked to do
Antibodies recognise small patches on the surface of the dengue envelope (E) protein. We collected 3D structures of E with antibodies attached, and marked every E amino acid that touches an antibody (within 4.5 Å). The model learns which kinds of positions tend to be touched, and then scores every position in E. It sees only the properties of each position (listed below), never the antibody itself.
| What the model sees at each E position (15 features) | Why a microbiologist would care |
|---|---|
| How exposed the position is on the virus particle (and on a lone E protein) | An antibody can only bind what it can reach. The virion surface hides a lot of E. |
| How much exposure changes between lone E and the assembled virus | Shows parts buried when E dimers pack together. |
| Exposure and water-loving character of the surrounding patch (within 10 Å), and how many neighbours it has | Antibodies bind patches, not single residues. |
| Distance to the nearest N-glycosylation site, and whether the position is itself one | Sugars on E (at N67 and N153) shape where antibodies can land. This was the strongest signal the model used. |
| Residue chemistry: hydrophobicity and charge (the residue and a 7-residue window) | Binding surfaces have typical chemistry. |
| Conservation across serotypes, variability within the serotype, and whether the position is identical in all four serotypes | Included to test the common idea that conserved means a good target. It did not help: antibody-touched positions were, if anything, less conserved. |
The training data
Structures are grouped into epitope classes: complexes whose antibodies touch largely the same set of E positions go in the same class (overlap of 50% or more). This grouping matters for testing, below.
| Class | Structures (PDB) | Serotype | Example | Positions touched | Model | Exposure alone | Conservation alone |
|---|
How we tested it four tests, one rule: write the pass mark first
| Test | In plain words | What it can and cannot tell you |
|---|---|---|
| E1: leave one class out | Hide every structure of one epitope class, train on the rest, and ask whether the model still ranks the hidden class's positions above the rest of E. Repeated for each class. | Closest to the real question: can it recognise a kind of epitope it has never seen? Only 13 classes, so intervals are wide. |
| E2 and E3: neutralization papers | Compare scores with residues that published mutagenesis studies found to matter for neutralization. E3 used 14 papers kept sealed and scored exactly once. | The structures and these papers describe many of the same well-studied antibodies, so this partly measures agreement between two label sources. The sealed set is now spent. |
| 4E11 hold-out | Remove all four structures of one well-known domain III antibody (4E11), and every position it touches, from training. | A clean test on one antibody; it does not generalise to every antibody. |
| Fresh complexes (v6) | Six more E-antibody structures that were never used for training, tuning or earlier tests. They were found by a written rule before any model was run. | The only evidence that goes beyond data we had already examined. Only three of the groups are truly new epitope classes, so the uncertainty is large. |
Reading the numbers. AUROC asks: pick one position that antibodies touch and one that they do not; how often does the model score the touched one higher? 0.5 is a coin flip, 1.0 is perfect. 95% interval is the range the true value plausibly lies in; a wide range means few data. Top-20% recovery: of the truly important positions, what share lands in the model's top fifth? By chance, 0.20.
Why you can trust the process pre-registration and fingerprints
Pre-registration means we wrote down the success criteria before running the test, then locked the document with a SHA-256 fingerprint (a string that changes if even one character of the file changes). Anything we changed later is logged as a dated deviation, never edited in place. That is why the failures are shown here.
| Protocol | Frozen | What it set up | Outcome | SHA-256 (first 16) |
|---|
The "fusion loop" rule a prior, not a finding
Some antibodies make a later infection worse instead of protecting (antibody-dependent enhancement, ADE). Antibodies against the fusion loop (E residues 98 to 110) are the classic example, so the ranking subtracts points from windows that overlap it. This is a fixed rule from published biology; the model did not learn it, and we have no data on which E positions cause enhancement. Recent structures show antibodies that touch the fusion loop in a larger, two-protein (dimer) site and still neutralize all four serotypes, so the rule may wrongly demote some of the best regions. Treat it as a caution flag, not a verdict.
The fresh complexes in detail v6
| PDB | Description | Positions touched | New epitope class? |
|---|
Ideas that did not work every variant we tried is listed
Show the table
| Idea | What it was | Held-out-class score (E1) [95% interval] | Outcome |
|---|
Glossary
| Serotype | One of the four dengue virus types (DENV-1 to DENV-4). |
| Epitope | The patch of a viral protein that an antibody binds. |
| Neutralizing antibody | An antibody that stops the virus from infecting cells. |
| ADE | Antibody-dependent enhancement: antibodies that bind but do not neutralize can help the virus enter immune cells and worsen a later infection. |
| Fusion loop | A short stretch of E (residues 98 to 110) that the virus uses to fuse with the cell membrane; antibodies to it are linked to ADE. |
| Conservation | How similar a position is across serotypes. High means the same amino acid in all four. |
| Held-out | Data the model never saw while learning. |
| Pre-registered | Success criteria written and locked before the test is run. |
| mRNA construct | A designed RNA sequence that would make cells produce the chosen E region; here a computer design only, not made or tested. |
Vaccine & antiviral targets
What each dengue countermeasure actually binds, mapped back to the proteins above. Live-attenuated vaccines target the surface proteins prM/E, which are only moderately conserved across serotypes; the newer antivirals target replication proteins with a higher share of identical columns.
| Candidate | Mechanism | Protein target | Status (as of Oct 2026) |
|---|
What actually drives severe dengue
Sequence differences between serotypes are part of the picture, but severity is not predicted by genetic distance alone. These factors matter at least as much, and mostly act independently of which serotype is involved.
Sources
Genome coordinates and sequences are fetched from each serotype's own NCBI RefSeq record, not estimated. Structures are real, deposited PDB entries — where no structure has been solved for a serotype (most of NS4B; NS5 for DENV-1/DENV-4), the viewer says so rather than substituting a stand-in.