Disulfide Engineering for Vaccine Constructs (DbD2)
Skip to content

Disulfide Engineering in a Vaccine Construct with DbD2

Disulfide Engineering in a Vaccine Construct with DbD2

Disulfide engineering adds a covalent cross-link to a validated multi-epitope vaccine model to reduce its flexibility before docking. Disulfide by Design 2 scans the backbone for residue pairs whose geometry could host a disulfide, then reports a torsion angle, a bond energy and a B-factor sum for each. You pick candidates, mutate both residues to cysteine, and re-validate.

This guide picks up exactly where modelling, refining and validating the construct’s 3D structure stops, and it runs before you hand anything to docking against TLR4. The whole series sits under the immunoinformatics pillar guide, and the broader skill order is in the computational biology skills roadmap.

What does disulfide engineering actually add to a vaccine construct?

A multi-epitope construct is a chimera. Epitopes from different proteins are strung together with short linkers and an adjuvant, so the assembled chain has no evolutionary history of folding as one unit. The validated model you get out of a predictor is usually confident at the domains it borrowed and floppy across everything you invented. A disulfide bond is the one post-modelling change that can reduce that flexibility without altering a single epitope sequence, because it cross-links two positions that were already close in space.

Be clear about what it does not do. A predicted disulfide is a hypothesis about geometry, not a measured increase in stability, and not an immunogenicity gain. The StemSkills Lab team’s own two published multi-epitope constructs went from a validated model straight to docking and molecular dynamics without a disulfide engineering step, which is the honest baseline most competing tutorials hide: this step is optional, and plenty of published work omits it.

So the useful question is not how to run the tool. It is when running it changes anything.

When does a disulfide bond earn its place in your construct?

Three signals decide it, and the source paper is explicit that bond energy alone is not enough. Craig and Dombkowski added B-factor analysis to DbD2 precisely so users could find disulfides likely both to form and to improve thermal stability, and they report prior experimental work in which stabilising mutations clustered at longer loop lengths, between 25 and 75 residues bridged by the bond, with few stabilising mutations reported below 25 residues.

SignalWhat it tells youGuideline from the sourceWhen it is unreliable
Bond energy (kcal/mol)How far the modelled disulfide sits from ideal geometryThe paper reports that “90% of native disulfides have an energy value less than 2.2 kcal/mol” and offers that as a working guidelineA low energy says the bond is geometrically plausible, not that it will form in expression
Loop length (residues between the pair)How much unfolded-state entropy the cross-link removesStabilising mutations were associated with loops of roughly 25 to 75 residues; few were reported under 25Pairs inside one rigid domain satisfy the geometry but have little to stabilise
Sum B-factorHow mobile the two residues are in the deposited structureStabilising disulfides predominantly spanned regions of relatively high mobilityMeaningless unless the PDB carries real temperature factors (see the pLDDT warning below)

Read the table as an AND, not an OR. A candidate with a 0.4 kcal/mol energy that bridges eleven residues inside a stable helix is a bond you can build and should probably skip. One more reason not to look for a single score: version 2.11 of DbD2 removed its own composite rank calculation after it was shown to be misleading, so the tool deliberately does not rank candidates for you.

How do you run Disulfide by Design 2 on a vaccine construct, step by step?

The server is at cptweb.cpt.wayne.edu/DbD2/ and it is hosted by the Center for Protein Therapeutics group at Wayne State University School of Medicine. One practical warning before you click: the site serves over plain HTTP only, so your browser will mark it as not secure. That is the same situation as several other academic immunoinformatics servers, and it is not a sign the tool is broken. There is nothing confidential in a PDB of your own model, but do not type credentials into any page that loads this way.

  1. Load the structure. Three routes: upload a local PDB, type a PDB identifier (the help page’s worked example is 7RSA) and fetch it from RCSB, or click Load Example PDB. Local uploads are capped at 2 MB, which a single construct never approaches.
  2. Pick the model if you have several. A Model drop-down offers 1 to 5 and is hidden for single-model files. For an NMR-style multi-model file the first model is used unless you change it, and for residues with alternate conformers only the first coordinate set is read.
  3. Leave the three option boxes checked. Intra-chain, Inter-chain and Build C-beta for Gly are all checked by default. The third one matters more than it looks: glycine has no C-beta, and DbD2 works from backbone and C-beta atoms, so unchecking it makes every glycine position invisible to the scan. Linker regions in multi-epitope constructs are often glycine-rich.
  4. Keep the default tolerances for the first pass. Two drop-downs control how far from ideal a candidate may sit: the chi3 angle tolerance, offered from 0 to 45 degrees in steps of 5 and set to 30 by default, around the two minima at +97 and -87 degrees; and the C-alpha-C-beta-S-gamma angle tolerance, offered from 0 to 12 degrees in steps of 2 and set to 10 by default, around 114.6 degrees.
  5. Run it, then read the tabs. Results arrive across four tabs: FILE INFO, ANALYSIS, 2-degree STRUCTURE and 3-D VIEWS. The ANALYSIS tab is the candidate list.
  6. Save what you will need later. Save Results writes the table as CSV, which is what belongs in your methods section. Create/View Mutant builds the cysteine pair in the viewer, and Save Mutant PDB writes a new file with SSBOND records added plus a REMARK line above every residue mutated to cysteine. For a multi-model PDB it saves only the modified model.

Reset restores every default, which is worth knowing once you start narrowing tolerances.

PDB 7RSA: ribonuclease a (1 chain, A) with 2-METHYL-2-PROPANOL (ligand code TBU) and 188 crystallographic water molecules, at 1.26 Å resolution.
PDB 7RSA: ribonuclease a (1 chain, A) with 2-METHYL-2-PROPANOL (ligand code TBU) and 188 crystallographic water molecules, at 1.26 Å resolution. Source: RCSB PDB entry 7RSA.

How do you read the results table?

Each candidate pair reports two kinds of information: who the residues are, and what the bond between them would look like.

ColumnMeaningHow to use it
Chain, Seq #, AAChain identifier, residue number and amino acid for each half of the pairCheck both positions against your construct map before you accept them
StructSecondary structure read from the PDB, printed as a sheet strand with its sense, or a helix with its typePrefer pairs that bridge different elements over pairs buried inside one helix
chi3Estimated torsion angle of the modelled bondSanity check: it should sit near +97 or -87 degrees
kcal/molCalculated bond energy from the DbD2 energy functionApply the 2.2 kcal/mol guideline as a filter, not as a ranking
Sum B-factorSum of the PDB temperature factors of the two residues, over backbone and C-beta atomsOnly meaningful on a structure with real temperature factors

A construct position that lands in a linker is usually the interesting one, because the linker is the part with no folding history. A position inside a borrowed epitope needs a second look: mutating an epitope residue to cysteine changes the epitope.

Why is the Sum B-factor column misleading on a predicted model?

This is the mistake almost every student makes with this tool on a vaccine construct, and no competing page warns about it.

Sum B-factor is useful because experimental temperature factors track mobility, and the stabilising disulfides in the literature spanned mobile regions. But your construct has no crystal structure. It has a predicted model, and AlphaFold-style predictors write their per-residue confidence score, pLDDT, into the B-factor field of the output PDB. pLDDT is a confidence value, so a high number means the predictor was sure, which is close to the opposite of flexible. Read that column naively on a predicted model and you will select for the regions the predictor was most confident about, which on a chimera are the borrowed domains, which is exactly where a cross-link has the least to do.

Two honest ways forward. Either ignore the Sum B-factor column on a predicted model and select on energy, loop length and secondary structure alone, saying so in your methods; or get a mobility estimate the right way, from a short molecular dynamics run, and use per-residue fluctuation instead. Whichever you choose, check what your own predictor wrote into the B-factor field before you trust a single number in that column.

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.

Join the waitlist (free) →

What geometry and energy does DbD2 actually model?

Knowing the model matters, because it tells you what the numbers can and cannot mean. DbD2 builds a disulfide with fixed bond lengths, 1.81 angstrom for C-beta to S-gamma and 2.04 angstrom for S-gamma to S-gamma, and a fixed C-beta-S-gamma-S-gamma bond angle of 104.15 degrees, consistent with the geometry survey of Petersen and colleagues from 1999. The chi3 torsion is defined by C-beta-S-gamma-S-gamma-C-beta, rotating about the S-gamma-S-gamma bond, while chi1 is defined by N-C-alpha-C-beta-S-gamma.

The algorithm rotates chi3 until the modelled C-beta to C-beta distance matches the measured distance between your two candidate residues, examines the possible S-gamma orientations and keeps the lowest-energy coordinates. The energy function the help page publishes is a sum of five terms:

  • Eij = E(chi1,i) + E(chi1,j) + E(chi3) + E(theta,i) + E(theta,j)
  • E(chi1) = 1.4[1 + cos(3 chi1)]
  • E(chi3) = 4.0[1 – cos(1.957[chi3 + 87.0])]
  • E(theta) = 55.0[theta – theta0]2, with theta the C-alpha-C-beta-S-gamma angle and theta0 = 114.6 degrees

Energies are in kcal/mol, the minima fall at chi3 of +97 and -87 degrees and chi1 of plus or minus 60 and 180 degrees, and there is no bond-stretching term because the bond lengths are held fixed. The observed C-alpha-C-beta-S-gamma angle distribution peaks near 115 degrees over a range of roughly 105 to 125 degrees, which is where the default tolerance of plus or minus 10 degrees comes from.

One trap to avoid when you write this up. The 2013 paper and the help page quote different survey sets, and averaging them is fabrication. The paper analysed 1,505 native disulfide bonds in 331 non-homologous proteins, finding chi3 peaks at -87 and +97 degrees and an energy distribution with a mean of 1.0 kcal/mol and a 90th percentile of 2.2 kcal/mol. The help page’s technical details describe a sample of 331 proteins containing 1,418 known disulfide bonds, with a mean energy of 0.89 kcal/mol and a maximum of 8.35 kcal/mol. The 1,418 figure is the paper’s count of bonds whose chi3 chirality was predicted correctly, not a second survey. Cite each number to the document you read it in.

How accurate is DbD2, and what does that number actually mean?

The paper reports a blind test on those 331 non-homologous PDB structures, selected for having at least one disulfide bond, less than 50% sequence identity and a resolution of 2.0 angstrom or better. Side-chain identities and locations of the native disulfides were hidden, which is possible because the algorithm reads only backbone and C-beta atoms. With default settings, of the 1,505 native disulfide locations DbD2 correctly predicted 1,479 (98%) as appropriate for disulfide formation, and of those, chirality of the chi3 torsion angle was correct in 1,418 (96%), with a reported R2 of 0.995 between predicted and true chi3 angles.

Now read that carefully, because the framing is where tutorials go wrong. Those percentages are recall of disulfides that already exist in crystal structures. They say the geometric model recognises real disulfide sites. They are not a measured success rate for engineered bonds forming in a recombinant protein, and the paper makes no claim that a designed bond will stabilise anything. If your dissertation says “98% accurate” without that sentence attached, a reviewer will ask, and they will be right to.

What do you do with the mutant PDB next?

The Save Mutant PDB output is an input to three things, not an endpoint.

  1. Sanity-check the two substitutions. You have replaced two residues with cysteine. Score that change the way you would score any point mutation, using the stability predictors covered in our guide to predicting the effect of a mutation on protein stability. DbD2 proposes a pair on geometry; it does not tell you what losing the original side chains costs.
  2. Re-validate the structure. Run the mutant back through the same validation you used on the original model, as described in the structure prediction and validation guide, and re-check the physicochemical profile with ProtParam, since two cysteines change the predicted properties of the sequence.
  3. Test whether the bond changed anything. The honest test is a comparison, and that is a standard workflow: see wild-type versus mutant MD simulation in GROMACS. If the cross-linked construct shows no reduction in fluctuation across the bridged loop, the bond bought you nothing, and reporting that is a result.

Only then does the construct go forward to receptor docking and the complex MD run. If you are still assembling the construct itself, start at designing the multi-epitope construct and come back here.

What goes wrong, and how do you fix it?

Every item here is a documented behaviour of the tool, not a guess.

SymptomCauseFix
Your PDB will not load at allThe help page’s Known Issues names this directly: some files fail because of “non-standard” residue numbering, and DbD2 expects numbering that starts at 1Renumber the model from 1 and reload. This is the top failure for a refined or re-chained construct, since refinement servers often return odd numbering
The page refuses to open, or the browser blocks itThe server has no HTTPS service, so an https address fails outrightUse the http address. Expect a not-secure warning and accept it for this host only
Upload rejectedLocal files are capped at 2 MBStrip waters, duplicate chains and extra models, or fetch by PDB ID instead
No candidate appears anywhere near your linkersBuild C-beta for Gly was unchecked, so glycine positions were skippedRe-run with the box checked. Glycine-rich linkers are common in these constructs
Far too many candidates to judgeThe default plus or minus 30 degree chi3 tolerance is deliberately permissiveThe help page’s own advice is to decrease the tolerance for a shorter list with preferential characteristics. Tighten chi3 first, then the C-alpha-C-beta-S-gamma tolerance
Results do not match the model you meant to analyseFor multi-model files the first model is used unless the Model drop-down is changed, and only the first conformer is read for alternate locationsSet the Model drop-down explicitly, or split the file before upload
The interface behaves oddlyInternet Explorer and Opera are flagged as not recommended by the tool’s own documentationUse Chrome, Firefox or Safari

Frequently asked questions

Is disulfide engineering a required step in a multi-epitope vaccine study?

No. It is optional, and plenty of published constructs go from a validated model to docking and MD without it. Treat it as a step you justify, using the energy, loop length and mobility criteria above, rather than a box to tick.

How many disulfide bonds should I add?

Neither the tool nor its paper prescribes a number, so any page that gives you one is guessing. Add the smallest number you can justify, test each against the original construct, and keep only the ones that measurably reduce fluctuation in a comparison run.

Can I use DbD2 on an AlphaFold model?

Yes for the geometry, with one caveat: the B-factor field of a predicted model usually holds pLDDT rather than experimental temperature factors, so the Sum B-factor column is not a mobility measure. Select on energy, loop length and secondary structure, and say in your methods which you used.

What is a reasonable energy cut-off for a candidate bond?

The paper’s own guideline is that 90% of native disulfides fall below 2.2 kcal/mol, which makes that a defensible filter. It is a guideline from a survey of native bonds, not a validated threshold for engineered ones, so cite it as such.

Does mutating two residues to cysteine change my epitopes?

It can, and this is the one check nobody should skip. If either position falls inside a predicted epitope rather than a linker or an adjuvant region, the substitution changes that epitope, and the epitope-level predictions need re-running before the construct is defensible.

Which version of the tool is this, and does it matter?

The interface reports version 2.13, whose release note records support for PDB files containing nucleic acid chains (displayed in the 3-D view, ignored for other processing). Version 2.12 moved the 3-D view to JSmol, so no Java plug-in is needed. Record the version you used in your methods.

Where this fits in the workflow

Disulfide engineering is a judgement step between validation and docking, not a stage every project needs. The reason to learn it now is that the decision rule generalises: a low-energy, geometrically plausible modification that bridges a mobile loop is worth testing, and anything you cannot measure against the original construct is not worth reporting. The StemSkills Lab team works on structural bioinformatics, drug design and multiscale molecular modelling, and this series follows the order a real project runs in, step by step, under the immunoinformatics pillar guide.

Primary sources for this guide: Craig DB and Dombkowski AA, Disulfide by Design 2.0: a web-based tool for disulfide engineering in proteins, BMC Bioinformatics 2013;14:346 (PMID 24289175), and the Disulfide by Design 2 help and technical details pages. The earlier method is described in Dombkowski AA, Disulfide by Design, Bioinformatics 2003;19(14):1852-3 (PMID 14512360), and the biology is reviewed in Dombkowski AA, Sultana KZ and Craig DB, Protein disulfide engineering, FEBS Letters 2014;588(2):206-12 (PMID 24291258).

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.

Join the waitlist (free) →

Get StemSkills certified, free
Take a free assessment and earn a verifiable certificate you can download and add to your LinkedIn profile.
Browse free certifications

Keep going

Dihedral Angle Analysis in GROMACS: gmx angle and gmx chi Track ligand torsions with gmx angle and side-chain rotamer flips with gmx chi, build the 4-atom index group… GROMACS MD With Two Ligands: Topology, Index, Restraints Set up a GROMACS run with two bound ligands: merge topologies, fix duplicate atomtypes, build index groups and… GROMACS Position Restraints: gmx genrestr and POSRES Restrain your ligand during NVT and NPT with gmx genrestr, place the #ifdef POSRES block correctly, and fix…
See live workshops