Peptide Design

The design target is a 23-mer that has to clear two receptors and a secretion pore

BoPEP4 is a plant elicitor peptide from Brassica oleracea, mature sequence GILIGSKKRPREPHSSGKPGGHN. Plants release peptides of this family when they are wounded or stressed; the receptor kinases PEPR1 and PEPR2 read them, recruit the co-receptor BAK1, and switch on defence and stress-tolerance responses [1, 3]. ReLEAF asks Bacillus subtilis to secrete one of these peptides into the root zone, so the molecule has to satisfy two judges at once: the receptor at the far end, and the secretion machinery it has to leave the cell through.

Those two judges disagree. The receptor wants the C-terminal two thirds intact and untouched. The secretion pathway wants a short, uncharged mature N-terminus at the signal-peptidase junction, and it does not care what happens downstream. Nearly every design decision on this page comes out of that disagreement, and the design space is small enough to enumerate: twenty-three positions, a handful of substitutions worth making, and seventeen possible N-terminal truncations.

The modelling question, stated so it can be answered wrongly

Two questions run through everything below, and keeping them apart is the single most useful thing this project did.

• Question A, structural. Does a given edit prevent the peptide from making the interface the crystal shows? This has a defined answer at the level of named, mutagenesis-validated contacts.
• Question B, affinity. Which variant binds most tightly? Our pipeline cannot answer this, and section 7 is the experiment that proves it cannot.

Every claim on this page is a Question A claim. Where the project made Question B claims, they have been withdrawn, and they are listed as withdrawn in section 8.

Three published results carry more weight than anything we computed

The modelling is anchored to experiments other people did, and it is worth being precise about which parts of the story we own and which parts we borrowed.

BoPEP4 is not a designed sequence in the de novo sense. It is a natural Brassica peptide chosen because it aligns onto AtPep1 with every activity-critical position conserved, and the design work is the decision about what to change from there.

The family varies everywhere except four positions

Seventeen mature Pep sequences, seven from Arabidopsis and nine from Brassica species alongside BoPEP4, aligned over twenty-four columns. The N-terminal third is close to unreadable as a family: at most six of seventeen sequences agree on any residue in columns 1-6. The C-terminal third is the opposite. Ser15, Ser16, Gly17 and Gly20 are the same residue in all seventeen.

Two things in that grid mattered to the design. The first is that the invariant block sits exactly where the crystal puts the interface, which is the subject of the next section. The second is Asn23: it is the last residue in twelve of the seventeen, and the others end in His, Ser, Asp or Thr. So the family does not conserve the asparagine side chain as strictly as it conserves Gly17, and Pearce measured the same thing directly, since replacing Asn23 with alanine is tolerated while removing it costs about 400-fold [2]. What is conserved is the free carboxylate at the end of the chain, not the side chain hanging off it. That distinction decides section 9.

Residues 7-23 make the interface and 1-6 touch nothing

BoPEP4 was docked onto the PEPR1 ectodomain taken from the 5GR8 crystal, in HADDOCK 2.5, with restraints transferred from the crystallographic interface [4]. Before trusting the result, the same protocol was run on AtPep1 itself: it returned the crystal pose at 0.77 Å ligand RMSD, top-ranked, with a 35-unit score gap to the first non-native pose. The PEPR2 interface has no crystal to stand on and was predicted twice, by AlphaFold3 co-folding and by docking, with the disagreements between them reported rather than averaged.

Below is the interface, residue by residue, as buried surface area on PEPR1 with the receptor partner and the PEPR2 status attached. Pick a residue to read it, and the matching column of the alignment above lights up with it.

The shape of that strip is the whole design brief. Nothing is buried until Lys7. From there the peptide runs a belt of basic residues along a row of receptor carboxylates: Lys7 to Glu199, Arg9 to Asp179, Arg11 to Asp294, Lys18 to Glu324 and Asp348, His22 to Glu439. It finishes with Asn23 buried in a pocket at 170 Å², more than any other residue, held by three receptor side chains at once.

PEPR2 reads the same footprint with two fewer grips

PEPR2 is 64% identical to PEPR1 and its groove superimposes to within about 2 Å at every anchor position. It keeps the C-terminal clamp and the Lys7, Arg9, Ser15, Ser16 and His22 partners. It loses the acidic partner of Arg11 and the basic partner of Glu12, and position 18 faces a hydrophobic patch instead of a carboxylate. A receptor that keeps the anchors and loses the middle of the belt is a plausible structural account of why PEPR2 binds two members of the family where PEPR1 binds six [3]. It is an account, not a measurement: there is no PEPR2 structure, and everything in this paragraph inherits the uncertainty of an AlphaFold model.

The interface map splits the peptide into a protected core and two editable ends

Reading the strip as a design rule gives the constraint every later decision was checked against.

The rule is stronger than a docking score because none of it depends on one: it is burial, plus family conservation, plus published mutagenesis. All three agree on the same block of the sequence.

The truncation ladder has a cliff, and the two methods put it five residues apart

If residues 1-6 do nothing at the receptor, how much further can the peptide be cut? Cutting is worth something: a shorter payload is cheaper to synthesise, and moving the mature N-terminus changes the charge the signal peptidase sees. So the whole ladder was built rather than one guess: every N-terminal truncation from 1-23 down to 17-23, docked against both receptors and co-folded in AlphaFold3. Seventeen fragments, two receptors, thirty-four complexes.

Read on the docking score, the curve is a ramp with no edge in it. Read on AlphaFold3 confidence, it is flat from 1-23 all the way to 13-23 and then falls off between 13-23 and 14-23, where interface confidence drops from 0.87 to 0.54 and the interface error jumps from 2.5 to 8.9 Å. The two methods are answering different questions: one is asking how much interface there is, the other is asking whether it can place the peptide at all.

Switch the axis to score per buried Ų and the cliff disappears

The third setting on that chart is the one that mattered. Across the seventeen rungs, the docking score correlates with peptide length at r = −0.96 and with buried surface area at r = −0.98; a regression on length and net charge alone explains 93% of its variance. Divide the score by the area it buries and the whole series lies between −72 and −87 with no step at any position, Arg9 included. The shortest fragment in the set, a documented non-binder, has the most favourable score per residue of any rung.

The verdict column in the source table was assigned by hard score cut-offs. Given that score runs at about −7.4 per residue, the cut-off at −160 is a length threshold at fifteen residues, and fifteen residues is 9-23. The cliff was reported at Arg9 because the threshold was set where Arg9 is.

None of this is surprising once it is looked for. Docking scores have been benchmarked against measured affinities on forty-six complexes and carry no useful correlation with them [8], and interface energy terms scale with interface size, which is why contact counts predict affinity where raw energies do not [9]. The project had been reading a size measurement as a strength measurement.

The docking failed the one benchmark that had measured activity attached

Everything above is a Question A claim about geometry. The project also wanted to rank designs, which is Question B, so it ran the test that decides whether it can. Pearce 2008 measured half-maximal activity for an alanine scan across the 9-23 scaffold [2]. Sixteen of those analogues sit on the same 9-23 scaffold, so length and buried surface are essentially constant and the substitution is the only large variable. Docking all sixteen against measured activity is as clean a test of a scoring function as this project could build.

The three analogues Pearce measured as catastrophic rank 7th, 8th and 9th out of sixteen. G17A, which is more than 1,600-fold less active than any analogue the docking ranked in its top six, scores as a better binder than all of them. Spearman ρ is +0.05 with p > 0.8. The correlation the source report originally quoted, +0.237, included the wild-type reference point in its own correlation; removing that one self-referential point takes it to +0.05, and it is not significant either way.

The co-folding arm fails the same control from the other direction. [A17]Pep1 is the Gly17→Ala analogue Pearce measured at more than 4,000-fold down, and it was included in the AlphaFold3 panel as a designed negative. It scored interface confidence 0.73 against wild-type BoPEP4's 0.74. This is the documented behaviour of AlphaFold-family models on single missense substitutions [11], and our own control reproduced it.

An audit of our own pipeline retracted five claims, one of them after the DNA had shipped

The audit started as a complaint. A team member pointed out that the docking had been run without an inspectable statement of its restraints and search space. That objection is partly a category error: HADDOCK has no search box to report, because its search space is the list of residues declared active at the interface [4]. It is partly correct, because that list had never been written down anywhere a reader could find it. Reconstructing it from the run trees turned up four problems, and the one that mattered was not the missing documentation.

The later campaigns start from the pose they are meant to be testing and derive their restraints by measuring which residues touch in that same pose. A run built that way cannot reject its own input. It is a legitimate operation, restrained refinement, and it was described in the downstream reports as independent physics-based confirmation of the AlphaFold3 models, which it is not. The runs also produced a single cluster containing every structure, so clustering carried no information, and no run was repeated at a second random seed, so no reported score difference had an error bar on it.

Two further findings came out of the project's own data with no new computation. The founding campaign's mutant table shows every point mutant scoring better than wild type, with the two mutants the literature calls most damaging scoring best of all. And the wild-type run's top-ranked cluster is a non-native pose at 5.80 Å; the complex carried forward as canonical is cluster three of four, chosen because it matches the crystal. That choice is defensible and it is now stated: the canonical wild-type pose is homology-guided, not blindly predicted.

The ledger

Every claim the docking programme produced, with what it now rests on. Five were withdrawn. The one that hurt was K18R, which had already gone into a DNA order and was pulled out of it on the morning of 16 August.

K18R is worth one more sentence, because it was withdrawn twice for different reasons. The audit killed the docking-score argument in July. What survived into August was a structural argument from the literature: Tang reports AtPep1's Arg18 both salt-bridging Asp348 and packing against Phe371, and BoPEP4 has a lysine there, which can make the salt bridge but not the stack [1]. Then the August panel measured the distances. BoPEP4's own Lys18 sits 2.62 Å from Asp348 and 4.17 Å from Phe371, against 2.50 and 4.09 Å in the crystal, and in three of four constructs it approaches the aromatic ring more closely than a re-docked arginine does. A lysine amine over an aromatic ring is a cation-π contact. The gap the substitution existed to close was not visible in our own models, so it lost its last rationale and its cloning slot.

A C-terminal tag deletes the bond the peptide binds with

Every BoPEP4 construct the project ordered before August fused six histidines directly onto Asn23, with no linker. The reason this is fatal needs no docking at all, and stating it without one is the point: peptide-bond geometry puts the backbone nitrogen of any residue 24 essentially where Asn23's terminal oxygen sits, which in the crystal is 2.40 Å from Arg487 NH1 and 2.80 Å from NE, one half of a bidentate salt bridge [1]. Substituting an amide nitrogen there does not weaken that bond, it removes it.

The literature has already run the experiment. AtPep2, AtPep4 and AtPep5 all carry natural C-terminal extensions past the conserved asparagine. Two of the three fail to bind PEPR1 at all, the third binds and cannot recruit the co-receptor, and truncating the extension restores activity in all three [1]. That is the base rate our seven tagged cassettes were drawing from.

The two ends of the peptide are not the same engineering site

The geometric question was asked with a flood fill over the receptor structure using a 2.4 Å backbone-clearance probe: if a chain leaves the bound peptide at either terminus, can it reach bulk solvent without the peptide moving? No force field, no scoring function, nothing the audit had criticised.

The co-receptor position comes from a ternary model built on SERK-family templates [16] and AlphaFold3, because no PEPR1-BAK1 crystal exists. Co-receptor binding does not touch the N-terminal side by a single grid point, and it closes the C-terminal side almost completely: the escape path becomes longer than the six-residue tag that has to fit through it. An N-terminal tag goes into space the peptide's own disordered N-terminus already occupies. A C-terminal tag goes into the receptor-co-receptor interface.

The docking agreed, under a test permissive enough to rescue a known non-binder

A thirty-pose panel was then run on the eight constructs with one shared wild-type-derived restraint set per receptor, sampling raised to 100/30/30, clustering restored to 5.0 Å, and every engineering addition left unrestrained, so where the tag ended up was an output rather than an assumption. The result is categorical. Every construct reaches at least three of the four crystallographic clamp contacts except the C-terminally tagged one, which reaches none of them, in none of its thirty poses, with Asn23 sitting 47 to 57 Å from all four of its partners.

The panel's own negative control shows how permissive that test was: the documented non-binder 15-23 also reaches three of four clamp contacts, because it contains Asn23 and Lys18 and can therefore form them. A test generous enough to rescue a peptide known to be inactive could not rescue the tagged one.

A reproducibility check then made the same point a third time. Three constructs were re-run at two further random seeds. Clamp integrity was identical across all three seeds for all three constructs, including the failure, which repeated at 0 of 4 every time. The docking score moved by 13 to 36 units. At one seed the tagged construct scored −181.9, the best score in the entire study and better than the crystallographic positive control, while still making none of the four contacts. Every score difference this project had ever reported is smaller than that seed spread.

Codon choice is scored on the worse of two accessibilities, never on their sum

A sequence the receptor accepts is worth nothing if the chassis will not initiate translation on its messenger RNA. The expression cassette is fixed: a green-light-inducible promoter, the MF001 ribosome binding site with its AAGGAGG core, the SamyQ secretion signal, the payload, the tag, the terminator. The only free variable is synonymous codon choice, and the objective is the unpaired probability of two regions computed from the ViennaRNA base-pair matrix at 37 °C: the fifteen bases from the start codon, and the Shine-Dalgarno core. The reference is a gene that works in this chassis, ho1, at 0.681.

The first optimiser ranked roughly 50,000 synonymous draws per payload by the sum of those two scores, and the design it returned had start accessibility at 0.911 and a Shine-Dalgarno core at 0.050, more occluded than the wild-type sequence it replaced. A sum lets the search buy a large gain in one term with a total loss in the other. Ranking moved to the bottleneck, J = min(start +15, SD core), and the random draw was replaced with a deterministic descent over the codons that actually base-pair with the initiation region. On BoPEP4 the bottleneck went from 0.067 to 0.528 at a codon adaptation index of 0.78.

Two results here contradict standard practice and are worth stating plainly. Maximising codon adaptation alone produces designs whose ribosome binding site is buried, which is the opposite of the intended effect. And the published low-GC heuristic we ran as a literature control [13] made one payload worse than its own wild type. The wild-type sequence separately fails a manufacturing rule outright, with a run of seven adenines.

What this model ranks is precursor supply. It says nothing about signal-recognition targeting, translocase throughput, or the quality-control proteases waiting on the far side of the membrane, and it is not a yield prediction.

The signal peptide was screened and then left alone

A panel of B. subtilis Sec signal peptides was ranked on the determinants known to matter: n-region charge, h-region hydrophobicity and length, and the −3/−1 cleavage rule, from verified UniProt sequences. SamyQ, the signal peptide already in our cassette, lands in the favourable zone with an n-region charge of +4, a hydrophobic core mean of 1.65 and a clean A-X-A cleavage site; AprE is the best-matched alternative and AmyE the weakest. That ranking is worth what such rankings are worth: signal-peptide performance is strongly cargo-specific, and empirical screens of more than a hundred signal peptides routinely find that the winner is protein-dependent [14]. The screen chose which three to put in a wet-lab panel, and it does not replace one.

Six cassettes went to synthesis, and three of them are controls with published effect sizes

The order placed on 20 August is the design endpoint. Every mature product is either the natural sequence exactly or the natural sequence with an N-terminal tag; nothing is fused after Asn23; no spacer was inserted, so each tagged construct has an untagged twin and the tag comparison is single-variable.

• D-01BoPEP4_9-23_NHis691 bp

  SamyQ₃₁▾HHHHHHRPREPHSSGKPGGHN

  Activity lead, and the only lead that a western blot can see. Cleavage +1 residue: H.

• D-02BoPEP4_G17A_1-23_NHis715 bp

  SamyQ₃₁▾HHHHHHGILIGSKKRPREPHSSAKPGGHN

  Dead control. Pearce measured G17A at more than four thousandfold down. Cleavage +1 residue: H.

• D-03BoPEP4_WT_1-23_NHis715 bp

  SamyQ₃₁▾HHHHHHGILIGSKKRPREPHSSGKPGGHN

  Full-length positive reference. Cleavage +1 residue: H.

• D-04BoPEP4_9-23_native673 bp

  SamyQ₃₁▾RPREPHSSGKPGGHN

  The same lead with no tag. Arg at +1 is the least favourable cleavage context in the batch. Cleavage +1 residue: R.

• D-05BoPEP4_WT_1-23_native697 bp

  SamyQ₃₁▾GILIGSKKRPREPHSSGKPGGHN

  Exactly the natural molecule, and the internal control for the +1 question. Cleavage +1 residue: G.

• D-06BoPEP4_S15A_1-23_NHis715 bp

  SamyQ₃₁▾HHHHHHGILIGSKKRPREPHASGKPGGHN

  Graded control, the middle rung. Pearce measured S15A at about a hundredfold down. Cleavage +1 residue: H.

N interface anchor · A substitution against wild type · ▾ signal peptidase I cleaves here

D-01, D-06 and D-02 are the point of the order. They are wild type, S15A and G17A on one backbone: about 1×, about 100× and more than 4,000× down in Pearce's measurements [2], four orders of magnitude of expected effect. Drop D-06 and the set collapses into a binary switch that can show the peptide works but cannot show the platform ranks variants. That distinction is the difference between a positive result and a measurement.

The residual risk in this design sits at the cleavage junction. Signal peptidase I prefers small residues immediately after the cut, and four of the six constructs present a histidine there, D-04 presents an arginine followed by a proline, and only D-05 presents the ideal glycine. D-05 is therefore the internal control: if it secretes and the tagged constructs do not, the +1 residue is the cause and the peptide is exonerated. A cheaper answer already exists in the freezer, since the sequence-verified AmilCP construct uses the same signal peptide and is a chromoprotein, so its secretion can be read by eye off the supernatant.

Three design axes were retired in writing

Retiring an idea on evidence is worth more than dropping it quietly, so all three are recorded here. The point-mutation axis went, because position 8 makes no receptor contact and position 18 already makes the contacts the substitution was meant to add. Tandem repetition went: a sweep of (SSGKPGGHN)_n for n = 1…8 found buried area per residue falling monotonically with every added copy, from 127 to 32 Å² on PEPR1, and the number of copies actually touching the receptor saturating at three regardless of n. Repetition does not rescue a motif that does not bind alone. The shallow truncations Δ1-2 and Δ1-3 went because their pre-registered predictions sat between two constructs we already had, which is close to no information for a cloning slot.

One tandem idea survives in a different frame. The sweep asked whether repetition rescues binding, and it does not. It never asked whether repetition raises yield, and a cleavable-linker polypeptide giving n peptides per translation event is a real proposal. It needs a protease with clean specificity in B. subtilis, which is not a two-week problem, so it is filed as future work.

Five predictions, written down before the assay exists

These were recorded on 16 August, before any construct had arrived, so they can be wrong in public.

The assay also has to carry four controls or it cannot be interpreted: vehicle and empty-vector supernatant, because B. subtilis supernatant contains proteases and surfactin; a chemically synthesised reference peptide at a known concentration, which is the only thing that separates “the peptide does not work” from “we secreted almost none of it”; the dead fragment at the same molar load; and quantification before activity, because activity per unit peptide is the number that means something and activity per millilitre of supernatant is not.

What this modelling cannot decide

• None of it is a measurement. Two structural methods agreeing is agreement, and the decisive next step is experimental.
• Neither method resolves single substitutions. Section 7 is the evidence, and it is our own. No design on this wiki is ranked by a docking score or by a co-folding confidence value.
• PEPR2 has no experimental structure. Every PEPR2 number rests on an AlphaFold model with no per-residue confidence filter applied before docking. More docking rigour cannot fix this; a structure or a binding assay can.
• The receptors are the wrong species. BoPEP4 is a Brassica peptide and we model it against Arabidopsis PEPR1 and PEPR2, because that is where the structural and mutagenesis evidence is. The Brassica orthologues will differ.
• The most informative published result is outside the pipeline. Pearce showed that Gly17→D-Ala and Gly17→2-methyl-Ala restore activity while L-Pro and D-Pro abolish it, which proves the requirement is backbone conformational freedom [2]. D-amino acids and Aib are not representable in stock CNS topology, so three rows of our own panel are marked as not modellable.
• The canonical wild-type pose was chosen, not predicted. The docking's own top cluster was non-native at 5.80 Å; the pose carried forward is cluster three of four, selected because it matches the crystal.

The remaining computational item is a narrow, pre-registered re-run: the sixteen equal-length analogues against PEPR1 only, from unbound conformer ensembles, with crystal-derived restraints, full sampling, three seeds, and cluster-mean scoring. The pass criterion is committed before the runs start: Spearman ρ above +0.5 at p < 0.05, with at least two of G17A, G17P and S15A in the four worst-scoring analogues. Anything less means the protocol has no substitution-level resolution and every remaining score-difference claim comes out, whatever the numbers look like.

Every parameter, and where the runs live

Three numbering conventions coexist in the source data and cross-study comparisons are only valid inside one of them: mature 1-23 as used by the truncation and Pearce panels, renumber-from-one as used by the co-folding truncation jobs, and construct-local numbering in the tandem study. Everything on this page is in mature numbering.

The methodology audit reproduces the complete parameter and restraint record for all 121 runs and all 44 co-folding jobs, lists twelve references split into five that support the protocol and seven that challenge it, and triages every downstream claim. The page you are reading is the summary; the audit is the document to check it against, and both are in the dry-lab record alongside the run trees, the restraint tables and the analysis scripts.