Stability explorer
Ten models, trained on peptide-HLA complex half-lives, scored on binding grooves none of them had seen. Pick a model, type a peptide, see what it predicted and what the experiment actually measured.
Team Fridaymerchants Paing Hein Htet · Radin Moradi · Sina Ahani · Andy Kaça
Drug and protein design track, set by Serova · Halkin, 1–2 Paris Garden, London SE1
Organised by Iterate and futurebio.xyz, part of London Deep Tech Week.
Pick a model
Bars show median per-allele Spearman across 21 leave-one-groove-cluster-out folds. Higher is better, 0 is no skill, 1 is a perfect ordering.
Why there is no NetMHCstabpan column
NetMHCstabpan trained on the entirety of this dataset, so Serova's own challenge brief calls it unfair as a direct comparator. Every prediction here is instead compared against the measured half-life, which is ground truth rather than another model's guess.
Predict a peptide
Pick an allele, then any 9-mer measured against it. Only measured pairs appear, because every prediction here is a stored held-out result.
In the groove
Your peptide's residues mapped onto a real peptide-HLA crystal structure, with the anchor positions marked.
Drag to rotate · swipe sideways or press ←→ to pan · pinch to zoom · scrolling up and down moves the page
This is a template, not a predicted structure
The coordinates are PDB 3GSN, HLA-A*02:01 presenting NLVPMVATV with the RA14 T cell receptor. Your residues are labelled onto its nine positions so you can see where each one sits. We do not fold your peptide, and the side-chain geometry shown is the crystal's, not yours.
Design: which substitution makes it stick?
Stability lives at positions 2 and 9. Swap an anchor for one that fits the pocket better and the complex survives longer, while positions 3 to 8 — the face a T cell actually reads — stay untouched. That is anchor optimisation, and modified peptides of this kind are used in real melanoma vaccine trials.
A real melanoma driver
We enumerate, we do not optimise
A 9-mer has 171 single mutants. We score all of them and sort. We do not run a search over sequence space, because a forward model at this accuracy would be driven straight into its own blind spots and hand back confident nonsense. Staying inside single mutants of a real measured peptide is the safety mechanism.
And the tool refuses when it should
Abacavir and HLA-B*57:01
A drug that reshapes a binding groove, and a self-peptide repertoire that was never tolerised against. The same measured dataset as the rest of this site, read for a different question.
The thing this tab exists to correct inference
Abacavir hypersensitivity is often described as the drug stabilising HLA-B*57:01 complexes. That is the wrong mechanism. Abacavir sits in the F pocket and changes which peptides the groove accepts at all, so the result is a different repertoire: self-peptides never previously presented, and therefore never tolerised against.
The precise claim matters. Abacavir does not globally stabilise the pre-existing repertoire. It does selectively slow the dissociation of particular complexes, and that selectivity is part of how a different repertoire comes to be loaded. Both halves of that sentence are load-bearing.
1 Abacavir in the F pocket tier 2
PDB 3VRI at 1.6 Å: HLA-B*57:01 with β2m, the self-peptide RVAQLEQVYI, and one molecule of abacavir.
Drag to rotate · swipe sideways or press ←→ to pan · pinch to zoom · scrolling up and down moves the page
Figure 1. PDB 3VRI, 1.6 Å, one complex in the asymmetric unit; ligand 1KX, confirmed as abacavir (C14H18N6O). Chains A heavy, B β2m, C peptide. The contact residues below were computed from these coordinates at a 4.5 Å heavy-atom cutoff, not read off a figure. Length mismatch: this peptide is a 10-mer, while every peptide in the tier-1 data is a 9-mer. Deposited by Ostrov et al., 10.1073/pnas.1207934109.
What abacavir is touching
Every HLA residue with a heavy atom within 4.5 Å of abacavir:
It also contacts the peptide directly. Closest approach to the C-terminal isoleucine is 3.74 Å, so the drug is in physical contact with the anchor residue whose preference it changes.
2 Where the groove actually grips tier 1
A raw frequency logo would mislead here, because the tested panel is pre-selected: before a single half-life is read, position 9 is already entirely aromatic or large-hydrophobic. So this asks a question the data can answer instead. Among peptides measured for B*57:01, which residues are over-represented in the most stable third?
Figure 2. Enrichment in percentage points of each residue's frequency in the most-stable third versus its frequency across all peptides measured for HLA-B*57:01. Source Rasmussen et al. 2016, computed here. n . Above the axis is enriched among stable peptides, below is depleted. Only residues making up at least 2% of the panel at that position appear. No error bars: these are exact frequency differences of a finite sample, not estimates, so a bar carries no confidence interval.
Two positions dominate and the rest is noise-level. They are P2 and P9, the two anchors, recovered from half-life alone with no structural input — the same result the design tab's model reaches by a completely different route.
3 The shift
What the groove prefers at its C-terminus with the pocket empty, and with abacavir in it. These two panels deliberately do not look alike: one is a measured distribution, the other is a direction reported in words. Matching their visual weight would overstate what the literature quantifies.
Pocket empty tier 1 measured
Bulky aromatics anchor best. Isoleucine and leucine are the two weakest of the six residues that appear at all.
Abacavir bound tier 2 literature
Novel drug-induced peptides lacked typical carboxyl (C) terminal amino acids characteristic of the HLA-B*57:01 peptide motif and instead contained predominantly isoleucine or leucine residues. Norcross et al., AIDS 2012 · 10.1097/QAD.0b013e328355fe8f
Mass spectrometry of peptides eluted from abacavir-treated B*57:01 cells. The source reports a direction, not a distribution: there are no published per-residue frequencies to plot opposite the panel on the left, so none are drawn.
Valine is not in this claim. The shift is reported as isoleucine or leucine. Valine gets added when the finding is restated; it is not in the source, and as the left panel shows, it was never tested here either.
Figure 3. Left: median half-life per C-terminal residue, HLA-B*57:01, all measured rows including left-censored ones; source Rasmussen et al. 2016, computed here, n printed on each bar. Bars are medians with no dispersion shown — see Figure 4. Right: direct quotation, no derived values.
4 The stability landscape tier 1
The full distributions behind those medians, on a log axis because half-life spans nearly three orders of magnitude.
Figure 4. Complex half-life by C-terminal residue, HLA-B*57:01. Source Rasmussen et al. 2016, computed here. n = 349 peptides, group sizes under each box. Boxes are Tukey: box spans the first to third quartile, heavy line is the median, whiskers reach the furthest point within 1.5 interquartile ranges, points beyond are drawn individually. Censoring: 46 of 349 measurements sit at exactly 0 h, meaning the complex dissociated faster than the assay could resolve. Zero has no place on a log axis, so those are open marks on the floor line at 0.05 h, counted per group. They are included in every quartile and median shown.
Why this is a hypersensitivity risk and not a binding story tier 3 inference
The residues abacavir newly favours, isoleucine and leucine, sit at the bottom of the drug-free landscape: 3.65 h and 3.00 h against 8.50 h for tryptophan. The novel peptides are not better binders the drug unmasks. They are peptides the groove would ordinarily pass over.
That is what makes them dangerous. The risk comes from novelty of presentation rather than tighter binding: self-peptides a B*57:01 carrier's T cells have never been shown, so there was no opportunity to tolerise. This paragraph is our reading of the two tiers side by side, not a claim either source makes in this form.
5 The natural control: B*58:01 tier 1
Abacavir binds B*57:01 and not its close relative B*58:01, which hands us a control for free. The comparison is sharper than it looks, because 218 peptides in this dataset were measured against both alleles. Paired, not two separate panels.
Figure 5. Every difference between the HLA-B*57:01 and HLA-B*58:01 peptide-binding domains. Source the α1 and α2 sequences shipped with Rasmussen et al. 2016, compared position by position here; n = 4 differences across 182 residues. Faint marks are the 14 residues within 4.5 Å of abacavir in PDB 3VRI. Not a statistical figure, so there is nothing to put an interval on.
One residue in common tier 3 inference
Four substitutions separate these two grooves, and exactly one, position 97, valine in B*57:01 and arginine in B*58:01, falls inside abacavir's contact shell. A small hydrophobic side chain replaced by a large positively charged one, in the space the drug has to occupy.
The coincidence is measured; reading it as the reason abacavir discriminates between the alleles is our inference. We have not sourced a mutagenesis experiment isolating residue 97, and we say "in the contact shell" rather than naming a pocket, because position 97 is assigned to pocket C, E or F depending on whose definition you use.
Figure 6. Median half-life by C-terminal residue for the same peptides measured on both alleles. Source Rasmussen et al. 2016, computed here. n = 218 paired peptides, per-residue n printed at each line. Lines join two medians for one residue group and show reordering, not significance. Groups with fewer than 8 pairs are faint and carry no interpretation.
| on the 218 shared peptides | B*57:01 | B*58:01 |
|---|---|---|
| median half-life | 7.20 h | 5.55 h |
| C-terminal tyrosine, median | 4.30 h | 1.30 h |
| C-terminal tryptophan, median | 8.60 h | 10.00 h |
| Spearman between the two alleles | 0.649 | |
| paired Wilcoxon | p = 0.090 | |
| B*57:01 the more stable of the pair | 47% of 218 | |
The two alleles are not globally different in stability: the paired test misses significance and B*57:01 wins fewer than half the pairs. What four substitutions do is reorder which peptides survive. C-terminal tyrosine loses more than three quarters of its half-life while tryptophan slightly gains.
Taking the unpaired medians instead, 6.50 h against 2.40 h, would have implied a large global difference. That gap is mostly an artefact of which peptides each allele happened to be tested against, which is exactly why this panel is paired.
6 Drug-bound half-life deliberate gap
This panel is empty and the reason is the point. There is no drug-bound complex half-life in this dataset, because every measurement in it is drug-free, and we did not find a published half-life in hours for an abacavir-bound HLA-B*57:01 complex to put beside it.
The closest sourced statement is kinetic rather than numeric. Illing et al. report that abacavir had
minimal impact on the maturation or average stability of HLA-B*57:01 molecules [but] was able to differentially enhance the formation, selectively decrease the dissociation, and alter tapasin loading dependency of certain HLA-B*57:01-peptide complexes. Illing et al., Frontiers in Immunology 2021 · 10.3389/fimmu.2021.672737
Which is the nuance the callout at the top turns on: average stability barely moves, while particular complexes dissociate more slowly. An interpolated number here would have looked better and would have been invented.
Sources
- Rasmussen et al. 2016, the NetMHCstabpan stability dataset. Every tier-1 number here is computed from it. 10.4049/jimmunol.1600582
- Illing et al., Nature 2012. Abacavir binds non-covalently to HLA-B*57:01, lying across the bottom of the cleft and reaching into the F pocket where a C-terminal tryptophan normally anchors the peptide. 10.1038/nature11147
- Ostrov et al., PNAS 2012. Abacavir binds within the F pocket, altering specificity; self-peptides presented only with the drug were recognised by patients' T cells. Depositors of PDB 3VRI. 10.1073/pnas.1207934109
- Norcross et al., AIDS 2012. Eluted-peptide mass spectrometry: drug-induced peptides carried predominantly isoleucine or leucine. Also reports those peptides bind with high affinity not altered by adding abacavir. 10.1097/QAD.0b013e328355fe8f
- Illing et al., Frontiers in Immunology 2021. Kinetics of the remodelling; source for the selective-dissociation nuance. 10.3389/fimmu.2021.672737
HLA-B*57:01 genotyping before abacavir prescription is required by the FDA and the EMA. Nothing here is clinical guidance.
Citations
Every DOI below was resolved live against Crossref and the returned title, first author and year compared against what we were calling it. Anything that failed that check is in the second list, not the first.
Verified, and used
| Used for | Reference |
|---|---|
| The dataset. Serova's brief makes this one mandatory | Rasmussen M, et al. Pan-Specific Prediction of Peptide-MHC Class I Complex Stability, a Correlate of T Cell Immunogenicity. J Immunol 2016;197(4):1517–24. 10.4049/jimmunol.1600582 |
| Why stability and not affinity | Harndahl M, et al. Peptide-MHC class I stability is a better predictor than peptide affinity of CTL immunogenicity. Eur J Immunol 2012;42(6):1405–16. 10.1002/eji.201141774 |
| Supertypes, the fold-design check | Sidney J, et al. HLA class I supertypes: a revised and updated classification. BMC Immunol 2008;9:1. 10.1186/1471-2172-9-1 |
| The 34 peptide-contact positions | Reynisson B, et al. NetMHCpan-4.1 and NetMHCIIpan-4.0. Nucleic Acids Res 2020;48(W1):W449–54. 10.1093/nar/gkaa379 |
| HLA sequences | Barker DJ, et al. The IPD-IMGT/HLA Database. Nucleic Acids Res 2023;51(D1):D1053–60. 10.1093/nar/gkac1011 |
| ESM-2 | Lin Z, et al. Evolutionary-scale prediction of atomic-level protein structure with a language model. Science 2023;379(6637):1123–30. 10.1126/science.ade2574 |
| Counter-evidence: binding prediction survives the allele gap | Atkins C, et al. Geographically Biased Composition of NetMHCpan Training Datasets and Evaluation of MHC-Peptide Binding Prediction Accuracy on Novel Alleles. bioRxiv 2023. 10.1101/2023.09.03.556092 |
| Abacavir binds the F pocket, non-covalently | Illing PT, et al. Immune self-reactivity triggered by drug-modified HLA-peptide repertoire. Nature 2012;486:554–8. 10.1038/nature11147 |
| Abacavir alters specificity; depositors of PDB 3VRI | Ostrov DA, et al. Drug hypersensitivity caused by alteration of the MHC-presented self-peptide repertoire. PNAS 2012;109(25):9959–64. 10.1073/pnas.1207934109 |
| The C-terminal shift to isoleucine or leucine | Norcross MA, et al. Abacavir induces loading of novel self-peptides into HLA-B*57:01. AIDS 2012;26(11):F21–9. 10.1097/QAD.0b013e328355fe8f |
| Selective dissociation, not global stabilisation | Illing PT, et al. Kinetics of Abacavir-Induced Remodelling of the MHC Class I Peptide Repertoire. Front Immunol 2021;12:672737. 10.3389/fimmu.2021.672737 |
| Structures rendered on this site | PDB 3GSN, HLA-A*02:01 with NLVPMVATV and the RA14 receptor · PDB 3VRI, HLA-B*57:01 with abacavir, 1.6 Å |
One error this check caught
We had been calling the geographic-bias paper “Barton et al.” throughout. The first author is Atkins. It was wrong in the written brief and was corrected once Crossref returned the real record. That is the whole reason for running the check rather than trusting a reading list.
Found, but deliberately not cited
These surfaced in searches and were never resolved, so they appear nowhere on this site, in the deck, or in the write-up.
- A bioRxiv preprint on peptide-MHC stability with protein language models. Found in a search, DOI never resolved, never read.
- MINT, an interaction-aware model pretrained on affinity and fine-tuned on half-life. Same: surfaced, never verified.
- Fasoulis et al. 2024, ImmunoInformatics, the source of the external NetMHCstabpan numbers in our baseline notes. Its DOI is unverified, so those numbers are not quoted anywhere.
Tools and credits
GDM Science Skills (Apache-2.0) · UniProt reviewed proteomes, for epitope source mapping · 3Dmol.js for the structure viewers · RCSB PDB for coordinates · Crossref for DOI verification · compute and API credits from Modal, Anthropic and Hugging Face, and literature search via Amass.
The rule we worked to
No reference goes on a slide or into the write-up unless it has been resolved against Crossref first. It takes ten seconds and it already caught one wrong author name.