AIKI-HLA

Measuring peptide–MHC generalization to unseen alleles across both HLA classes
Class I + Class II 421 / 507 HLA alleles +0.22 to +0.37 vs SOTA 0.816 strictly novel Apache-2.0 + CC-BY-4.0
Research use only All three demos below serve the 5.8 M / 507-allele research-use ensemble (CC-BY-NC-4.0) — outputs cannot be used for commercial purposes. For commercial use, pip install aiki-hla ships the 4.66 M / 421-allele open ensemble (CC-BY-4.0) with the same architecture and recipe.
Tool 01 / 03
Scan a whole protein
Sliding-window peptides × HLA panel. Per-residue hot-spot map, ESMFold structure, ranked risky peptides, per-position deimmunization.
Try it
Tool 02 / 03
Rank cancer-vaccine candidates
Somatic mutations → ranked neoantigens. Differential Agretopicity Index, Patient HLA Binding Rank, 27-mer Long Synthetic Peptide cassette.
Try it
Tool 03 / 03
Score one peptide × HLA
For unit-testing or one-off scoring. Reveals the factorization pligand, pbinding, pcomposite directly, plus the predicted peptide–MHC complex.
Try it
Tool 01 / 03

Scan a whole protein for HLA hot-spots

Paste a sequence (or use the influenza M1 default) → per-residue immunogenicity heat-map, contiguous hot-spots, ranked risky peptides, and the ESMFold-predicted 3D structure, all in one view.

Default: influenza A/Puerto Rico/8/1934 M1, residues 1–104. The published validation case — the A*02:01-restricted GILGFVFTL epitope falls at residues 58–66 within this window.
Pre-loaded example shown below — click Run scan to compute live with your current settings.
Promiscuity per residue (MAX-then-COUNT across the panel): safe → amber → orange → high risk. Risky peptides ranked by pcomposite = √(pligand·pbinding).
↑ Top Next: rank neoantigens →
Tool 02 / 03

Cancer-vaccine neoantigen ranking DAI · PHBR · LSP

Paste a wild-type protein, list somatic missense mutations (e.g. G12D), and optionally the patient's HLA type. Ranks mutant peptides by Differential Agretopicity Index (DAI = log2 pMT/pWT; Duan Nat Genet 2014, Łuksza Nature 2017) so the most immune-visible mutants surface at the top — the field-standard signal for an immunogenic neoantigen.

Default: KRAS WT (P01116-2) with the G12D driver mutation pre-loaded.
Accepts G12D or 12:G>D.
Leave blank to score against a standard supertype panel.
Ready — sample KRAS G12D pre-loaded.
3D structure
Loading structure…
ESMFold-predicted backbone. Mutation residue highlighted in gold; the top neoantigen peptide span (containing the mutation) is the immune-visible candidate.
DAI > 0: mutant more visible than WT. DAI > 1: 2× presentation gain (Łuksza 2017 cut-off). PHBR < 2 %: presented by at least one patient HLA (Marty 2017). LSP column = 27-mer cassette for BNT122 / mRNA-4157 format.
← Back: whole-protein scan Next: score one peptide × HLA →
Tool 03 / 03

Score a single peptide × allele pair

For unit-testing or one-off scoring — reveals the factorization (pligand, pbinding, pcomposite) directly, alongside an ESMFold-predicted peptide–MHC complex for the same pair.

Ready
3D peptide–MHC complex
Loading default GILGFVFTL × HLA-A*02:01 complex…
Single-chain construct (MHC G‑domain + (GGGGS)4 linker + peptide) folded by ESMFold. The linker is a tethering device; the peptide pose is illustrative of where ESMFold places it under that constraint, not a docking prediction. Confidence (mean peptide pLDDT) shown above the viewer.
p_aiki = pbinding (allele-specific groove fit) · p_gate = pligand (peptide-only ligand-likeness) · p_composite = √(p_aiki·p_gate) is the released model. Safety screening ≥ 0.99; neoantigen ranking 0.3–0.7.
← Back: rank neoantigens Read the paper ↓

About AIKI‑HLA

0.765
strictly novel (39 alleles)
0.913
conventional split
+0.22 to +0.37
vs 8 open tools (p<10-9)
421 / 507
HLA alleles · I + II

AIKI‑HLA factorises presentation into a peptide-only ligand-likeness term pligand and an allele-specific binding term pbinding, combined as the parameter-free geometric mean √(pligand·pbinding). Both Class I and Class II flow through one head over the 182-residue MHC G-domain on a frozen ESM-2 650M backbone, deployed as a 3-seed ensemble. Open ensemble 4.66M / 421 alleles (CC-BY-4.0). Research-use 5.8M / 507 alleles (CC-BY-NC-4.0).

Caveats: Class I overconfident (ECE 0.15) vs Class II well-calibrated (0.05). Off-target screening ≥ 0.99. Cluster-novel generalisation is heterogeneous (HLA-A*01:01 cluster collapses to chance when purged).

API

All routes under https://aikium--aikihla-landing-page.modal.run/

GET /score

One (peptide, allele) pair. ?mode=both returns p_aiki, p_gate, p_composite.

curl '.../score?peptide=GILGFVFTL&allele=HLA-A*02:01&mode=both'

POST /score_batch

Up to 1000 pairs. JSON or CSV.

curl -X POST '.../score_batch' -d '{"pairs":[...], "mode":"both"}'

POST /scan · /score_mutations · /score_neoantigens

The three demos above, programmatic. Tool 01 / 02 take a sequence; Tool 02 also takes mutations and patient HLA.

POST /fold_complex · /fold_structure

Peptide–MHC complex (single-chain construct via ESMFold), or any protein sequence alone.

GET /gate · /alleles

Peptide-only viability score; supported allele list.

Install locally

pip

pip install aiki-hla
aiki-hla score --peptides p.csv --alleles a.csv --out o.csv

Docker

docker pull ghcr.io/aikium-public/aiki-hla:latest
docker run -v aikihla-models:/models \
  ghcr.io/aikium-public/aiki-hla:latest \
  score --peptides p.csv --alleles a.csv --out o.csv

Cite

Mysore, V. (2026). Measuring peptide–MHC generalization to unseen alleles across both HLA classes. Submitted, PLOS Computational Biology. Data + weights: 10.5281/zenodo.20520819 (concept DOI → latest; v1.1.0). Code: github.com/aikium-public/aiki-hla.

@article{aikihla2026,
  author = {Mysore, Venkatesh},
  title  = {Measuring peptide-{MHC} generalization to unseen alleles across both {HLA} classes},
  year   = {2026},
  journal= {PLOS Computational Biology},
  note   = {Submitted. Zenodo 10.5281/zenodo.20520819 for data + code.}
}

Built on

IEDB · HLA Ligand Atlas · CEDAR · NetMHC / NetMHCpan / NetMHCIIpan / NetMHCstabpan (DTU; via IEDB Tools API) · VDJdb · TransPHLA · BigMHC · MHCflurry 2.0 · IPD-IMGT/HLA · IPD-MHC · UniProt · NCBI · ESM-2 650M (Lin 2023, Meta AI — MIT) · ESMFold Atlas (Meta AI) · mmseqs2 · 3Dmol.js · Modal · Zenodo.