ImmunoStruct

The same nine-residue peptide can provoke a CD8+ T-cell response on one HLA allele and pass unnoticed on another, and a single somatic substitution can flip a tumor peptide from immunologically silent to visible. What differs is often not the sequence but the conformation the peptide adopts in the MHC class I groove — how far it bulges, which side chains face outward, how much surface a T-cell receptor can reach. Predictors built on amino acid strings alone cannot see that, which is part of why only a small fraction of computationally nominated neoantigens elicit a measurable response.

ImmunoStruct makes structure a first-class input. Every peptide-MHC pair is jointly folded by a high-throughput AlphaFold 2 pipeline, then converted into a graph whose nodes are residues carrying amino acid identity and 3D coordinates and whose edges mark peptide bonds, hydrogen bonds, hydrophobic contacts, and ionic interactions. E(n)-equivariant graph layers process that graph, so the representation survives rotation and translation of the complex. The structural embedding is fused with a variational autoencoder over the full peptide-MHC sequence and a learned embedding of computed biochemical properties — two meta-features distilled from the ten highest-variance properties of the peptide plus its peptide-MHC solvent-accessible surface-area score.

It was built at Yale University across Smita Krishnaswamy's computational lab and Akiko Iwasaki's immunobiology lab, with Howard Hughes Medical Institute support, posted as a preprint in November 2024 and published in Nature Machine Intelligence after peer review.

#Key Features

• Attention over three modalities: A 32-dimensional sequence latent, a 64-dimensional structural embedding, and an 8-dimensional biochemical embedding concatenate into a 104-dimensional vector that an eight-head self-attention layer reweights per sample, rather than treating the modalities as interchangeable.
• Foreignness pretraining, then transfer: The network first regresses a thermodynamic foreignness score — how closely a peptide resembles known pathogen sequences — and is then fine-tuned on binary immunogenicity with the final MLP reinitialized. Ablations show this beats training from scratch.
• Cancer-wildtype contrastive learning: A paired objective pushes a mutant peptide-MHC away from its wildtype counterpart in latent space only when the mutation is immunogenic, writing central tolerance directly into the geometry of the representation.
• Readable structural attention: Attention weights from the equivariant graph module are mapped back onto the folded complex, so a call can be traced to specific peptide and MHC residues; for KRAS G12V versus G12D the difference localized to the peptide C-terminus rather than the mutated position.

#Technical Details

The released checkpoints are compact — roughly six million parameters, most of them in the autoencoder's 512-unit hidden layers over a 283 × 21 one-hot encoding of the concatenated peptide and MHC sequence. Six E(n)-equivariant graph convolution layers with 64 hidden channels handle the structure graph, and a two-layer MLP maps the fused vector to a single logit. Training used 26,049 peptide-MHCs across the 27 most common class I HLA alleles, drawn from IEDB for infectious-disease epitopes and CEDAR for cancer neoepitopes. On the held-out IEDB split, ImmunoStruct reached an AUROC of 0.882 ± 0.005, an AUPRC of 0.696 ± 0.020, and a mean PPVn of 0.514 ± 0.020; the strongest sequence-based comparator in the paper's benchmark, BigMHC-IM, reached 0.462 AUPRC, with DeepNeo, MHCflurry 2.0, MHCnuggets, NetMHCpan-4.1, and PRIME2.1 between 0.212 and 0.269.

#Applications

Two evaluations exercise the frozen checkpoint on data disjoint from training. An IFN-γ ELISpot assay run for 19 SARS-CoV-2 peptide-HLA pairs matched 15 of 19 measured immunogenicities, at 79.0% accuracy and 0.780 AUROC. Scoring a cancer patient cohort with no retraining stratified both overall and progression-free survival, with hazard ratios of 0.54. The practical use is ranking candidate epitopes before synthesis and wet-lab screening: a vaccine construct such as a lipid nanoparticle carries only 20-30 peptides, so the order of the candidate list decides whether the construct contains anything immunogenic at all.

#Impact

Code, the two pretrained checkpoints, and the full multimodal corpus — AlphaFold 2 structures, prebuilt graphs, and the curated label tables — are all public, under a Yale non-commercial license that rules out commercial use without a separate agreement; both HuggingFace repositories are gated, though approval is automatic once an account agrees to share contact information. That release makes the expensive half of the pipeline reusable, since the folded peptide-MHC dataset no longer has to be regenerated to build on this work. The limitations the authors state are real ones. Immunogenicity labels come from assays that vary between protocols and cell lines; TCR information, which would explain much of the residual variance, is absent; coverage stops at 27 alleles, so rare HLA types are out of scope; and the survival cohort is small and mixes immunotherapy regimens. There is also no end-to-end tool for scoring arbitrary peptides — new candidates must be pushed through the folding pipeline first.