BeitAI-pHLA

Mass spectrometry of eluted ligands is the most direct way to see which peptides a cell actually displays, but it never says which HLA molecule was doing the displaying. A single human cell expresses up to six HLA class I alleles, so every peptide recovered from that experiment arrives with six candidate assignments and no label. Predictors that want single-allele training data must first guess those labels — NetMHCpan 4.1 annotates them with a semi-supervised pass, MHCflurry 2.0 trains on single-allele and affinity data instead — and whatever that pre-decomposition gets wrong propagates downstream.

BeitAI-pHLA, from Kindstar Biotech with Peking University First Hospital, treats the ambiguity as the modelling problem rather than a preprocessing step. Each peptide becomes a bag of the alleles its sample expresses, and a learned attention module over that bag is trained end to end against the bag-level binding label. The attention weights are the deconvolution: whichever allele the model leans on hardest is the one the peptide is assigned to, and that assignment falls out of the same gradient step as the binding score. Published in Computational and Structural Biotechnology Journal in September 2026, the name is the authors' own acronym — Binding Estimation using Immune Technology of Artificial Intelligence — and is unrelated to the BEiT vision transformer.

#Key Features

• Bag-of-alleles multi-instance learning: The set of alleles a sample expresses forms a weakly supervised bag with a single binary label, and a three-head attention module pools the instances into one bag representation. Parameters update every batch rather than once per epoch as in NNAlign_MA.
• Protein language model embedding: Peptide and HLA pseudo-sequence are concatenated and read through a frozen ESM-2 150M encoder, replacing the BLOSUM matrix and sparse encodings that NetMHCpan and MHCflurry rely on.
• Capsule feature extractor: A residual convolutional stack feeds a capsule layer whose vector-valued units encode relative position and orientation, which the ablations show contributes more than either the language model or the MIL head.
• Motif deconvolution for free: Peptides grouped by their highest-attention allele are passed to Seq2Logo, producing per-allele, per-length binding motifs without presetting a cluster count.
• Hosted web server: A public tool accepts pasted peptides and an HLA selection, with a three-checkpoint ensemble available for local batch runs.

#Technical Details

Training used 10.67 million mass-spectrometry-derived records spanning 144 alleles drawn from IEDB, with affinity measurements discarded in favour of eluted ligands alone, roughly two multi-allele entries for every single-allele one, and Swiss-Prot decoys at a 1:20 ratio, length-matched so that peptide length cannot become a shortcut. Leakage was blocked by CD-HIT at 85% identity, then merging clusters sharing any of six anchor residues into indivisible superclusters. Peptides up to 15 residues and the 34-residue HLA pseudo-sequence concatenate into a 52-token input, embedded at 640 dimensions per token; the capsule layer emits 20 ten-dimensional capsules (a 200-dimensional vector) and the attention module a 1×600 bag representation before a sigmoid head. On an external multi-allele set of 78,011 ligands and 312,044 decoys from 21 meningioma samples, it reached AUROC 0.988, AUPRC 0.967 and PPV 0.954, ahead of MixMHCpred 2.2, NetMHCpan 4.1, MHCflurry 2.0, TripHLApan and CapsNet-MHC, and it degraded most slowly as decoy ratios rose from 1:4 to 1:50. Held-out alleles absent from training gave AUROC 0.969 and AUPRC 0.876. Against MHC Motif Atlas references its deconvolved motifs matched on PSSM correlation at 0.485 versus NetMHCpan 4.1's 0.427 — the 13.6% gain the paper headlines — while PFM correlation was indistinguishable (0.961 versus 0.963). On 349 immunogenic and 1,838 non-immunogenic neoepitopes it placed second on AUPRC at 0.248, just behind TripHLApan's 0.252.

#Applications

The model is aimed at immunopeptidomics groups sitting on multi-allele eluted-ligand data that they would otherwise have to deconvolve with GibbsCluster or MixMHCpred before use, and at neoantigen pipelines for cancer vaccines and cell therapy, where the ranked shortlist handed to synthesis and T-cell assays is the expensive step. Small jobs run through the web server; larger screens run the released ensemble locally.

#Impact

BeitAI-pHLA's contribution is architectural rather than a headline accuracy number: it shows that the allele-assignment ambiguity in multi-allele immunopeptidomics can be carried through training as weak supervision instead of resolved in advance, and that the resulting attention weights recover binding motifs closer to reference than the semi-supervised alternative. The scope is deliberately narrow — HLA class I only, peptides of at most 15 residues, and binding alone, with proteasomal cleavage, TAP transport and T-cell receptor recognition all left to other tools. Absolute immunogenicity performance remains low for every predictor tested, BeitAI-pHLA included — presentation is necessary but far from sufficient to provoke a response. Source code and the three ensemble checkpoints are public, though the repository carries no licence and the web server is offered for academic use.