ImmuneCLIP

A peptide epitope displayed on an MHC molecule is not recognized by one T-cell receptor but by a whole family of them, and the same receptor often responds to several related peptides. Cast as a classification problem, that many-to-many structure forces a separate model per antigen and demands negative examples that binding assays rarely supply. ImmuneCLIP recasts it as retrieval instead: embed receptors and peptides into a single space, and let cosine similarity say which pairs belong together, so an unseen receptor can be scored against an arbitrary pool of candidate epitopes.

The model was built by Chiho Im, Ryan Zhao, Scott Boyd and Anshul Kundaje in the Departments of Computer Science, Pathology and Genetics at Stanford University, posted as a preprint in October 2024 and published in the RECOMB 2025 proceedings. It adapts the CLIP recipe from vision-language modeling to immune recognition: one encoder reads the peptide epitope, another reads the TCR CDR3 α and β chains, and both project into a shared latent space trained with a contrastive objective.

What the encoders are is the second design choice. Rather than training sequence models from scratch on the few thousand labeled pairs available, ImmuneCLIP takes pretrained protein language models — ESM-2 or ESM-3 for epitopes, TCR-BERT or TCRLang-Paired for receptors — freezes their weights, and inserts low-rank adapters. That places it in the same family as CLIPepPI, which applies contrastive fine-tuning to domain-peptide binding, and sets it against TULIP and STAPLER, the sequence-based TCR specificity models it benchmarks against.

#Key Features

• Shared receptor-epitope latent space: Scoring is the cosine similarity between two embeddings, so a trained checkpoint ranks any epitope pool for a receptor it has never seen without refitting.
• Frozen backbones with LoRA adaptation: Rank-8 adapters on the key and value matrices of the self-attention layers leave roughly 2M trainable parameters, matching full fine-tuning while keeping GPU memory under 4 GB.
• A loss built for many-to-many binding: Because one epitope binds many receptors, off-diagonal pairs in a minibatch are not all negatives; a binary cross-entropy objective over the true label matrix replaces the standard CLIP softmax.
• Learns from small labeled sets: With only 5% of the training pairs the model still reaches top-1, top-5 and top-10 hit rates of roughly 0.22, 0.30 and 0.41, and performance plateaus well before the full set is used.
• Sequence-only inputs: CDR3 α/β and peptide sequences are all that is required, sidestepping the scarcity of resolved TCR-pMHC structures.

#Technical Details

Training uses the MixTCRpred-curated dataset, which pools TCR-pMHC records from IEDB, VDJdb, McPAS-TCR and 10x Genomics. From 17,715 initial pairs, duplicate filtering leaves 14,245, and restricting to human data yields the final 8,439 unique TCR/epitope pairs covering 127 epitopes, split 70:15:15 with low-count epitopes upsampled during training. Per-residue embeddings from the final hidden layer are pooled and projected through MLP layers into a 512-dimensional space; tokens are masked individually with probability 0.15, and a batch size of 16 keeps redundant epitopes out of each minibatch. On epitope ranking, ImmuneCLIP reaches top-1, top-5 and top-10 hit rates of 0.506, 0.661 and 0.748, against 0.464/0.614/0.700 for tcrdist3 and 0.453/0.603/0.699 for Levenshtein distance. On binary binding prediction over VDJdb test splits curated by other groups, with overlapping pairs removed, it reaches AUROC 0.88 and average precision 0.70 versus TULIP's 0.87 and 0.66, and AUROC 0.81 versus STAPLER's 0.79. The ablation shows the pretrained backbone carries the result: a one-hot receptor encoder collapses to 0.002 top-1, while TCR-BERT (0.506) beats ESM-2 (0.459) and TCRLang-Paired (0.392).

#Applications

The natural use is epitope deconvolution — given TCR sequences from repertoire sequencing, rank which antigens a clonotype is likely to recognize, or run the query in reverse to shortlist receptors against a target peptide for immunotherapy and vaccine work. Because scoring needs only CDR3 and peptide sequences, it applies to the large body of repertoire data collected without paired structural or MHC-multimer information, and its behavior under data scarcity makes it usable for rare epitopes where only a handful of binding receptors are known.

#Impact

ImmuneCLIP is a compact demonstration that immune recognition benefits from the same contrastive alignment that connected images to captions, and that parameter-efficient adaptation of general protein language models can match or beat purpose-built architectures trained on the same labels. Its limits are stated plainly by the authors: the training set spans only 127 epitopes, and performance degrades on test epitopes dissimilar to those seen during training, so it is a tool for interpolating within known antigen space rather than a zero-shot predictor for novel epitopes. Code is released under an MIT license and the trained checkpoint is deposited on Zenodo.