DapPep

Known TCR-peptide pairs are distributed with a long tail. A handful of well-studied viral epitopes account for thousands of experimentally validated receptors, most peptides in the databases carry only a few, and the peptides that matter most clinically — a somatic neoantigen unique to one patient's tumor — carry none at all. A predictor trained on the binding patterns of one specific peptide therefore cannot be applied in exactly the case worth predicting. PanPep, the method DapPep is benchmarked against, handles this by treating each peptide as a meta-learning task: fine-tune a peptide-specific learner on whatever small support set exists, and distill a mapping from peptide encoding to learner when none exists. DapPep instead makes the peptide an input to a single fixed model rather than a task the model must be adapted to.

The architecture is deliberately small. A receptor encoder, initialized from ESM-2, embeds the TCR sequence; a shallow self-attention stack embeds the peptide, which is typically under ten residues and needs far less capacity. The two meet in a cross-attention block where the key and value come from the receptor and the query from the peptide, so a new peptide interrogates learned receptor structure residue by residue instead of the two sequences simply being concatenated. A mean-pooling linear decoder turns the joint representation into a binding score in [0, 1].

DapPep was developed by Jiangbin Zheng, Qianhui Xu, Ruichen Xia and Stan Z. Li at Westlake University, Zhejiang University and Fudan University, posted to arXiv in November 2024 and published at ICASSP 2025.

#Key Features

• One checkpoint for every data regime: the same trained weights are applied to peptides with abundant known binders, with a handful, and with none, with no continued training or fine-tuning at inference.
• Peptide as query, receptor as context: taking the query from the peptide and the key and value from the TCR lets an unfamiliar peptide be scored against receptor representations the model already holds, rather than requiring a learner of its own.
• A peptide prior learned by reconstruction: no large pretrained language model exists for peptides, so the cross-attention module is first trained as an asymmetric autoencoder reconstructing peptide sequences from the training pairs, which also initializes the cross-attention weights.
• Protein language model initialization: the receptor encoder starts from ESM-2 weights, carrying representations learned over large-scale protein data into the CDR3 sequences that dominate TCR diversity.
• Continuous affinity rather than a binary call: the sigmoid output is a score in [0, 1] fit with mean squared error, so predictions rank candidate receptors instead of only labeling them.

#Technical Details

Training runs in two stages. Stage 1 initializes the receptor module from ESM-2 and pretrains the cross-attention module as a sequence-reconstruction autoencoder over the peptide sequences of the training pairs, using Word2Vec residue embeddings, encoder-decoder weight sharing, a lower-triangular decoder mask, and a cross-entropy recovery loss. Stage 2 transfers both modules into the binding affinity pipeline and optimizes the framework end to end. Evaluation follows PanPep's splits — MajorSet, FewSet and ZeroSet — with binding receptors balanced against a control set of 60,333,379 non-binding TCRs. DapPep reaches 0.846 ROC-AUC and 0.864 PR-AUC in the majority setting, 0.784 and 0.814 few-shot, and 0.787 and 0.815 zero-shot; PanPep on the same splits reaches 0.792 and 0.796 in the majority setting and 0.708 and 0.715 zero-shot. On the curated ZeroSet of peptides absent from every tool's training data it reaches 0.816 ROC-AUC and 0.836 PR-AUC, against 0.744 and 0.755 for PanPep, 0.564 and 0.555 for pMTnet, 0.504 and 0.524 for ERGO2, and 0.483 and 0.481 for DLpTCR. Neither the parameter count nor the ESM-2 size variant is reported.

#Applications

The intended use is triage for neoantigen-directed immunotherapy, where the antigen is patient-specific and no binding data exist for it. On tumor-infiltrating lymphocytes from patients with metastatic gastrointestinal cancer, screened against 10 experimentally validated neoantigens, DapPep separated immunoreactive from non-reactive T cells at 0.835 ROC-AUC and 0.834 PR-AUC — the ranking step that decides which lymphocytes are worth expanding for adoptive cell transfer. The same checkpoint applies to exogenous and viral antigens in vaccine design and to bulk repertoire screening.

#Impact

DapPep's contribution is an argument about where peptide generality should live: a fixed peptide-agnostic checkpoint outperforms per-peptide adaptation on the very regimes meta-learning was introduced to handle, removing the fine-tuning step that made earlier tools awkward to deploy on one-off clinical antigens. The caveats are real. No code or weights have been released, so the results cannot be reproduced or the model run by others. All evaluation is in silico, and every benchmark balances binders against the same 60-million-sequence control set, so absolute AUC values depend on that negative-sampling choice. The model consumes sequence alone; the authors identify structural and gene expression features as the obvious next inputs, and note that no prospective wet-lab validation has been carried out.