DUET

Two computational problems sit next to each other in cancer immunotherapy pipelines and are almost always solved apart. The first is prioritizing which T-cell receptor (TCR)-epitope pairs are worth validating experimentally. The second is deciding which T cells in a tumor biopsy are actually tumor-reactive, rather than bystanders. Binding predictors treat each TCR-epitope pair as an independent observation and ignore the fact that similar receptors tend to share binding preferences; tumor-reactivity workflows lean almost entirely on transcriptomic signature scoring and barely use receptor sequence at all. DUET closes that gap by putting both tasks inside one heterogeneous graph.

DUET was developed at Algoma University in Brampton, Ontario, with a collaborator at the University of Guelph, and posted to bioRxiv in May 2026. The first version of the preprint carried the name TRACE (TCR-epitope pRioritization And T-Cell idEntification); the August 2026 revision renamed the same method DUET, for Dual Unified Evaluation of TCR-Epitopes and Tumor-reactive T cells.

Where pair-scoring models such as TULIP, LANTERN, and DeepPROTECTNeo embed one receptor against one peptide, DUET represents TCRs, epitopes, and individual T cells as typed nodes in a single graph. Similarity edges connect entities of the same type, association edges connect across types, and attention propagates information along both.

#Key Features

• One architecture, two prediction targets: TCR-epitope prioritization is edge-level prediction and tumor-reactivity is node-level prediction, but only the input node types and the final head change between them.
• Repertoire context through similarity edges: K-nearest-neighbor edges among TCRs and among epitopes are built from k-mer composition and physicochemical similarity, so receptors that resemble each other exchange information during message passing.
• Spectral structure encoding: Laplacian positional encoding gives each node a topological fingerprint, and ablation shows it is the single most consequential structural component in sparse graphs.
• Paired chain handling: alpha and beta CDR3 sequences are embedded separately and concatenated, with the beta chain weighted at 0.7 in the similarity computation to reflect its larger contribution to antigen specificity.
• Cell state joined to receptor sequence: in the single-cell setting, cells enter as nodes carrying differential-expression features and connect to the clonotype they express through a membership edge.

#Technical Details

TCR CDR3 sequences are encoded with BertTCR and epitope sequences with ESM-2; cell nodes use expression features from genes found differentially expressed between reactive and non-reactive cells after Seurat v5 preprocessing. A structure-injection step adds an 8-dimensional Laplacian positional encoding and an edge-weight bias before two edge-aware graph transformer layers (256 hidden channels, two attention heads, dropout 0.3, learning rate 5e-4, K = 20 for TCRs and 15 for epitopes), implemented in PyTorch Geometric. Benchmarks cover human MHC class I records: 8,187 IEDB pairs across 7,682 TCRs and 980 epitopes, 71,686 VDJdb pairs across 1,352 epitopes, and a gastrointestinal cancer single-cell cohort of 376 cells spanning 84 clonotypes.

Across five repeated runs of 10-fold cross-validation, DUET reached AUROC/AUPR of 0.937/0.922 on IEDB and 0.992/0.990 on VDJdb, ahead of NetTCR-2.0 (0.870/0.892) and HeteroTCR (0.814/0.843). On the single-cell cohort it reached 0.984/0.984 against a best signature-based baseline of 0.735/0.310. Under TCR-disjoint cross-validation, where test receptors are withheld from training entirely, performance held at 0.888/0.906 on IEDB and 0.983/0.986 on VDJdb, declining to 0.893/0.733 on the small single-cell cohort. Negative pairs are sampled from unobserved TCR-epitope combinations rather than measured non-binders.

#Applications

The workflow targets immunotherapy discovery: shortlisting receptor-antigen pairs before tetramer staining or functional assays, and annotating tumor-infiltrating lymphocytes from paired single-cell TCR and transcriptome data so that reactive clonotypes can be selected for expansion. That covers adoptive cell transfer, TCR-engineered T-cell therapy, and neoantigen vaccine programs, where the bottleneck is deciding which of thousands of observed clonotypes deserves a wet-lab experiment.

#Impact

DUET argues that receptor recognition is better modeled relationally than as isolated sequence pairs, and its precision-recall margins on the sparser IEDB graph support that claim where class imbalance is realistic. It complements generative approaches such as TCRDiff and EpitopeGen, which propose candidates that a scorer like this one can rank. The work remains a preprint awaiting peer review, and its constraints are real: the tumor-reactivity evaluation rests on a single 376-cell cohort, MHC identity is not modeled explicitly, and the authors state that code and data will be released on GitHub upon acceptance, so no implementation or trained weights are available yet. The preprint is licensed CC BY-NC.