TridentTCR

TridentTCR predicts which antigens a T-cell receptor (TCR) will recognise, and — unusually for this class of model — it also predicts whether the recognised antigen is a self-derived, autoimmune-related antigen (arAg). Most computational TCR tools answer a binary question: does this receptor bind this peptide? TridentTCR reframes the task as a trinary classification over general antigen binding, arAg binding, and non-binding, which is the distinction that actually matters when engineering TCR-T or CAR-T therapies. A TCR that cross-reacts with a self-peptide presented on healthy tissue can cause severe on-target, off-tumour toxicity, so a model that flags autoimmune risk directly is more clinically actionable than one that only scores binding.

The framework was developed by Rui Niu, Xiaoying Kong and Xuequn Shang at Northwestern Polytechnical University in Xi'an, with Jiren Zhou and Yanli Li at the John Curtin School of Medical Research, Australian National University. It was posted to bioRxiv in June 2025 and has not yet been peer reviewed.

TridentTCR pairs two complementary signals. Sequence identity comes from a frozen protein language model, ESM-2, which encodes antigen peptides and TCR beta CDR3 loops into high-dimensional embeddings. Relational context comes from a topology-aware graph neural network that propagates information over the observed antigen–TCR interaction network and over latent neighbourhoods built from embedding similarity. A geometric multi-view aggregation scheme fuses the two views before classification.

#Key Features

• Trinary specificity output: The model separates therapeutic antigen recognition from autoimmune-related antigen recognition rather than collapsing both into a single "binding" class, giving an explicit read-out of off-target autoimmune risk.
• Dual-view graph aggregation: One view propagates over the empirical interaction graph of known antigen–TCR pairs; a second propagates over latent neighbourhoods induced by embedding similarity, so receptors with no recorded partners still inherit relational context.
• False-positive-averse training: A hybrid loss combining cross-entropy with a focal Tversky term penalises false positives more heavily than false negatives, reflecting the asymmetric clinical cost of wrongly nominating a TCR for cell therapy.
• Cross-individual generalisation: Performance was assessed on three external cohorts — a 10x Genomics tetramer-stained single-cell dataset spanning four donors, 44 viral-antigen-specific TCRs from one donor, and 14 autoreactive TCRs from 12 type 1 diabetes donors.
• Antigenic immune response entropy: The accompanying AIRE metric combines predicted recognition probabilities with clonotype frequencies and repertoire diversity to quantify antigen-specific immune activity from single-cell immune profiling data.

#Technical Details

Antigen peptides and TCR beta CDR3 sequences are encoded with the 650M-parameter, 33-layer ESM-2 checkpoint, with per-residue representations pooled into fixed-length vectors. The benchmark set was assembled from IEDB, VDJdb and McPAS-TCR, yielding 41,732 high-confidence binding pairs after filtering, of which 11,054 involve autoimmune-related antigens and 30,678 involve general antigens; an equal number of negatives was generated within each fold by randomly mismatching TCRs and antigens.

Under five-fold cross-validation, the trinary model reaches a macro AUROC of 0.904 ± 0.006, a macro AUPRC of 0.854 ± 0.010, and accuracy of 0.743 ± 0.009, with one-vs-rest AUROCs of 0.893 ± 0.006 for general antigen binding and 0.979 ± 0.003 for arAg binding. A binary variant reaches AUROC 0.846 ± 0.002 and AUPRC 0.867 ± 0.002, exceeding the next-best method by 5.3% in AUROC, 6.6% in AUPRC and 3.3% in accuracy across NetTCR-2.0, pMTnet, HeteroTCR and single-modality ablations. Ablating the graph module costs 5.5% AUROC; replacing ESM-2 with BLOSUM50 encoding costs 8.9%, indicating the language model contributes more than the topology.

#Applications

The primary use case is safety triage in T-cell engineering: ranking candidate receptors from tumour-infiltrating lymphocyte repertoires by predicted antigen recognition while simultaneously screening them for predicted reactivity against autoimmune-related self-peptides. The AIRE metric extends the model to descriptive immunology, letting groups working with single-cell immune profiling quantify how strongly an individual's repertoire responds to a given antigen. In the paper's disease-focused analysis, TridentTCR predicted cross-reactivity between TCRs specific for the type 1 diabetes autoantigen ZnT8(186–194) and a mimotope from the gut commensal Bacteroides stercoris, a molecular-mimicry hypothesis for disease initiation.

#Impact

TridentTCR is the first published method to model autoimmune-related antigen recognition as an explicit class rather than treating all binding events alike, which is a meaningful reframing for immunotherapy safety assessment. Its ablations also add evidence to an ongoing question in the field: for TCR specificity, high-fidelity protein language model embeddings matter more than interaction-network topology, though combining both is better than either alone. Two limitations temper adoption. No code or trained weights accompany the preprint, so the results are not currently reproducible by third parties. And the underlying databases skew heavily toward a small number of HLA alleles and immunodominant epitopes, which bounds how far any model trained on them — including this one — can be expected to extrapolate to rare alleles and novel antigens.