LoFT-TCR

T cells recognise diseased cells by matching their receptors against short peptides presented on MHC class I molecules, and knowing which T-cell receptor (TCR) binds which antigen underpins neoantigen vaccine design and TCR-T cell therapy. Tetramer assays resolve those pairings experimentally but are slow and reagent-limited. Computational prediction runs into the scale of the repertoire: a healthy adult carries five to six million distinct TCRs, and even siblings share under one percent of them.

LoFT-TCR attacks that generalisation problem by adapting a protein language model rather than working around one. Most supervised TCR-antigen predictors encode sequences with shallow, hand-designed schemes such as BLOSUM50 matrices feeding a CNN, which capture common motifs but struggle on rare receptor fragments. Language model embeddings are richer, but a frozen general-purpose model is not tuned to immune sequences and full fine-tuning on a dataset this small invites overfitting. LoFT-TCR takes the middle path: low-rank adaptation of ESM-2, producing a TCR-specialised encoder whose backbone weights never move.

The framework was developed by Rui Niu, Xiaoying Kong and Xuequn Shang at Northwestern Polytechnical University in Xi'an and posted to bioRxiv in September 2025; it has not been peer reviewed. The same group earlier released TridentTCR, which pairs a frozen ESM-2 encoder with a topology-aware graph over a benchmark of the same scale. LoFT-TCR keeps its trinary framing — separating general antigens from autoimmune-related antigens (arAgs) — but swaps that frozen encoder for an adapted one.

#Key Features

• Low-rank adaptation of ESM-2: LoRA matrices of rank 4 and scaling factor 16 are injected into the query, key, value and output projections of all 33 transformer layers, leaving 3.05 million trainable parameters against a 650M backbone.
• Self-supervised specialisation: Adaptation uses a masked language modelling objective over 37,054 antigen-specific TCR beta CDR3 sequences, with no binding labels; masked-token accuracy rises from 0.176 to 0.598 in one epoch.
• Asymmetric graph encoding: TCR nodes are embedded with the adapted model while antigen nodes use the original pretrained ESM-2, combining task-specific and general representations in one graph.
• Cross-reactivity as topology: One antigen binds many receptors and vice versa, so interactions are modelled as a heterogeneous bipartite graph whose message passing supplies relational context sequence alone cannot.
• Autoimmune-risk head: A trinary classifier separates non-binding, general antigen recognition and arAg recognition, surfacing the off-target toxicity risk binary predictors leave invisible.

#Technical Details

The benchmark was assembled from IEDB, McPAS-TCR and VDJdb, restricted to human TCR beta CDR3 sequences of 10-20 residues paired with HLA class I peptides of 8-15 residues, deduplicated, and filtered with iSMART to drop receptors that cluster with nothing else. That yields 41,732 high-confidence binding pairs covering 972 distinct peptides and 37,054 unique CDR3 sequences, plus 786 curated arAgs; negatives are generated per fold by randomly mispairing antigens with receptors. After graph propagation, TCR and antigen representations are concatenated and passed through a two-layer MLP of 512 and 256 units; adaptation ran for a single epoch on NVIDIA A800 GPUs.

Under five-fold cross-validation on binding prediction, LoFT-TCR reaches AUROC 0.859 ± 0.002, AUPRC 0.877 ± 0.002 and accuracy 0.758 ± 0.004, against 0.793 / 0.801 / 0.709 for HeteroTCR, 0.758 / 0.763 / 0.680 for NetTCR-2.0, and 0.615 AUROC for pMTnet. On trinary specificity discrimination it reaches macro AUROC 0.912 ± 0.004, macro AUPRC 0.869 ± 0.007 and macro accuracy 0.757 ± 0.008. Ablations attribute a 13.8% AUROC drop to replacing the ESM-2 encoder with BLOSUM50 encoding and a 10.3% drop to removing the graph module — both components contribute, the encoder more.

#Applications

The direct use case is prioritising receptor-antigen pairs before tetramer validation: given a repertoire from tumour-infiltrating lymphocytes or a vaccine trial cohort, the model scores which receptors are likely to engage a peptide of interest, concentrating scarce experimental capacity on the best candidates. The trinary head extends this to safety triage, flagging receptors predicted to recognise autoimmune-related self-peptides — the failure mode behind on-target, off-tumour toxicity in engineered T-cell therapy. The adaptation recipe is reusable on its own, giving a drop-in feature extractor for other immune-sequence tasks.

#Impact

LoFT-TCR's most transferable result is methodological: it quantifies what parameter-efficient adaptation buys over frozen protein language model embeddings when labelled biological data is scarce. Several constraints bound the claims. The model uses only TCR beta CDR3 sequences, with alpha chain and V(D)J gene usage left to future work. Negatives are synthetic rather than confirmed non-binders, a limitation shared with most models trained on these databases, and all evaluation is five-fold cross-validation on one curated benchmark with no external cohort or unseen-antigen split reported. Training and evaluation code plus the curated splits are on GitHub, but the repository carries no licence file and no adapted checkpoint, so reuse terms are undefined and the LoRA weights must be reproduced from scratch.