LANTERN

Predicting which T-cell receptor binds which peptide-MHC complex is a bottleneck for neoantigen vaccines, TCR-T therapy, and repertoire interpretation, and it is hard for a specific reason: labelled binding data is scarce and heavily skewed toward a handful of well-studied epitopes, so models that fit the training epitopes tend to collapse on new ones. LANTERN — Large lAnguage model-powered TCR-Enhanced Recognition Network — attacks the generalisation problem by borrowing representations from two pretrained models in different domains rather than learning them from the small labelled corpus.

The asymmetry it exploits is that TCRs and epitopes are not the same kind of object. A TCR beta chain is a protein, encoded with ESM-1b. An epitope is a peptide of typically 8 to 11 residues — too short for a sequence model to build much context around — so LANTERN converts it into a SMILES string and encodes it with MoLFormer, a chemical language model. The SMILES form is longer than the amino-acid string and makes molecular connectivity and functional groups explicit, turning a data-poor sequence problem into one a pretrained chemical model already has priors for.

LANTERN was developed at the New Jersey Institute of Technology, posted as a preprint in April 2025 and published in PeerJ in March 2026. The peer-reviewed version adds the cross-attention fusion module that defines the architecture.

#Key Features

• Two pretrained backbones, one task: A protein language model handles the receptor and a chemical language model the epitope, so both modalities arrive with representations learned from far more data than the binding labels provide.
• SMILES encoding of epitopes: Converting short peptides to SMILES yields longer, chemically explicit sequences; randomising the SMILES strings while preserving their distribution destroys most of the benefit, confirming the gain comes from real chemical structure.
• Multi-head cross-attention fusion: Rather than concatenating the two embeddings, the receptor representation queries the peptide representation, letting the model attend selectively to interaction-relevant chemical features.
• Alignment regularisation: An auxiliary mean-squared-error term keeps the fused representation close to the original receptor embedding, stabilising the TCR-centric view while allowing peptide-specific modulation.
• Hard epitope splits: Peptide-level grouped splits keep test epitopes out of training and validation entirely — the zero-shot condition that matters clinically.

#Technical Details

Both encoders emit 256-dimensional representations. The cross-attention module takes the receptor embedding as query and the peptide embedding as key and value; its output is concatenated with the original receptor embedding and passed to an MLP with a sigmoid output. Training uses binary cross-entropy plus the weighted alignment term, with the Adam optimizer and early stopping on validation AUC. The cross-attention module and MLP are trained from scratch; the pretrained encoders are held frozen or fine-tuned depending on the experimental setting.

Data comes from the TCHard benchmark, comprising roughly 160,000 unique TCR beta chains and 1,341 unique peptides. Only the beta-chain CDR3 and the peptide are used; the alpha chain, MHC allele, and V/J gene annotations are excluded, the last because they are present for only about 4.8% of samples. Four variants are evaluated, crossing the original reference-control (NA) and random-control (RN) negatives with regenerated versions of each, under five-fold cross-validation with three seeds per fold. LANTERN reaches ROC-AUC 0.903 on original NA and 0.883 on generated NA, versus 0.663 on generated RN and 0.571 on original RN. It does not win uniformly: TEINet scores higher on both reference-control settings, while LANTERN leads on generated random control, where most baselines fall toward chance. Removing the SMILES branch costs 0.018 to 0.035 ROC-AUC across settings.

#Applications

The model is aimed at immunologists and immuno-oncology groups screening candidate epitopes against a receptor repertoire, or ranking receptors against a target neoantigen, when the epitope of interest has no binding data behind it. Because the epitope is handled as a small molecule, the framework also transfers to peptide analogues and chemically modified epitopes that a purely sequence-based encoder cannot represent. A Hopkins score of 0.90 for the pretrained representation, against a much weaker score without pretraining, supports using it for repertoire clustering as well as pairwise scoring.

#Impact

LANTERN's contribution is methodological: it shows that cross-modal fusion of a protein language model with a chemical language model beats naive concatenation for TCR-epitope prediction, and that the gain is largest where the field struggles, on random-control negatives and unseen epitopes. The honest reading of the benchmark table is that no method is yet reliable under random-control sampling, LANTERN included, with its 0.571 ROC-AUC on the original random-control split close to chance. Release is thin: the paper's availability statement points to an anonymised review repository that is no longer accessible, and the maintained repository holds training, preprocessing and embedding-generation notebooks along with result tables, with data behind a Google Drive link, but no stored embeddings, no trained checkpoint and no license file.