Enhancing TCR-Peptide Interaction Prediction with Pretrained Language Models and Molecular Representations

TL;DR

This work presents LANTERN (Large lAnguage model-powered TCR-Enhanced Recognition Network), a deep learning framework that combines large-scale protein language models with chemical representations of peptides to advance TCR-pMHC binding prediction and support the development of personalized immunotherapies.

Abstract

Understanding the binding specificity between T-cell receptors (TCRs) and peptide-major histocompatibility complexes (pMHCs) is central to immunotherapy and vaccine development. However, current predictive models struggle with generalization, especially in data-scarce settings and when faced with novel epitopes. We present LANTERN (Large lAnguage model-powered TCR-Enhanced Recognition Network), a deep learning framework that combines large-scale protein language models with chemical representations of peptides. By encoding TCR \b{eta}-chain sequences using ESM-1b and transforming peptide sequences into SMILES strings processed by MolFormer, LANTERN captures rich biological and chemical features critical for TCR-peptide recognition. Through extensive benchmarking against existing models such as ChemBERTa, TITAN, and NetTCR, LANTERN demonstrates superior performance, particularly in zero-shot and few-shot learning scenarios. Our model also benefits from a robust negative sampling strategy and shows significant clustering improvements via embedding analysis. These results highlight the potential of LANTERN to advance TCR-pMHC binding prediction and support the development of personalized immunotherapies.

Figures (5)

• Figure 1: Framework of the LANTERN Model. LANTERN encodes TCR $\beta$ chains using ESM and peptide sequences via SMILES and MolFormer. A Multi-Head Cross-Attention module aligns TCR and peptide embeddings before final classification. The model supports both random and reference-based negative sampling, and generalizes well to unseen epitopes.
• Figure 2: Performance of LANTERN on four datasets. |a. AUC-ROC scores across models for NA dataset. b. AUC-ROC scores across models for RN dataset. c. Accuracy scores across models for the four datasets.
• Figure 3: t-SNE visualization of the embeddings from LANTERN model. |a. With pretrain of 0.857 h score. b. Without pretrain of 0.607 h score. c. Show frequency of the pretrained embeddings. d. Show frequency of the no pretrained embeddings.
• Figure 4: Comparison of cluster metrics between embeddings with or without pretraining. The SI and CHI scores are scaled by sigmoid function to make them more readable.
• Figure 5: Few-Shot Learning performance of LANTERN | a. Performance comparison of models over different numbers of seen pairs. b. Performance comparison of models over different seen ratios.