LYnet: Computational identification of tumor T cell antigens using convolutional and recurrent neural networks

Abstract

Background

Immunotherapy represents a paradigm shift in oncology, offering advantages in efficacy and specificity over traditional therapies. Key to its success is the identification of T-cell antigens, which are essential for triggering an effective antitumor immune response. Current methodologies for antigen prediction, however, lack the precision required for optimal vaccine development.

Purpose

This study aims to address this gap by introducing a novel deep learning model for the accurate prediction of tumor T-cell antigens. It seeks to improve the identification process, thereby facilitating the creation of more effective therapeutic cancer vaccines.

Methods

A hybrid architecture, designated LYnet, was constructed by integrating one-dimensional Convolutional Neural Networks with bidirectional Long Short-Term Memory layers, thereby capturing both local motif patterns and long-range sequence dependencies. Nineteen complementary sequence-derived descriptors—including AAindex, AAK-mer, CKSAAP/CKSAAGP, and physicochemical composition vectors—were concatenated to form the input feature space. Class imbalance in the training set was mitigated with the SMOTE-Tomek resampling strategy. Model robustness was quantified by stratified 10-fold cross-validation, and generalisation was verified on two independent benchmark collections (TAP 1.0 and iTTCA-RF).

Results

Across 10-fold validation on the LYnet benchmark, the proposed model achieved an AUC of 0.992, together with a sensitivity of 0.863, specificity of 0.925 and MCC of 0.776. Independent evaluation confirmed the advantage: LYnet yielded AUCs of 0.949 on the TAP 1.0 set and 0.836 on the iTTCA-RF set, surpassing the strongest competing method by 2.4–6.9 percentage points in AUC and up to 10.6 percentage points in MCC. These results demonstrate that LYnet attains state-of-the-art accuracy and balanced prediction for tumour T-cell antigen identification.

Conclusions

The introduction of this deep learning model represents a significant advancement in the prediction of tumor T-cell antigens. Its superior accuracy and robustness offer substantial potential to enhance the development and efficacy of cancer immunotherapies. This work not only underscores the importance of precise antigen identification in immunotherapy but also provides a powerful computational tool to aid in the design of next-generation cancer vaccines.