Enhancing tumor T-cell antigen prediction using cross-attention-based multimodal representations with hybrid CNN-GLU model

Tumor T-cell antigens play a significant role in cancer immunotherapy by enabling the immune system to recognize and eliminate malignant cells. However, existing experimental methods for identifying these antigens are costly, time-consuming, and labor-intensive, highlighting the need for an accurate computational prediction model. In this study, we propose iTTCA-DeepGLU, a deep-learning model for predicting tumor T-cell antigens from peptide sequences. Each peptide sequence was initially transformed into an SMR-based two-dimensional matrix, after which local phase quantization was applied to generate novel intrinsic local descriptors, termed SMR-LPQ features. In parallel, semantic and contextual features were extracted using the pretrained ProtT5 protein language model. The SMR-LPQ and ProtT5 features were then integrated through cross-attention-based fusion to construct a multimodal feature representation. Subsequently, a gradient-boosted decision tree was applied to retain the most informative features. Among the applied classifiers, our hybrid CNN-GLU model demonstrated high predictive performance. The proposed iTTCA-DeepGLU model using Train-DS1 and Train-DS2 datasets achieved accuracies of 95.87% and 96.72%, representing improvements of 8% and 11%, respectively, with corresponding AUC values of 0.97 and 0.98. To evaluate the model's generalizability, it was further assessed using two independent test sets. On the imbalanced Ind-DS1 dataset, the iTTCA-DeepGLU model achieved an accuracy of 90.39%, representing a 14% improvement, with an AUC of 0.94. Similarly, on the balanced Ind-DS2 dataset. iTTCA-DeepGLU attained an accuracy of 95.30%, approximately 2% higher than that of existing tumor T-cell antigen prediction models, with an AUC of 0.92. Overall, iTTCA-DeepGLU demonstrated strong predictive performance and generalization capability, highlighting its potential as a reliable computational tool for identifying tumor T-cell antigens and supporting cancer immunotherapy research.