Xu Ying, Qian Xinyang, Tong Yao, Li Fan, Wang Ke, Zhang Xuanping, Liu Tao, Wang Jiayin
Department of Computer Science and Technology, School of Electronic and Information Engineering, Xi'an Jiaotong University, Xi'an, China.
Geneplus Beijing Institute, Beijing, China.
Front Genet. 2022 Aug 22;13:942491. doi: 10.3389/fgene.2022.942491. eCollection 2022.
T-cell receptors (TCRs) are formed by random recombination of genomic precursor elements, some of which mediate the recognition of cancer-associated antigens. Due to the complicated process of T-cell immune response and limited biological empirical evidence, the practical strategy for identifying TCRs and their recognized peptides is the computational prediction from population and/or individual TCR repertoires. In recent years, several machine/deep learning-based approaches have been proposed for TCR-peptide binding prediction. However, the predictive performances of these methods can be further improved by overcoming several significant flaws in neural network design. The interrelationship between amino acids in TCRs is critical for TCR antigen recognition, which was not properly considered by the existing methods. They also did not pay more attention to the amino acids that play a significant role in antigen-binding specificity. Moreover, complex networks tended to increase the risk of overfitting and computational costs. In this study, we developed a dual-input deep learning framework, named AttnTAP, to improve the TCR-peptide binding prediction. It used the bi-directional long short-term memory model for robust feature extraction of TCR sequences, which considered the interrelationships between amino acids and their precursors and postcursors. We also introduced the attention mechanism to give amino acids different weights and pay more attention to the contributing ones. In addition, we used the multilayer perceptron model instead of complex networks to extract peptide features to reduce overfitting and computational costs. AttnTAP achieved high areas under the curves (AUCs) in TCR-peptide binding prediction on both balanced and unbalanced datasets (higher than 0.838 on McPAS-TCR and 0.908 on VDJdb). Furthermore, it had the highest average AUCs in TPP-I and TPP-II tasks compared with the other five popular models (TPP-I: 0.84 on McPAS-TCR and 0.894 on VDJdb; TPP-II: 0.837 on McPAS-TCR and 0.893 on VDJdb). In conclusion, AttnTAP is a reasonable and practical framework for predicting TCR-peptide binding, which can accelerate identifying neoantigens and activated T cells for immunotherapy to meet urgent clinical needs.
T细胞受体（TCR）由基因组前体元件随机重组形成，其中一些介导对癌症相关抗原的识别。由于T细胞免疫反应过程复杂且生物学实证证据有限，识别TCR及其识别肽的实际策略是从群体和/或个体TCR库进行计算预测。近年来，已经提出了几种基于机器学习/深度学习的方法用于TCR-肽结合预测。然而，通过克服神经网络设计中的几个重大缺陷，这些方法的预测性能可以进一步提高。TCR中氨基酸之间的相互关系对于TCR抗原识别至关重要，而现有方法并未对此进行适当考虑。它们也没有更多地关注在抗原结合特异性中起重要作用的氨基酸。此外，复杂的网络往往会增加过拟合风险和计算成本。在本研究中，我们开发了一种名为AttnTAP的双输入深度学习框架，以改进TCR-肽结合预测。它使用双向长短期记忆模型对TCR序列进行稳健的特征提取，该模型考虑了氨基酸与其前体和后体之间的相互关系。我们还引入了注意力机制，赋予氨基酸不同的权重，并更多地关注起作用的氨基酸。此外，我们使用多层感知器模型而非复杂网络来提取肽特征，以减少过拟合和计算成本。AttnTAP在平衡和不平衡数据集上的TCR-肽结合预测中均实现了高曲线下面积（AUC）（在McPAS-TCR上高于0.838，在VDJdb上高于0.908）。此外，与其他五个流行模型相比，它在TPP-I和TPP-II任务中的平均AUC最高（TPP-I：在McPAS-TCR上为0.84，在VDJdb上为0.894；TPP-II：在McPAS-TCR上为0.837，在VDJdb上为0.893）。总之，AttnTAP是一种合理且实用的预测TCR-肽结合的框架，可加速识别新抗原和活化T细胞用于免疫治疗，以满足紧迫的临床需求。
