Tumor-specific neoantigens are mutated self-peptides presented by tumor cell major histocompatibility complex (MHC) molecules and are necessary to elicit host’s anti-cancer cytotoxic T cell responses. It could be specifically recognized by neoantigen-specific T cell receptors (TCRs). However, current wet-lab assays for identifying peptide MHC binding are too expensive and time-consuming to meet the clinical needs. In this study, we developed an in silico method with a deep convolutional neural network (CNN) model, iConMHC, to predict peptide MHC binding affinity. Unlike other in silico methods that only learn from properties of amino acid in neoantigen peptides alone and/or MHCs alone, iConMHC learns from physical and chemical interaction properties between pairwise amino acids from the two molecules. These properties, such as contact potentials and distances in folded proteins, directly affect neoantigen-MHC binding affinity. In addition, IConMHC is a pan-allele model that is capable of making predictions for all the MHC alleles. Even for those rare MHC alleles without training data, iConMHC can make predictions with reasonable accuracy. We benchmarked iConMHC with other commonly used MHC-I binding predictors and found our model performs better than most of the pan-allele models.

肿瘤特异性新抗原是由肿瘤细胞主要组织相容性复合物（MHC）分子呈递的突变自身肽，对于引发宿主的抗癌细胞毒性T细胞应答是必需的。它可以被新抗原特异性T细胞受体（TCR）特异性识别。然而，当前用于鉴定肽MHC结合的湿实验室测定法过于昂贵且费时，无法满足临床需求。在这项研究中，我们开发了一种具有深度卷积神经网络（CNN）模型iConMHC的计算机模拟方法，以预测肽MHC结合亲和力。与其他仅从单独的新抗原肽和/或单独的MHC中学习氨基酸特性的计算机方法不同，iConMHC从两个分子的成对氨基酸之间的物理和化学相互作用特性学习。这些属性 例如折叠蛋白中的接触电位和距离，直接影响新抗原与MHC的结合亲和力。此外，IConMHC是一个泛等位基因模型，能够对所有MHC等位基因进行预测。即使对于那些没有训练数据的罕见MHC等位基因，iConMHC也可以以合理的准确性进行预测。我们将iConMHC与其他常用的MHC-1结合预测因子进行了基准比较，发现我们的模型比大多数泛等位基因模型表现更好。