BiLSTM with novel feature matrix predicts the binding affinity between MHC-I and peptides

Major histocompatibility complex I (MHC-I) molecules have the characteristic of de-homogeneity which can be recognized by CD8+T cells and plays an important role in antigen delivery. The loss or downregulation of the MHC-I may prevent the immune-mediated reaction. Some tumors may use the mechanism to become invisible in the immune system, causing MHC-I-based immunotherapy to be essential for cancer treatment. To raise the accuracy of targeted immunotherapy, precise and robust prediction of MHC-I affinity can be indispensable. Although traditional experimental methods focused on the comparison of mass spectrometry and flow cytometry to determine the sequence information, experimental technology was expensive and inefficient. Due to the weakness of second-generation sequencing, small-length sequences such as MHC-I may face a high risk of base deletion and mismatch. While machine learning and deep learning methods are mainly based on data mining and statistical inference, without costing too much on biological experiments. In addition, these models have robust generalization ability and accurate performances in several classification and regression tasks, offering confidential results of MHC affinity prediction. In this thesis, we considered 25 biological and sequence features including amino acid (AA) residues, hybrid orbital theory, and 3D physical structure. These computed properties in higher dimensions were further deduced by the specific feature selection algorithm. The final feature maps were integrated into a features matrix (9 x 25), regarded as a large labeled dataset. 222,875 samples were divided into training (178,300 samples) and testing datasets (44,575 samples). With this high-quantity feature matrix, we build several machine learning and deep learning models. Using MSE as the assessment, Bi-LSTM gained the best performance in both training dataset and testing dataset (Training_MSE: 0.0517; Testing_MSE: 0.0594) which was better than the SVM (Training_MSE: 0.0759; Testing_MSE: 0.0757) and DNN (Training_MSE: 0.0736; Testing_MSE: 0.0735). Our results mentions that the Bi-LSTM model may have a robust and stable ability in MHC-I affinity prediction, which can improve the study of bioinformatic study and increase the development of sequence prediction.