Deep convolutional neural networks for peptide-MHC binding predictions

Abstract

Background: Determining peptides that bind specific MHC molecules can facilitate the development of peptide-based vaccines and design of immunotherapies. Recently, machine-learning-based methods have generated successful results by training large amounts of experimental data. However, many machine learning-based methods are generally less sensitive in recognizing locally-clustered interactions, which can synergistically stabilize peptide binding. Deep convolutional neural network (ConvNet) is a deep learning method inspired by visual recognition process of animal brain and it is known to be able to capture meaningful local patterns from 2D images. Once the peptide-MHC interactions can be encoded into image-like matrix(ILM) data, ConvNets can be employed to build a predictive model for peptide-MHC binding prediction. In this thesis, we demonstrate that ConvNet model are able to not only reliably predict peptide-MHC binding, but also sensitively capture locally-clustered interactions without the prior knowledge of binding modes.

Results: For MHC-I, the ConvNet model for pan-specific peptide-HLA-I binding predictions was trained using ILM data encoded from peptide-HLA-I experimental data and showed the reliable performance in nonapeptide binding predictions through the independent evaluation of IEDB external datasets which consist of 43 datasets for 15 HLA-A alleles and 25 datasets for 10 HLA-B alleles. In particular, the model outperformed other tools for alleles belonging to the HLA-A3 supertype. The F1 scores of the DCNN were 0.86, 0.94, and 0.67 for HLA-A*31:01, HLA-A*03:01, and HLA-A*68:01 alleles, respectively, which were significantly higher than those of other prediction tools. We developed ConvMHC, a web server(http://jumong.kaist.ac.kr:8080/convmhc) to provide user-friendly web interfaces for peptide-MHC class I binding predictions using the ConvNet model. For MHC-II, the ConvNet model for pan-specific peptide-HLA-II binding predictions was trained on ILMs encoded from experimental data for the binding of variable length peptides to MHC-II molecules. The nine ConvNet models were trained on the ILM datasets encoded using different amino acid encoding schemes. The ConvNet model showed the reliable prediction performance through the independent evaluation on external datasets covering 13 HLA-DR alleles. In particular, we report that the ConvNet model outperformed the NetMHCIIpan method in predicting the 18-mer peptide KKAGLVGVLAGLAFQEMD-binding to four different HLA-DR alleles, including HLA-DRB1*11:01, HLA-DRB1*13:01, HLA-DRB3*03:01, and HLA-DRB4*01:03.

Conclusions: We developed a novel method for pan-specific peptide-MHC binding prediction using the ConvNet prediction model trained on ILM data encoded from experiment data. We showed the reliable performance of the ConvNet models in predicting both the peptide-MHC-I bindings and peptide-MHC-II bindings. We anticipate that our ConvNet models will be significantly reliable in predicting peptide binding to MHC molecules through further evaluations on more experimental data. Moreover, our approaches described herein will be useful for recognizing locally-clustered interactions without the prior knowledge of binding modes in molecular binding structures, such as protein/DNA, protein/RNA, and drug/protein interactions.