DeepHLApan: A Deep Learning Approach for Neoantigen Prediction Considering Both HLA-Peptide Binding and Immunogenicity

Abstract

Neoantigens play important roles in cancer immunotherapy. Current methods used for neoantigen prediction focus on the binding between human leukocyte antigens (HLAs) and peptides, which is insufficient for high-confidence neoantigen prediction. In this study, we apply deep learning techniques to predict neoantigens considering both the possibility of HLA-peptide binding (binding model) and the potential immunogenicity (immunogenicity model) of the peptide-HLA complex (pHLA). The binding model achieves comparable performance with other well-acknowledged tools on the latest Immune Epitope Database (IEDB) benchmark datasets and an independent mass spectrometry (MS) dataset. The immunogenicity model could significantly improve the prediction precision of neoantigens. The further application of our method to the mutations with pre-existing T-cell responses indicating its feasibility in clinical application. DeepHLApan is freely available at https://github.com/jiujiezz/deephlapan and http://biopharm.zju.edu.cn/deephlapan.

Keywords: cancer immunology; deep learning; human leukocyte antigen; neoantigen; recurrent neural network.

Figure 1

The architecture of DeepHLApan. Two types of data (437,077 binding data points, 32,785 immunogenicity data points) are collected for model training, and one-hot encoding is used for amino acid representation. Three layers of bidirectional GRU with attention have been employed as the model framework. The immunogenic score is used as a filter (>0.5), and peptides with binding scores ranked within the top 20 are predicted as high-confidence neoantigens.

Figure 2

Binding model performance on never-seen HLA alleles. (A) Prediction AUC on alleles with both positive HLA-peptide pairs and negative pairs in descending order. (B) Prediction ACC on alleles with single-labeled HLA-peptide pairs in descending order.

Figure 3

The comparison of actual motifs and predicted motifs on 16 HLA alleles. The motif logo is created by Weblogo. The actual motifs are based on their binding peptides, the predicted motifs are generated by taking top 1% predicted peptides out of 100,000 random peptides.

Figure 4

Model comparison between the binding model of DeepHLApan with other tools. (A) Performance of the binding model compared with the other 12 well-acknowledged tools on the latest IEDB benchmark datasets. (B) Performance of the binding model compared with the other 5 binding tools on the independent MS dataset. The detailed information of each sub-dataset is listed in Tables S7A,B.

Figure 5

For 26 mutations with pre-existing T-cell responses, we ranked them in order of probability of presentation within their corresponding patients. The mutation rank of NetMHCpan 4.0 was measured by taking the minimum predicted rank across all mutation-spanning peptides. The number of predicted mutations ranked in the top 5, 10, and 20 by EDGE and MHCflurry were derived from Bulik-Sullivan et al. (27).