MATHLA: a robust framework for HLA-peptide binding prediction integrating bidirectional LSTM and multiple head attention mechanism.

Background Accurate prediction of binding between class I human leukocyte antigen (HLA) and neoepitope is critical for target identification within personalized T-cell based immunotherapy. Many recent prediction tools developed upon the deep learning algorithms and mass spectrometry data have indeed showed improvement on the average predicting power for class I HLA-peptide interaction. However, their prediction performances show great variability over individual HLA alleles and peptides with different lengths, which is particularly the case for HLA-C alleles due to the limited amount of experimental data. To meet the increasing demand for attaining the most accurate HLA-peptide binding prediction for individual patient in the real-world clinical studies, more advanced deep learning framework with higher prediction accuracy for HLA-C alleles and longer peptides is highly desirable. Results We present a pan-allele HLA-peptide binding prediction framework-MATHLA which integrates bi-directional long short-term memory network and multiple head attention mechanism. This model achieves better prediction accuracy in both fivefold cross-validation test and independent test dataset. In addition, this model is superior over existing tools regarding to the prediction accuracy for longer ligand ranging from 11 to 15 amino acids. Moreover, our model also shows a significant improvement for HLA-C-peptide-binding prediction. By investigating multiple-head attention weight scores, we depicted possible interaction patterns between three HLA I supergroups and their cognate peptides. Conclusion Our method demonstrates the necessity of further development of deep learning algorithm in improving and interpreting HLA-peptide binding prediction in parallel to increasing the amount of high-quality HLA ligandome data.