Abstract
Abstract
Peptide-HLA (pHLA) is a peptide that may attach to human leukocyte antigen (HLA) and be presented to specialized immune cells, then initiate an immune response. Computational prediction of peptide and HLA binding is an important tool in studying T cell immunity that can assist in the design of neoantigen vaccines. However, the majority of current prediction approaches are limited to single allele (SA) HLA data, predictive tools to optimize mutant peptides with higher affinity for multi allelic (MA) HLA are lacking. Here, we describe ResMAHPan (https://github.com/Luckysoutheast/ResMAHpan.git), which integrates long short term memory (LSTM) network and Resnet network with coordinate attention (CA) for pHLA binding and presentation prediction. ResMAHPan considerably outperforms the standard predictors NetMHCpan 4.0 and MHCflurry 2.0 by enriching for current MA HLA presentation prediction algorithms on held-out mass spectrometry experiments. We propose a mode based on existing MA-pHLA encoding that allows incorporation of neoantigen prediction tasks into computer vision methods—which can aggregate MA HLA molecules into a multichannel matrix and incorporated peptide sequences to capture binding signals efficiently. Finally, the integrated model could be employed as a independent neoantigen recognition approach to improve neoantigen identification accuracy, or in conjunction with other methods to achieve the maximum level of accuracy.
Publisher
Research Square Platform LLC