Transformer-Based Models for pHLA Binding Affinity Prediction

The prediction of peptide- HLA binding affinity is a crucial step in cancer immunotherapy based on neoantigens. Currently, machine learning algorithms for predicting pHLA binding affinity primarily rely on neural networks, particularly neural network models based on convolutional neural networks (CNNs). However, in the peptides capable of binding to class I HLA molecules, the majority of sequences have a length of 9. Due to the difficulty in handling variable-length inputs, these models mostly consider fixed lengths, specifically predicting the binding affinity of 9-mer peptides. This paper discusses a model based on Transformers that can handle variable-length inputs. Experimental results demonstrate that the Transformer-based model exhibits excellent performance in predicting pHLA binding affinity.