Research on Prediction of Pan-Cancer Immune Checkpoint Blockade Based on Multi-Head Self-Attention Classification Model

The development of immune checkpoint inhibition has changed the way we treat various cancers, but there is a large degree of variability in treatment response and many patients can not benefit clinically. There is currently a lack of pan-cancer multi-omic research on the efficacy of immune therapy, and in order to determine predictive biomarkers for immune checkpoint inhibition, we need to uncover transcriptional features associated with response to pan-cancer immune checkpoint inhibitors and decode the mechanisms of immune therapy. In this study, we propose an innovative algorithm to predict the efficacy of immune checkpoint inhibition, we first preprocess the transcriptomic features using COMBAT and SMOTE algorithm, and then use a Multi-Head Self-Attention model for binary classification to predict the sample's response to immune therapy. Finally, we use LIME algorithm for model explanation. Our method is validated on dataset, with an accuracy of 88%, outperforming other excellent models. The results showed that our method can effectively predict the sample's response to immune therapy.