Deep learning model for immune repertoire-based cancer prediction

Abstract

T-cells and B-cells are the major components regarding tumor-specific immune activation. Complementarity determining region 3 (CDR3) immune receptor repertoires constructed from activated immune cells represent the unique cancer status in individuals. However, an approach based on immunological characteristics has yet to be widely used in the diagnosis of multiple cancers, especially in the liquid biopsy field. Hence, we apply a deep neural network model reflecting 124 cancer-specific immunological features from more than 5,000 tumors and 4,000 normal tissue-derived transcriptomes. The model successfully distinguishes 530 blood samples of various cancer types from 701 normal blood samples resulting in a ROC-AUC of 0.93. Interpretation of the model unveils the significance of features related to the B-cell receptor repertoire. Our research highlights the application of immune-derived features in a noninvasive, blood-based multiple cancer prediction.