Abstract
AbstractClassifiers trained on adaptive immune receptors (AIRs) have been reported to accurately identify a broad range of diseases. However, when using the conventional “clonotype” representation of AIRs, different donors exhibit vastly different features, limiting the generalizability of these classifiers. To address this problem, we developed a representation of AIRs based on paratope adjacency that significantly improved donor sharing and classification performance on new donors across a wide range of diseases (mean ROC AUC 0.893) compared to classifiers trained on clonotypes (0.714). Surprisingly, in cancer data, we observed that some of the AIRs that were important for the classification were significantly more abundant in healthy controls than in patients. Such healthy-biased, but not disease-biased, AIRs were predicted to target known cancer-associated antigens at dramatically higher rates than background values (Z-scores > 75), constituting an overlooked reservoir of cancer-targeting immune cells that are diagnostic and identifiable from a routine blood test.
Publisher
Cold Spring Harbor Laboratory