Our immune system can destroy most cells in our body, an ability that needs to be tightly controlled. To prevent autoimmunity, the thymic medulla exposes developing T cells to normal "self" peptides and prevents any responders from entering the bloodstream. However, a substantial number of self-reactive T cells nevertheless reaches the periphery, implying that T cells do not encounter all self peptides during this negative selection process. It is unclear if T cells can still discriminate foreign peptides from self peptides they haven't encountered during negative selection. We use an "artificial immune system"-a machine learning model of the T cell repertoire-to investigate how negative selection could alter the recognition of self peptides that are absent from the thymus. Our model reveals a surprising new role for T cell cross-reactivity in this context: moderate T cell cross-reactivity should skew the post-selection repertoire towards peptides that differ systematically from self. Moreover, even some self-like foreign peptides can be distinguished provided that the peptides presented in the thymus are not too similar to each other. Thus, our model predicts that negative selection on a well-chosen subset of self peptides would generate a repertoire that tolerates even "unseen" self peptides better than foreign peptides. This effect would resemble a "generalization" process as it is found in learning systems. We discuss potential experimental approaches to test our theory.
Keywords: T cell repertoires; artificial immune system; central tolerance; learning by example; negative selection; self-nonself discrimination.