Machine learning analysis of the T cell receptor repertoire identifies features that impact thymic positive selection and self-reactivity. Mus musculus

The T cell receptor (TCR) sequence determines the specific peptides a given T cell recognizes. TCR-antigen interactions determine cell fates during thymic development and immune responses, but it remains unclear to what extent these fates are pre-disposed by the TCR sequence itself. We used machine learning (ML) to identify murine CDR3beta sequence features that impact how strongly naive CD4+ T cells recognize self peptides. Our ML model revealed that non-templated nucleotide additions and acidic amino acids weaken self-reactivity, reducing the likelihood of positive selection, whereas V12 gene usage and hydrophobic amino acids strengthen self-reactivity. We corroborated these findings by testing self-reactivity predictions using TCR retrogenic experiments. Our ML analysis also robustly showed differences between conventional and regulatory T cells, as well as cells reacting to acute versus chronic viral infection. Our results highlight how CDR3beta sequence features give rise to a trade-off between repertoire diversity and recognition strength.