Machine learning analysis of the T cell receptor repertoire identifies sequence features that predict self-reactivity

The T cell receptor (TCR) determines the specificity and affinity for both foreign and self-peptides presented by MHC. It is established that self-pMHC reactivity impacts T cell function, but it has been challenging to identify TCR sequence features that predict T cell fate. To discern patterns distinguishing TCRs from naïve CD4+ T cells with low versus high self-pMHC reactivity, we used data from 42 mice to train a machine learning (ML) algorithm that predicts self-reactivity directly from TCRß sequences. This approach revealed that n-nucleotide additions and acidic amino acids weaken selfreactivity. We tested our ML predictions of TCRß sequence self-reactivity using retrogenic mice. Extrapolating our analyses to independent datasets, we found high predicted self-reactivity for regulatory CD4+ T cells and low predicted self-reactivity for T cells responding to chronic infection. Our analyses suggest a potential trade-off between repertoire diversity and self-reactivity intrinsic to the architecture of a TCR repertoire. Overall design: We generated a dataset of 1.5x10^7 unique CDR3ß sequences from a total of 42 mice, investigating patterns among TCRß chain sequences between mature CD5lo and CD5hi naïve CD4+ T cells, as well as sequences in the double positive (DP, pre-selection) and single positive (SP, post-selection) stage in the thymus.