Hesham ElAbd, Aya K.H. Mahdy, Eike Matthias Wacker, Maria Gretsova, David Ellinghaus, Astrid Dempfile, Andre Franke
In the Emerson2017 CMV cohort, the ClareV Random Forest (RF) Fusion model improved AUC by 14.3% over the strongest V-family usage-only baseline (linear SVM) and by 13.3% over the strongest V-gene usage-only baseline (Random Forest). These results indicate that ClareV embeddings capture additional repertoire-level information beyond V-frequency summaries. Ablation studies showed that frequency-weighted bagging and contrastive V-bag learning both contributed to performance, outperforming simple pooling and non-contrastive learned-bag controls. In the CMV cohort, downstream embedding analyses showed that ClareV representations partially recapitulated IMGT-defined V-gene structure while retaining a CMV-responsive, non-sequence component that was largely orthogonal to classical V-usage signals.
Conventional T cells recognize peptides presented by the human leukocyte antigen (HLA) proteins through their T cell receptors (TCRs). Given that thousands of HLA proteins have been discovered, each presenting thousands of different peptides, decoding the cognate HLA protein of a TCR experimentally is a challenging task. To address this problem, we combined statistical learning methods with a unique dataset of paired T cell repertoires and HLA allotypes for 6,794 individuals. This enabled us to discover 34,206 T cell receptor alpha (TRA) and 891,564 beta (TRB) clonotypes that were associated with 175 unique HLA alleles. The identified clonotypes target prevalent infections, e.g. influenza, cytomegalovirus and Epstein-Barr virus. Utilizing these clonotypes, we develop statistical models that impute the carriership of common HLA alleles from the TRA- or the TRB- repertoire. In conclusion, the identified allele-associated clonotypes encode the HLA fingerprints and the antigenic exposure history of individuals and populations.