Deep Learning Predictions of T Cell Receptor Epitope Affinity in COVID-19 Inform Repertoire Biases Associated With Disease Severity

Le résumé fourni par la source

Abstract Rationale: There is heterogeneity in the manifestations of patients with coronavirus disease 2019 (COVID-19). The T cell receptor (TCR) repertoire is responsible for pathogen recognition, and selection of suboptimal TCRs is implicated in an impaired response. We leveraged a deep learning pipeline to characterize the epitope affinities of TCRs in patients with COVID-19 to elucidate the relationship between TCR-epitope affinity and disease severity. Methods: We enrolled 351 patients presenting to the emergency room at a tertiary care hospital with febrile respiratory illness of which 302 patients had confirmed SARS-CoV-2 infection by nasal PCR. Clinical features and peripheral blood were collected at enrollment and on days 3, 7, 14, and 28. Single-cell RNA sequencing and TCR profiling were performed for each sample. Utilizing TCRconv, a previously reported deep learning pipeline combining an embedding model and convolutional neural network (CNN), we created embeddings from the CDR3 and TCRβ chain sequences in the VDJ database, and trained the CNN to predict affinity for 71 available epitopes across 14 species. We generated embeddings for each sequence in our samples and applied the CNN to predict epitope affinities. We modeled the association between TCRβ V/J expression with severity on a modified WHO ordinal scale using linear models adjusting for age, sex, and COVID-19 status and examined epitope specificity of severity-associated TCRβ V/J combinations. Results: Clonality was associated with age(p=2e-9) and age-adjusted severity(p=4e-4). Across 85,355 TCR sequences, epitopes from SARS-CoV-2 were predicted for 15,341(18.0%) sequences. Limiting the analysis to the 2,227 sequences with high-confidence predictions(ŷ>0.5) revealed 548(24.6%) sequences associated with SARS-CoV-2 epitopes. Across these sequences we identified 78 unique V/J gene combinations associated with severity (FDRBenjamini-Hochberg<0.05, FC>2) of which 30(38.5%) were associated with SARS-CoV-2 epitopes (TFE [nsequences=75, nV/J =12], YLQ [nsequences =72, nV/J=10], NYN [nsequences =17, nV/J =7], QYI [nsequences =1, nV/J=1]). Specificity within this set of genes was defined by the J gene with TRBJ1-2, TBRJ2-2, TRBJ2-7, TRBJ2-1 corresponding with TFE(10/12), YLQ(7/7), NYN(10/10), and QYI(1/1), respectively. However, expression of TRBV5 family genes corresponded with the strongest acuity-associated combination for YLQ(p=1.0E-3), NYN(p=8.26E-4), and QYI(p=1.0E-3). Conclusion: The diversity and composition of the TCR repertoire are associated with disease severity in COVID-19. However, the majority of severity-associated TCRs have affinity for non-SARS-CoV-2 epitopes. Although the TRBJ genes drive epitope specificity, expression of genes from the TRBV5 family yielded the strongest acuity association, suggesting that TRBV5 family expression may yield TCRs with suboptimal specificity across multiple SARS-CoV-2 epitopes.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.