UCSF

T-cell Receptors (TCRs) play a pivotal role in antigen recognition and binding, and their sequence similarity significantly impacts the breadth of antigen recognition. Network analysis is employed to explore TCR sequence similarity and investigate the architecture of the TCR repertoire. Network properties hence could be utilized to quantify the structure of the TCR network. However, the heterogeneous nature of TCR network properties poses challenges in performing statistical learning across subjects directly, particularly when assessing their relationship with disease states, clinical outcomes, or patient characteristics. To overcome this challenge, a powerful method is developed, TCR-NP (TCR Network properties Prioritization), that aggregates the raw heterogeneous network properties and conducts grouped feature selection using a pseudo-variables-assisted penalized group Lasso model. Unlike the traditional parameter-tuning using cross-validation, a novel tuning strategy is introduced by incorporating permutation and pseudo-variables to improve the selection performance. The effectiveness of the proposed method is demonstrated through comprehensive evaluation, including simulation studies and real data analysis. By comparing the performance of the different approaches, the advantages of the proposed methodology in capturing the underlying relationships between TCR network properties and clinical outcomes or patient characteristics are highlighted.