Abstract
Early detection of lung cancer remains challenging due to limitations of current methods. We developed LCPBert, a deep learning framework leveraging peripheral blood T cell receptor beta (TCRβ) repertoires for early detection of lung cancer. LCPBert accurately discriminated lung cancer-associated TCRs (test AUC = 0.82). Based on LCRI (lung cancer risk index), LCPBert robustly stratified lung cancer risk in the external validation cohort: healthy donors (0.111 ± 0.058), benign pulmonary nodules (0.184 ± 0.113), lung cancer (0.296 ± 0.166; p < 0.001) and showed a spatial gradient from peripheral blood (0.296) to tumor tissue (0.384, p < 0.084). At an LCRI cutoff of 0.1465, LCPBert achieved 75% sensitivity and 70% specificity in discriminating lung cancer patients from healthy individuals. In a longitudinal cohort, LCRI elevation (Δ > 0.15) exceeding 0.30 after SBRT predicted distant metastasis (DM) in 75% of patients who developed DM. In addition, LCRI predicted lung cancer independently of age, sex, and TCR diversity (D50). LCPBert provides an accurate and non-invasive approach for early-stage lung cancer detection.