Smoking History Tells Who, the Plasma Proteome Tells When: Toward Protein-Triggered Lung Cancer Screening

Methods In 52,953 participants of the UK Biobank Olink proteomics subcohort without lung cancer at baseline, we screened 2{,}920 plasma proteins against incident lung cancer (cancer registry, ICD-10 C34; 548 events) and death from lung cancer (460 deaths), over a median follow-up of 13.2 and 15.3 years. Endpoint-specific Cox panels (Detection, Lethality) and stability-selected full-proteome models were evaluated by nested cross-validation with Harrell's concordance (C^*){C^* denotes Harrell's concordance statistic for censored time-to-event data: the probability that, of two participants whose order of events is known, the one with the earlier event has the higher score, against a smoking-history model fitted on the same folds (age, sex, smoking status, pack-years, body mass index, education, chronic obstructive pulmonary disease, previous cancer, ethnicity), by time from blood draw to the event. We measured how far the cases' protein scores rose above smoking history as diagnosis approached, and tested protein-triggered CT: within a pool defined by smoking history (American Cancer Society 2023 criteria), scanning only the people with the highest protein scores, compared in the same people with scanning by smoking-history rank, with a paired test whose primary contrast was specified after the unpaired results had been seen and before the paired analysis was run. Findings Detectable and lethal lung cancer were one proteomic disease: the ten leading proteins were the same for both endpoints (CXCL17 first, C^* = 0.761 and 0.764), the rankings were correlated across the proteome (Spearman ρ = 0.862), and a panel selected for death predicted death no better than a panel selected for diagnosis (difference $-0.000$, 95% CI $-0.004$ to $+0.003$). Over all follow-up, smoking history predicted incident lung cancer better than the proteins (out-of-fold C^* 0.868 against 0.854), and among never-smokers no protein model differed from smoking history, and the full-proteome models selected no protein. Among ever-smokers, the Detection panel exceeded smoking history for diagnoses within 2 years of the blood draw (0.881 against 0.819; difference $+0.062$, $+0.012$ to $+0.116$) but not after 5 years ($-0.004$, $-0.022$ to $+0.012$). Beyond smoking history, the panel score of future cases was 1.34 SD above that of non-cases within 2 years of diagnosis and 0.61 SD above it more than 5 years before (slope $-0.059$ SD per year, $-0.093$ to $-0.026$), while the smoking-history score of the same cases did not change with time to diagnosis ($-0.007$). The rise was concentrated in the last 2 to 3 years before diagnosis. In the 6{,}729 people meeting the 2023 criteria, scanning the fifth with the highest panel scores found 29 of the 45 cancers diagnosed within 2 years, against 17 when the same number was scanned by smoking-history rank (difference $+26.7$ percentage points, $+9.8$ to $+42.5$; exact McNemar $P = 0.008$), and 39 against 24 of the 61 lung cancer deaths within 5 years. Removing everyone diagnosed in the first year did not remove the advantage (27 against 15 of 48 cancers diagnosed 1 to 3 years after the draw; $P = 0.002$). A portable 2-year model of smoking history, CXCL17 and WFDC2 was calibrated (observed to expected 0.98). Interpretation Smoking history fixes who belongs in a lung screening pool; a single plasma draw indicates who in the pool is likely to be diagnosed in the next one to three years. Above a stable excess present a decade ahead, the protein signal rises as diagnosis approaches, and it ranks people for imaging better than smoking history does at the same number of scans. Protein-triggered CT within smoking-history-defined pools is ready for a prospective trial; the equations are provided for external validation.