Detectable and Lethal Colorectal Cancer are Partially Distinct Proteomic Diseases

Unlike prostate cancer, where PSA anchors screening, colorectal cancer has no circulating protein with comparable discriUnlike prostate cancer, where PSA anchors screening, colorectal cancer has no circulating protein with comparable discrimination, and CEA is used for monitoring rather than detection. Whether the circulating proteins that precede a colorectal cancer diagnosis are the same as those that precede death from colorectal cancer has not been examined across the plasma proteome with time-to-event methods. Methods.In 52{,}752 participants of the UK Biobank Olink proteomics subcohort without colorectal cancer at baseline, we screened 2{,}920 proteins against incident colorectal cancer (cancer registry, ICD-10 C18 to C20; 699 events) and death from colorectal cancer (underlying cause C18 to C20; 230 deaths) over a median follow-up of 13.2 and 15.3 years. Each protein was residualized on age and sex and ranked by Harrell's concordance ($C^*$)\footnote{$C^*$ 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.}. Endpoint-specific Cox panels (Detection, Lethality), a reference model of age, sex and the leading shared protein, stability-selected full-proteome models, a four-gate discovery against CEACAM5 (CEA) and 25 clinical covariates, and a cross-cancer test against prostate and lung cancer models were evaluated by nested cross-validation. The screen was repeated with a flexible age-by-sex adjustment. Findings. No protein dominated either endpoint, and none reached $C^* = 0.60$. Seven of the ten leading proteins were shared between the two endpoints. Two of the leading proteins, RLN1 (first for mortality, $C^* = 0.590$) and RLN2, ranked high only because a linear age adjustment did not follow their steep decline with age in women; with a flexible adjustment they fell out of the leading proteins and carried no weight beside age and sex. In the flexible screen, six of the ten leading proteins were shared (REG4, CRYBB2, KRT19, TFF2, POF1B, GDF15). Age and sex alone reached nested $C^* = 0.666$ for incidence and 0.671 for mortality. The Detection panel reached 0.672 and 0.690, the Lethality panel 0.679 and 0.685. The Lethality panel did not predict death better than the Detection panel (difference $-0.005$, 95\% CI $-0.023$ to $+0.014$), and on incidence it was slightly ahead ($+0.007$). Over age and sex, the panels gained between $+0.005$ and $+0.019$; only the Lethality panel on incidence was significant ($+0.013$, $+0.004$ to $+0.022$). CEACAM5 alone reached 0.553 and 0.590, and age, sex and CEACAM5 0.670 and 0.682. Full-proteome models were below the panels (0.653 and 0.660); the mortality model kept three stable proteins, and none in three of five training folds. The panels and the reference model were calibrated in the large on both endpoints (observed to expected 0.961 to 0.985). nterpretation. The circulating proteins that rank highest years before a colorectal cancer diagnosis also rank highest before death from it, and a panel selected for diagnosis predicts death as well as a panel selected for death. Unlike prostate cancer, colorectal cancer shows largely shared leading proteins for the two endpoints; as in prostate cancer, no lethality-specific model survived nested validation. The proteomic models discriminate modestly and add little to age and sex. A linear age adjustment can promote proteins with sex-specific age patterns to the top of a residualized screen; such screens should be checked with a flexible adjustment and against Cox models with age and sex. The panels are provided with full coefficients for external validation.