Detection–death dissociation in the pre-diagnostic cancer proteome tracks screening, not lethality
Year: 2026
Type: article
Abstract: Background. In companion studies, the circulating proteins predicting prostate cancer detection and prostate cancer death were almost disjoint, colorectal cancer (CRC) showed partial dissociation, and lung cancer none. The natural hypothesis (that dissociation tracks case fatality, the biological distance between initiation and lethality) was pre-registered and tested here across the pre-diagnostic proteome of sixteen cancers. Methods. In 52,949 UK Biobank participants with Olink Explore 3072 proteomics, all 2,922 evaluable proteins were screened against incident disease and cause-specific death for every ICD-10 C-code with sufficient events (seven cancers powered at $\ge$100 incidence / $\ge$80 mortality events; nine analyzed at exploratory floors). The 7x2 full-length screen vectors were correlated pairwise across cancers and endpoints; a single general factor was fitted to the cross-cancer cells by alternating least squares; each cancer's \emph{specific-axis concordance} $\rs$ (the detection$\leftrightarrow$death rank agreement of its own axis after removal of the shared factor) was computed with person-bootstrap confidence intervals (100 replicates; attenuation bias diagnosed and reported). Detection route was coded ordinally (0 = symptomatic to 3 = heavily screened) and tested head-to-head against case fatality. Three out-of-sample calls (melanoma, kidney, endometrial) were pre-registered before the exploratory rows were computed. Findings. A single shared axis absorbed 92\% of all cross-cancer covariance. Its top proteins (TFF3, GDF-15, suPAR/PLAUR, TNFRSF10B, EDA2R, IL-6, adrenomedullin, KIM-1) constitute the inflammaging/frailty proteome, the same stratum that dominates proteomic aging clocks and multi-disease protein scores [tanaka2018,lehallier2019,oh2023,gadd2024], and the axis correlated $\rho=0.70$ with an independent all-cause-mortality screen: the shared component of cancer prodromes is the general pre-mortality state, not cancer biology. Prostate cancer was orthogonal to it entirely (mean cross-correlation $-0.003$). After removing the shared factor, $\rs$ split the powered cancers into two bands that case fatality cannot explain: the three cancers detected by screening programs (prostate, breast, CRC) at $\rs = 0.13$--$0.15$, and the four symptomatically detected cancers (NHL, pancreas, esophagus, lung) at $0.41$--$0.79$, with NHL and CRC sharing identical case fatality ($\sim$24--25\%) on opposite sides. The band-means difference was $+0.395$ (bootstrap 95\% CI $+0.214$ to $+0.457$; bias-reflected $+0.334$ to $+0.576$). Detection route out-predicted case fatality (Spearman $-0.777$ vs $+0.464$; leave-one-out $-0.74$ to $-0.83$; all sixteen cancers $-0.658$), and the residual case-fatality effect within the symptomatic band was null-to-negative ($-0.40$). Of the pre-registered calls, melanoma confirmed ($\rs=0.05$, the paradigm overdiagnosis cancer at the bottom of the table), endometrial was directionally supportive under severe attenuation, and kidney missed ($\rs=0.30$), sharpening the law from a binary to a gradient in the asymptomatic-detection fraction. Interpretation. The pre-diagnostic proteome of cancer decomposes into one frailty terrain axis plus cancer-specific signatures whose detection$\leftrightarrow$death concordance is set by how a cancer is found, not how lethal it is. Detection--death dissociation is the proteomic shadow of overdiagnosis: screen-detected incidence contaminates the detection axis with lead-time and never-progressing biology, visible in blood a median of 6--8 years before diagnosis. This reinterprets the companion trilogy, cautions that early-detection biomarkers discovered against screen-era incidence labels partly learn \emph{detectability}, and yields a decisive external test: in screening trials with labeled interval cancers, interval-detected disease should show unified architecture while screen-detected disease dissociates.
Author: Jacobs, Laurence A.
Language: English
Subfield: Oncology
Field: Medicine
Domain: Health Sciences
Open Access status: green
OpenAlex ID: W7225662399
FWCI:
Cites:
Cited by:
Related to: