Liver Disease and Liver Cancer are Different Proteomic Diseases

Background.Chronic liver disease and hepatocellular carcinoma (HCC) arise in the same organ and share risk factors, but whether their prodromal circulating protein signatures overlap or diverge---and whether the architecture of detection-versus-death dissociation depends on the disease mechanism or the tissue of origin---has not been tested. Methods.We screened 2,922 proteins (Olink Explore 3072) against four endpoints in $\sim$53,000 UK Biobank participants: incident liver disease (K70/K73/K74/K76 from hospital episodes; $n = 1{,}583$), liver-related death ($n = 150$), incident HCC (C22; $n = 76$), and HCC death ($n = 77$). For each disease, two endpoint-specific portable panels were constructed using forward selection with shared foundation proteins forced in. Findings.Liver disease showed moderate detection--death dissociation (3/10 top proteins shared) with a strong inflammatory prodrome: GDF-15, IL-6, and KRT18 led the incidence ranking ($C^* = 0.671$--$0.666$), and GGT1 (Olink) replicated routine $\gamma$-GT exactly ($C^* = 0.658$ vs.\ $0.657$). HCC showed near-complete dissociation (1/10 shared) with dramatically stronger monovariate signals (ITGBL1 $C^* = 0.757$; Detection panel $C^* = 0.807$). The two diseases shared only 2/10 incidence proteins despite arising in the same organ. Coefficient behaviors diverged qualitatively: the sex coefficient \emph{flipped} between liver disease panels (female risk for detection, male risk for death), while HCC showed sex-null panels dominated by EGFR depletion ($\beta_\text{raw} = -2.79$, the single largest coefficient across four diseases studied). All eight panels were well calibrated (ICI $\leq 0.001$). Interpretation.Chronic liver disease and liver cancer engage fundamentally different prodromal proteomic architectures within the same organ. The degree and character of detection--death dissociation is a property of the disease mechanism---continuous fibrotic progression versus neoplastic transformation---not the tissue of origin. Six deployable logistic regression models are provided for external validation.