A dynamic proteomic signature for the prediction of lung Cancer: a longitudinal analysis in the UK Biobank cohort

Background
To understand the molecularly obscure pre-diagnostic phase of lung cancer, we mapped the temporal evolution of the plasma proteome for new biological insights and improved risk prediction.
Methods
Leveraging the UK Biobank prospective cohort, we analyzed 2,921 plasma proteins from 37,759 participants, including 342 incident lung cancer cases identified over a median follow-up of 11.7 years. We employed time-stratified Cox models, locally weighted scatterplot smoothing (LOESS) trajectory modeling, and hierarchical clustering to characterize protein dynamics relative to the time of diagnosis. A multi-algorithm machine learning pipeline was used to develop a predictive signature, and two-sample Mendelian randomization was performed to infer causal relationships.
Results
We identified 340 risk-associated proteins showing significant temporal heterogeneity. Long-term risk (>5 years pre-diagnosis) was linked to proteins like CEACAM5, indicating early dysregulation of cell adhesion. Imminent risk (<5 years) was marked by a surge in inflammatory proteins like IL6. These dynamics were resolved into four distinct trajectory patterns, creating a molecular timeline of carcinogenesis. A machine learning-derived 28-protein signature, integrated with clinical factors and Polygenic Risk Score (PRS), achieved outstanding predictive performance (AUC = 0.830). Mendelian randomization also suggested a causal role for some proteins of 340 risk-associated proteins in lung cancer development.ConclusionOur findings establish that lung cancer evolves through a dynamic sequence of protein changes. This provides a new model for understanding pre-diagnostic disease, and our 28-protein signature is a powerful tool for precision screening to identify individuals with active disease progression.