Proteomic prediction of common and rare diseases

Background For many diseases there are delays in diagnosis due to a lack of objective biomarkers for disease onset. Whether measuring thousands of proteins offers predictive information across a wide range of diseases is unknown. Methods In 41,931 individuals from the UK Biobank Pharma Proteomics Project (UKB-PPP), we integrated ∼3000 plasma proteins with clinical information to derive sparse prediction models for the 10-year incidence of 218 common and rare diseases (81 – 6038 cases). We compared prediction models based on proteins with a) basic clinical information alone, b) basic clinical information + 37 clinical biomarkers, and c) genome-wide polygenic risk scores. Results For 67 pathologically diverse diseases, a model including as few as 5 to 20 proteins was superior to clinical models (median delta C-index = 0.07; range = 0.02 – 0.31) and to clinical models with biomarkers for 52 diseases. In multiple myeloma, for example, a set of 5 proteins significantly improved prediction over basic clinical information (delta C-index = 0.25 (95% confidence interval 0.20 – 0.29)). At a 5% false positive rate (FPR), proteomic prediction (5 proteins) identified individuals at high risk of multiple myeloma (detection rate (DR) = 50%), non-Hodgkin lymphoma (DR = 55%) and motor neuron disease (DR = 29%). At a 20% FPR, proteomic prediction identified individuals at high-risk for pulmonary fibrosis (DR= 80%) and dilated cardiomyopathy (DR = 75%). Conclusions Sparse plasma protein signatures offer novel, clinically useful prediction of common and rare diseases, through disease-specific proteins and protein predictors shared across multiple diseases. (Funded by Medical Research Council, NIHR, Wellcome Trust.)