Network-guided prioritization of plasma protein signatures across 23 diseases

Population-scale plasma proteomics has begun to map the circulating proteome across human diseases, yet reducing these large datasets to interpretable compact protein panels that remain accurate across independent cohorts is a major challenge. Here, we developed a network-guided framework that links the Human Disease Blood Atlas (HDBA) to the UK Biobank Pharma Proteomics Project (UKB-PPP) through 1154 proteins quantified in both cohorts and evaluates protein signatures across 23 human diseases. For each disease, we considered two classification tasks: disease versus healthy and disease versus diseases. Compact protein panels achieved a mean area under the receiver operating characteristic curve (AUROC) of 0.741 and 0.736 for the two tasks, compared with 0.827 and 0.797 for a baseline model comprising all shared proteins, and outperformed this baseline in 7 of 23 diseases per task. Integrating panel membership with differential protein levels and disease-associated co-expression modules reveals disease-enriched candidates from disease-shared proteins associated with multiple conditions. This framework provides a structured, systems-level basis for prioritizing plasma protein signatures across diverse diseases.