Systematic comparison of affinity plasma proteomic technologies across multiple cohorts reveals generalizability of proteomic prediction

Emergence of biobank level proteomic datasets with extensive phenotype linkage has significantly improved our understanding of proteomic drivers in health and ushered in a flurry of disease risk prediction models. However, differences in cohorts and proteomic assays can significantly undermine transferability of proteomic models, severely limiting their general use. In this study, we generate and harmonize comparisons of two leading multiplex affinity-based proteomic platforms (Olink and SomaScan) across 11 independent cohorts. We find key factors contributing to platform disparities and establish a high-quality set of 757 protein probe-pairs consistently correlated across cohorts as a reference for new emerging studies. To investigate the impact on proteomic prediction, we train models to predict age, body mass index and 15 common diseases, evaluating performance across multiple independent cohorts and platforms. We find systematic deterioration in model performance across platforms, especially without filtering for probe correlations, and a U-shaped effect in model transferability reflecting a trade-off between cross-platform protein consistency and biological relevance. The optimal cross-platform correlation cut-off varied between traits, with 0.3–0.4 as most prevalent. Our study highlights practical considerations for assessing proteomic model performance across platforms and cohorts, offering guidance for robust protein biomarker development, disease prediction, and drug targeting. Proteomic prediction models often fail to transfer across cohorts and assay platforms. Here, the authors compare two leading proteomic technologies across 11 cohorts, identifying reliable proteins and correlation thresholds that yield more generalizable disease prediction models.