Plasma proteomic profiles predict individual future health risk

Comprehensive risk assessments for various diseases often require numerous individual predictors, making clinical application challenging due to time and cost constraints. A single-domain assay capable of simultaneously assessing multiple disease risks is highly desirable. Routine blood tests are becoming increasingly common, and proteomics-based risk scores offer a promising approach to improve multi-disease risk prediction. While proteomics has shown promise in understanding gene-protein interactions, biomarkers in individual diseases, aging, and drug pharmacology, its potential for simultaneously assessing multiple future health risks requires further investigation. Most existing proteomic studies are cross-sectional or utilize case-control approaches, potentially biased by reverse causality. Longitudinal studies are needed to identify molecular signatures before disease onset. Furthermore, the shared molecular pathways underlying various diseases are not well understood, hindering systematic understanding of different incident diseases. This study aimed to explore the potential of proteomic profiles to predict multi-disease and mortality risk by constructing a disease/mortality-specific proteomic risk score (ProRS) using data from the UK Biobank. The predictive power of ProRS would be compared with established clinical predictors to assess its potential clinical utility across multiple diseases and mortalities.

Existing literature highlights the potential of proteomic profiling in predicting various health outcomes. Studies have shown its high predictive value for cardiovascular events, obesity, dementia, and cancer. However, these studies often focus on specific diseases or utilize cross-sectional or case-control designs which may not fully capture the predictive power of proteomics for a broad spectrum of disease outcomes. The use of metabolomics to predict multi-disease outcomes has been explored, but its predictive power alone may not surpass simpler models combining age and sex. This study aimed to address these limitations by utilizing a large, longitudinal cohort and a comprehensive set of disease outcomes to assess the broad applicability and predictive ability of blood proteomics for predicting future health problems. The study also aimed to fill a gap in knowledge on shared pathways underlying various diseases.

This retrospective study used data from 52,006 UK Biobank participants with available blood proteomics data and a median follow-up of 14.1 years. 45 endpoints were defined, encompassing 14 disease categories, 26 specific diseases, and all-cause mortality with four cause-specific mortalities, using ICD-10 codes. 1461 Olink plasma proteins were measured. A multilayer perceptron neural network (ProNNet) was developed to generate a proteomic risk score (ProRS) for each endpoint. ProNNet consisted of two modules: a comorbid network to estimate overall health conditions and an endpoint-specific network to customize risk prediction for each endpoint. The comorbid network was pre-trained and its weights passed to the endpoint-specific network. The model used ReLu and Sigmoid activation functions. Cox proportional hazard (CPH) models were fitted using ProRS alone, and in combination with three sets of clinical predictors (Age+Sex, Serum, PANEL), which includes 54 clinical predictors covering demographic information, lifestyle factors, physical measurements, disease and medication history, family disease history, and serum biochemistry data. Model performance was evaluated using Harrell's C-index, calibration plots, and net benefit curves. Shapley Additive explanations (SHAP) values were used to identify important proteins for each endpoint, and their associations were explored with CPH models. Leave-one-region-out cross-validation was used for model development and evaluation.

ProRS effectively stratified the risk for all 45 conditions studied across 14 disease categories, including infectious, hematological, endocrine, psychiatric, neurological, sensory, circulatory, respiratory, digestive, skin, musculoskeletal, and genitourinary diseases, cancers, and mortality. Participants with higher ProRS percentiles showed elevated observed event rates. Age was positively correlated with ProRS for most endpoints. Males had a higher risk of cancer, circulatory system disease, and all-cause mortality than females at the same ProRS percentile. Kaplan-Meier survival curves showed significantly different survival probabilities across ProRS tertiles. ProRS alone achieved Harrell's C-indexes exceeding 0.80 for several endpoints (all-cause mortality, and seven specific diseases). In most cases, ProRS alone demonstrated better or comparable predictive performance compared to using Age+Sex, Serum, or PANEL alone. Combining ProRS with other clinical predictors led to improved predictive performance for most endpoints but only marginal improvements over ProRS alone. Proteins like GDF15 showed important discriminative value for various diseases. Model calibration and decision curve analysis demonstrated the clinical utility of ProRS, often exceeding clinical predictors alone and their combinations.

This study demonstrates that plasma proteomic profiles, integrated into a risk score (ProRS), effectively stratify risk across a wide range of common diseases and mortality. The superior performance of ProRS, often surpassing established clinical predictors, underlines its potential as a single-domain assay for comprehensive risk assessment. The limited additional benefit of combining ProRS with clinical predictors suggests that ProRS captures significant information not readily available through conventional methods. The identification of key proteins like GDF15 with strong discriminative value across various diseases opens avenues for targeted interventions. The translation of the improved discrimination to improved clinical utility further strengthens the potential of ProRS for clinical application. These results highlight the potential of blood proteomics as a streamlined approach to comprehensive risk assessment, simplifying the complex process of gathering multiple predictors for individualized risk stratification.

This study demonstrates the significant potential of using plasma proteomic profiles to predict individual future health risks across a wide range of diseases and mortality. The developed ProRS outperforms existing clinical predictors in many instances and offers a promising single-domain assay for comprehensive risk assessment. Future research should focus on independent external validation in diverse populations and explore the potential for using ProRS in clinical decision-making. Further investigation into the underlying biological mechanisms associated with key proteins such as GDF15 could lead to targeted preventative interventions.

The study's limitations include the predominantly white European population of the UK Biobank, limiting generalizability to other ethnicities. The exclusion of proteins not included in the Olink panels might have affected the predictive performance of the model. The UK Biobank population might be healthier than other populations, potentially underestimating the incidence rates of certain diseases. Finally, while internal cross-validation was performed, independent external validation is crucial before clinical implementation.