Generalizability of proteomic risk prediction across biobanks reveals dependence on phenotype definitions
- 1Analytic and Translational Genetics Unit, Massachusetts General Hospital, Boston, MA, USA; Program in Medical and Population Genetics, Broad Institute of MIT an
Abstract
Advances in high-throughput proteomics technologies have enabled the assessment of dynamic health states across biobank-scale cohorts. Disease prediction models built on these data have higher accuracy than baseline clinical models for a broad range of diseases, and provide avenues to understand disease pathogenesis. However, the generalizability of these prediction models across cohorts has yet to be established at scale. Here, we train models for 15 diseases in the UK Biobank (UKB; n=53,026) based on Olink proteomics data, and find high accuracy for disease prediction (mean AUC=0.74; range 0.56-0.89). These models have improved accuracy over models using only clinical factors (mean {Delta}AUC = 0.03), and do not depend on model architecture, with simpler models (e.g. L2) performing as well as complex models (e.g. transformer). We assess generalizability of the UKB-trained models in two external cohorts: FinnGen (FG; n=5,865) and All of Us (AoU; n=7,405), spanning two Olink platforms. We find that proteomics-based prevalent (classification) and incident (prediction over next 5 years) disease models largely generalize across cohorts, but performance varies across diseases. Specifically, 10 of 14 prevalent and 13 of 15 incident disease models show no significant decrease in performance across any biobank. Adjusting for demographic, ancestry, and technical covariates, we demonstrate that differences in phenotyping quality are likely the major drivers of variability across cohorts. These results indicate that proteomic risk models can be powerful and generalizable predictors of disease across multiple cohorts.
Related
- Generalizability of a prediction model for walking independence at discharge in older adults after hip fracture surgery: a multicenter study
- Electronic health record phenotype choice and real-world outcome estimates: A propensity score–matched comparison of three invasive breast cancer definitions in TriNetX
- Cross-cohort generalizability of machine learning models for early-stage node-negative breast cancer relapse prediction
- Early Right Heart Failure After Left Ventricular Assist Device Implantation: A Narrative Review of Definitions, Risk Prediction, and Management
- Large-scale proteomic characterization reveals clinically relevant subtypes of early-stage papillary thyroid cancer for prognostic prediction
BibTeX
@misc{parsa2026generalizability,
title = {{Generalizability of proteomic risk prediction across biobanks reveals dependence on phenotype definitions}},
author = {S. J. Parsa and Y. H. Jee and M. A. Argentieri and J. Guez and W. Lu and E. M. Padhi and E. Kiernan and R. Jabal and M. I. Kurki and FinnGen and B. M. Neale and L. Lichtenstein and M. E. Talkowski and S. B. Montgomery and N. J. Lennon and A. R. Martin and M. J. Daly and K. J. Karczewski},
howpublished = {medRxiv},
year = {2026},
month = oct,
doi = {10.64898/2026.10.01.26364038},
url = {https://www.medrxiv.org/content/10.64898/2026.10.01.26364038v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - Generalizability of proteomic risk prediction across biobanks reveals dependence on phenotype definitions
AU - S. J. Parsa
AU - Y. H. Jee
AU - M. A. Argentieri
AU - J. Guez
AU - W. Lu
AU - E. M. Padhi
AU - E. Kiernan
AU - R. Jabal
AU - M. I. Kurki
AU - FinnGen
AU - B. M. Neale
AU - L. Lichtenstein
AU - M. E. Talkowski
AU - S. B. Montgomery
AU - N. J. Lennon
AU - A. R. Martin
AU - M. J. Daly
AU - K. J. Karczewski
PY - 2026
DA - 2026/10/01
PB - medRxiv
DO - 10.64898/2026.10.01.26364038
UR - https://www.medrxiv.org/content/10.64898/2026.10.01.26364038v1
AB - Advances in high-throughput proteomics technologies have enabled the assessment of dynamic health states across biobank-scale cohorts. Disease prediction models built on these data have higher accuracy than baseline clinical models for a broad range of diseases, and provide avenues to understand disease pathogenesis. However, the generalizability of these prediction models across cohorts has yet to be established at scale. Here, we train models for 15 diseases in the UK Biobank (UKB; n=53,026) based on Olink proteomics data, and find high accuracy for disease prediction (mean AUC=0.74; range 0.56-0.89). These models have improved accuracy over models using only clinical factors (mean {Delta}AUC = 0.03), and do not depend on model architecture, with simpler models (e.g. L2) performing as well as complex models (e.g. transformer). We assess generalizability of the UKB-trained models in two external cohorts: FinnGen (FG; n=5,865) and All of Us (AoU; n=7,405), spanning two Olink platforms. We find that proteomics-based prevalent (classification) and incident (prediction over next 5 years) disease models largely generalize across cohorts, but performance varies across diseases. Specifically, 10 of 14 prevalent and 13 of 15 incident disease models show no significant decrease in performance across any biobank. Adjusting for demographic, ancestry, and technical covariates, we demonstrate that differences in phenotyping quality are likely the major drivers of variability across cohorts. These results indicate that proteomic risk models can be powerful and generalizable predictors of disease across multiple cohorts.
ER -