Blood Protein Tests Face a Reality Check Across the World’s Biggest Biobanks

Blood is the body’s open ledger. Every protein circulating in plasma carries a message about what the immune system is fighting, what the liver is manufacturing, what the heart and kidneys are quietly enduring. Over the past decade, two technologies — Olink and SomaScan — have made it possible to read thousands of those messages at once, cheaply enough to apply to hundreds of thousands of biobank volunteers. The result has been a flood of discoveries linking specific proteins to disease risk, and a wave of prediction models that claim to forecast everything from diabetes to dementia from a single blood draw. But a fundamental question has shadowed the field: do these platforms actually measure the same thing, and do models built on one platform work when applied to another?

A new study published in Nature Communications by Anthony Barente, Zijie Zhao, Zhili Zheng, Tai Wang, Benjamin Sun and colleagues, spanning researchers at Bristol Myers Squibb, Massachusetts General Hospital, the Broad Institute, the University of Helsinki, Queen Mary University of London and the Berlin Institute of Health at Charité, delivers the most systematic answer to date. The team generated and harmonized comparisons of the two leading multiplex affinity-based proteomic platforms across 11 independent cohorts, including data from the UK Biobank, FinnGen and the Genes & Health study. Their conclusion is both cautionary and constructive: platform differences are real and can seriously undermine the transferability of proteomic models, but with careful filtering, a substantial core of protein measurements is robust enough to carry predictions across cohorts and technologies.

The technical heart of the problem lies in how each platform captures proteins. Olink uses proximity extension assays, in which pairs of antibodies bearing DNA tags bind a target protein and only generate a measurable signal when both antibodies attach, producing highly specific readouts. SomaScan instead deploys slow off-rate modified aptamers — synthetic DNA-like molecules selected to bind specific protein epitopes with high affinity. Both approaches can quantify thousands of proteins from a few microliters of plasma, but they rely on different capture reagents, different normalization schemes and different dynamic ranges. When the same protein is measured by an antibody pair on one platform and an aptamer on the other, the two signals may correlate strongly, weakly, or not at all, depending on factors such as epitope choice, cross-reactivity, and how each assay handles protein isoforms and complexes.

To quantify these disparities, the researchers systematically correlated protein measurements across the 11 cohorts, identifying the key factors that drive platform disagreement. From this analysis they distilled a high-quality reference set of 757 protein probe-pairs that are consistently correlated across cohorts — a kind of gold-standard bridge between the two technologies. This curated set, released alongside the paper, gives emerging studies an immediate benchmark: proteins on this list can be trusted to behave comparably whether measured on Olink or SomaScan, making them natural candidates for cross-platform biomarker panels and for validating findings generated on either system.

The stakes become vivid when the team turns to prediction. They trained machine learning models to predict age, body mass index and 15 common diseases, then evaluated how well each model performed when applied to independent cohorts and, crucially, to the other platform. The results showed a systematic deterioration in model performance when models crossed platforms, and the degradation was especially severe when no filtering for probe correlations was applied. In other words, a model trained on thousands of unfiltered protein features on Olink may look impressive within its own dataset, yet quietly fall apart when handed SomaScan data — not because the underlying biology differs, but because the features themselves are not measuring the same quantities in the same way.

Perhaps the most intriguing finding is a U-shaped effect in model transferability that reflects a genuine trade-off between cross-platform protein consistency and biological relevance. Proteins with very high cross-platform correlation are technically reliable but tend to be abundant, stable, homeostatic molecules — the kind that carry little predictive signal for disease. Proteins with poor cross-platform agreement are often the most biologically interesting, capturing acute inflammatory signals, tissue leakage products or disease-specific isoforms, but their measurements are too platform-dependent to transfer. The sweet spot lies in between: proteins with moderate cross-platform correlation that remain biologically informative. Filtering models to retain probes above a certain correlation threshold dramatically improved transferability, and the optimal cut-off varied between traits, with values between 0.3 and 0.4 being the most prevalent across the diseases and quantitative traits examined.

This U-shaped relationship carries a practical message that could reshape how proteomic prediction studies are designed and reported. Simply maximizing the number of protein features, a common strategy in the current literature, is a recipe for fragile models. Instead, developers should explicitly consider which of their predictive proteins have been validated across platforms and cohorts, and should report transferability metrics alongside within-cohort performance. The study’s framework — training on one platform, testing on the other, and sweeping correlation thresholds — offers a reproducible template for doing exactly that, and the authors provide extensive supplementary data to support reanalysis.

The implications extend beyond prediction into drug development, which is where the industrial muscle behind the study becomes relevant. Pharmaceutical companies increasingly use plasma proteomics to identify drug targets, to find biomarkers that report on target engagement, and to stratify patients in clinical trials. A target whose protein signal cannot be measured consistently across platforms is a risky foundation for a companion diagnostic, and a biomarker panel that collapses when moved between assay systems can derail a translational program. By establishing which probe-pairs are dependable and which thresholds preserve predictive power, the study offers guidance for robust protein biomarker development, disease prediction and drug targeting — the three pillars of proteomic translational science.

The collaborative structure of the work is itself notable. Combining Bristol Myers Squibb’s informatics and translational teams with academic groups embedded in the FinnGen biobank in Finland, the UK Biobank, and the Genes & Health cohort of British South Asian volunteers allowed the analysis to span diverse ancestries, geographies and sample-processing protocols. That diversity matters, because a platform comparison conducted in a single population might conflate technical disagreement with genuine biological variation. Across 11 cohorts, consistent correlations emerged for a substantial fraction of the proteome, suggesting that the reference set of 757 probe-pairs is not an artifact of any one study population but a durable feature of how these technologies measure plasma proteins.

For the rapidly growing field of biobank-scale proteomics, the study arrives at a pivotal moment. Proteomic datasets linked to rich electronic health records are expanding worldwide, and protein-based risk scores are being proposed as the next generation of clinical screening tools. This analysis makes clear that the promise is real — proteomic prediction does generalize, provided researchers respect the technical boundaries of their assays. The era of treating every protein measurement as interchangeable is over; in its place stands a more mature discipline, armed with a validated cross-platform reference set, a principled filtering strategy, and a clearer understanding of where the biology and the technology meet. As blood-based protein tests edge closer to the clinic, that understanding may prove to be the most important biomarker of all.

Subject of Research: Systematic comparison of Olink and SomaScan affinity plasma proteomic platforms across 11 cohorts to assess cross-platform consistency and the generalizability of proteomic disease prediction models.

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

Article References: Barente, A., Zhao, Z., Zheng, Z., Wu, C., Singh, K., Wang, Z., Hall, A. O., Menard, L., Neuhaus, I., Kurki, M., Palotie, A., Daly, M., Finer, S., van Heel, D. A., Pietzner, M., Langenberg, C., Maranville, J., Wang, T., & Sun, B. B. (2026). Systematic comparison of affinity plasma proteomic technologies across multiple cohorts reveals generalizability of proteomic prediction. Nature Communications. https://doi.org/10.1038/s41467-026-78443-7

Image Credits: AI Generated

DOI: 10.1038/s41467-026-78443-7

Keywords: proteomics, plasma proteins, Olink, SomaScan, biobanks, disease prediction, biomarkers, machine learning, UK Biobank, FinnGen, cross-platform comparison, precision medicine