The Blood Disease Atlas - Methods summary

Key publications

Álvez MB et al. (2025) "A human pan-disease blood atlas of the circulating proteome" Science 390, 6779 (2025).

Álvez MB et al. (2023) "Next generation pan-cancer blood proteome profiling using proximity extension assay" Nat Commun 14, 4308 (2023).

Kotol D et al. (2023) "Absolute quantification of pan-cancer plasma proteomes reveals unique signature in multiple myeloma" Cancers 15(19), 4764 (2023).

What can you learn from the Disease Blood Atlas?

Learn about

• Protein levels in blood covering 70 diseases
• Proteins associated with each disease, compared with healthy individuals and with other diseases based on differential expression and machine learning analyses
• How measurements of the same protein compare across technologies

Each protein’s gene summary page shows its plasma levels across all diseases, and whether the protein is differentially abundant in any disease.

Data overview

The Human Disease Blood Atlas currently includes five datasets, each generated with a different protein profiling technology (Table 1). They differ in how many proteins were measured, how many samples were analyzed and which diseases are included.

Table 1. Datasets in the Human Disease Blood Atlas. For each dataset: the technology used, the number of proteins measured, the number of plasma samples and the number of phenotypes (diseases and healthy controls).

Technologies

The Human Disease Blood Atlas combines data from affinity-based and mass spectrometry-based technologies (Table 2). The affinity-based data were generated with two versions of the Olink Proximity Extension Assay (PEA), Explore 1536 and Explore HT, and with the SomaScan 11K aptamer-based assay. These methods measure thousands of proteins with high sensitivity and report relative protein levels. Targeted mass spectrometry measures fewer proteins but gives absolute concentrations. Because the technologies use different detection principles and units, values for the same protein cannot be compared directly across datasets. They can also differ in which diseases show changes in that protein.

Table 2. Overview of the technologies in the Human Disease Blood Atlas. For each dataset: the detection principle, readout, unit, type of quantification and number of proteins.

How was the Proximity Extension Assay data generated?

Proximity Extension Assay (PEA) uses pairs of antibodies labelled with DNA oligonucleotides. When both antibodies bind the same target protein, the oligonucleotides come close enough to hybridize. They are then extended by a DNA polymerase, creating a unique DNA barcode for each protein. The barcodes are amplified and quantified by next generation sequencing (Wik L et al. (2021)). Only a few microliters of plasma are needed. Results are reported as Normalized Protein eXpression (NPX), a relative unit on a log2 scale. Two versions were used in the atlas: Olink Explore 1536 covers ~1,500 assays across four panels (cardiometabolic, inflammation, neurology, and oncology), and Olink Explore HT covers ~5,400 assays.

How was the SomaScan data generated?

The SomaScan 11K platform uses ~11,000 SOMAmer reagents (Slow Off-rate Modified Aptamers). These are chemically modified DNA aptamers that bind their target proteins with high specificity (Gold L et al. (2010)). After binding, the protein–SOMAmer complexes are captured and washed. The SOMAmers are then released and quantified by hybridization to a DNA microarray. Thousands of proteins can be measured from a small volume of plasma. Results are reported as Relative Fluorescence Units (RFU) and normalized with Adaptive Normalization by Maximum Likelihood (ANML), which adjusts each sample against a healthy reference population. Values are shown on a log2 scale.

How was the Targeted Proteomics data generated?

In targeted proteomics, proteins are digested into peptides with trypsin. A predefined set of peptides is then measured by liquid chromatography–tandem mass spectrometry (LC-MS/MS). Heavy isotope-labelled protein standards are spiked into each sample in known amounts before digestion. They are processed together with the endogenous proteins, so they correct for losses and digestion bias throughout the workflow. This allows absolute quantification. The atlas includes two targeted datasets, both measured in undepleted plasma. The pan-cancer dataset used 273 stable isotope-labelled PrEST standards (SIS-PrESTs) produced from Human Protein Atlas protein fragments (Zeiler M et al. (2012)) in plasma from 1,469 cancer patients. The pan-disease dataset used the DiscoveryEdge175 (DE175) panel (ProteomEdge), which contains heavy-labelled recombinant protein standards (qRePS), in 2,981 samples. For more information on the targeted proteomics data visit: Method details.

Analyses

Differential abundance analyses

To identify disease-associated proteins, each disease was compared against up to three control groups:

• Healthy individuals: each disease compared with healthy controls
• Diseases from the same class: each disease compared with the other diseases in its class (e.g. pancreatic cancer compared against all other cancer types)
• All other diseases: each disease compared with all other diseases in the dataset

All three comparisons are available for the Olink Explore 1536 dataset. For Olink Explore HT and SomaScan 11K, each disease was compared with all other diseases.

Differential abundance was assessed with linear models (limma, Ritchie ME et al. (2015)), adjusted for age and sex where applicable. Proteins with a Benjamini-Hochberg adjusted P value below 0.05 were considered differentially abundant. The results for each disease are summarized in volcano plots on the disease pages. Find more information in: Method details.

Machine learning analyses

To identify proteins that together distinguish a disease from others, lasso regularized logistic regression models were trained for diseases with sufficient sample size. The control groups were the same as in the differential abundance analyses: all three settings for Olink Explore 1536, and healthy controls for Olink Explore HT and SomaScan 11K. Models were trained and tested on repeated random splits of the data (70% training, 30% testing). For each disease, the top-ranking proteins are reported with their importance scores. Importance is calculated from the absolute model estimates, averaged across splits and scaled to 0–100. Find more information in: Method details.

Comparing across platforms

The same 2,621 plasma samples were profiled with both Olink Explore HT and SomaScan 11K, so the two platforms can be compared directly. Proteins were matched across platforms by gene symbol. In SomaScan, a protein can be targeted by several SOMAmers with different epitope specificities. Each Olink–SOMAmer pair was therefore treated separately, which gave 4,294 protein-aptamer pairs.

Baseline correlation. For each protein-aptamer pair, agreement between the platforms was measured as the Spearman correlation between Olink (NPX) and SomaScan (log2 RFU) values across all 2,621 samples.

Disease-specific correlation. Many proteins measured by both platforms are present at low levels in the blood of most individuals and rise only in certain diseases. Correlations across all samples can therefore underestimate agreement for proteins that are well measured when their levels are elevated. To capture this, the Spearman correlation was also calculated separately within each disease. The highest of these per-disease correlations is reported as the protein’s disease-specific correlation. The difference between the disease-specific and baseline correlations (Diff) shows how much agreement improves in a specific disease context. The heatmap on the cross-platform page compares baseline and disease-specific correlations for each pair.

Detectability. For each protein, a sample was counted as below the limit of detection (LOD) if either the Olink or the SomaScan measurement fell below that platform’s LOD. Olink LOD was calculated from negative controls and SomaScan LOD from buffer samples. A protein’s detectability is the percentage of samples above LOD on both platforms. Proteins with low detectability on either platform should be interpreted with caution when comparing measurements across technologies.

Downloadable results