Five to twenty proteins in the blood, and the risk of 218 diseases

If you could measure thousands of proteins in a single blood sample, would that tell you, better than a doctor's usual questions and tests, who is likely to develop a given disease in the next ten years? And for which diseases? The researchers asked this question for 218 diseases at once, common and rare, and insisted on a fair comparison: against what a family doctor already knows (age, sex, weight, smoking, family history) and against 37 ordinary blood tests.

The UK Biobank is a group of about half a million British adults aged 40 to 69 at enrolment, followed through their health records. In a subset of 41,931 of them, about 3,000 proteins had been measured in a blood sample taken at enrolment, as part of a project funded by a consortium of pharmaceutical companies. The researchers built, for each of 218 diseases, a model that uses a handful of proteins (between 5 and 20, chosen automatically) on top of basic clinical information, and measured how well it ranked people by their chance of a diagnosis within ten years. The yardstick is the C-index: 0.5 means guessing, 1 means perfect ranking. Each model was trained on part of the data and tested on a part kept aside. For six diseases, the models were then tried on a different British cohort, EPIC-Norfolk.

Basic clinical information alone gave a median C-index of 0.64 across diseases. Adding 5 to 20 proteins improved the ranking for 67 of the 218 diseases, by a median of 0.07 points (from 0.02 to 0.31 depending on the disease). The largest gains were for diseases of the blood and immune system, such as multiple myeloma (+0.25) and coeliac disease (+0.31), and for motor neurone disease, pulmonary fibrosis and dilated cardiomyopathy. For 52 of those 67 diseases, the protein models also beat models that added 37 standard blood tests. For most of the 67, five to ten proteins were enough: adding more did not help. Genetic risk scores, by comparison, improved prediction for only seven diseases and by much less. In the second cohort, the improvement was confirmed for five of the six diseases tested. The observation is the ranking gain; the interpretation the authors propose is that small protein panels could become affordable tests for selected diseases, with proteins specific to one disease pointing to its biology.

The size of the gain is easy to overstate. A median improvement of 0.07 on the C-index means the model ranks people somewhat better; it does not mean it predicts their fate. For 151 of the 218 diseases there was no significant gain at all, and the authors say the improvement was generally larger for rarer diseases, which have clearer biology, than for common complex ones such as heart failure or type 2 diabetes. The study reports no figures on calibration (whether a predicted 10% risk means 10 people in 100 really fell ill), and no attempt to show that acting on these predictions helps anyone.

For prevention, the study's value is in choosing where to look. Knowing that a few proteins predict multiple myeloma or coeliac disease years ahead suggests where an earlier diagnosis might be possible; whether finding those people earlier changes their outcome is a separate question, which needs its own trials. For diseases where lifestyle prevention is proven, such as type 2 diabetes, this study does not show that a protein signature would help choose who to treat.

Where this fits the Digital Twin: not as data the Twin reads today (these proteins require a research-grade laboratory platform), but as a map. It tells which diseases a small protein panel might one day flag and which it will not, and it sets the bar any new signal must clear: better than what a doctor already knows, and better than ordinary blood tests. If such panels become available as clinical tests, their results would be one more dated entry in a person's record. The study measured each person once; it says nothing about following proteins over time.

It does not provide a test: the protein panels exist as research models, not as assays a clinic can order, and the authors themselves list the first limitation as the need for validation in external studies, in ethnically diverse populations and in groups with different baseline risk; UK Biobank participants are healthier than the general population. External replication covered six diseases, not 218. The ten-year horizon is the model's, not a proof that the proteins change ten years before illness. Nothing here applies to a person reading their own blood results. Several authors are employees of a pharmaceutical company that co-funded the proteomic measurements; this is stated in the paper and matters for how the results are read.

Prospective cohort (UK Biobank Pharma Proteomics Project, random subset) with derivation of sparse prediction models for 10-year incidence of 218 diseases; internal validation on a held-out test set; external validation in EPIC-Norfolk for six diseases.

41,931 participants (random subset of UKB-PPP; UK Biobank recruited adults aged 40–69 in 2006–2010); 218 diseases with > 80 incident cases within 10 years (81–6,038 cases); 193 diseases in the random subset, 25 adding consortium-selected cases. EPIC-Norfolk: 25,639-person cohort from Norfolk, England; n = 295–1,116 and 5–236 incident cases per disease tested.

EDTA plasma at baseline (non-fasting); Olink Explore 1536 + Expansion (2,923 proteins, 2,941 assays, normalised protein expression). Clinical benchmark: age, sex, BMI, self-reported ethnicity, smoking, alcohol, family history (15 diseases); second benchmark adds 37 clinical blood assays. Feature selection by LASSO over 200 subsamples; regularised Cox models with 5-fold cross-validation; performance on held-out test set with 1,000 bootstrap samples (C-index, delta C-index with 95% CI, likelihood ratios, detection rate at 10% false-positive rate). Polygenic scores compared where available.

Clinical models: median C-index 0.64 (IQR 0.58–0.72). Sparse protein models (5–20 proteins) improved on clinical models for 67 diseases (median delta C-index 0.07, range 0.02–0.31), e.g. multiple myeloma 0.25 (95% CI 0.20–0.29), non-Hodgkin lymphoma 0.21 (0.14–0.28), coeliac disease 0.31 (0.21–0.38), dilated cardiomyopathy 0.17 (0.10–0.22), pulmonary fibrosis 0.09 (0.03–0.14), motor neuron disease 0.11 (0.04–0.16). Clinical-assay models improved on clinical models for only 28 diseases (median delta 0.08); protein models exceeded them for 52 of the 67. Saturation at 5–10 proteins for 64% of improved diseases; average 5.4% C-index gain of sparse signatures over the single best protein. Polygenic scores: significant improvement for 7 diseases only (median delta 0.03 vs 0.08 for proteins). External validation (EPIC-Norfolk, 6 diseases): C-indexes correlated r = 0.81, improvements r = 0.97; improvement replicated for 5 of 6 (COPD only at 20-year incidence). Calibration not reported in the sections read. Implicit negative result: no significant protein gain for 151 of 218 diseases.

PROBAST (preliminary, on full text): participants low-to-moderate (healthy-participant effect acknowledged); predictors low (standardised platform, measured before outcome); outcome low-to-moderate (EHR-based definitions, first 6 months excluded); analysis low (adequate events per disease by design threshold, penalised selection, held-out test, bootstrap; calibration not reported). Generalisability: predominantly European ancestry; external validation limited to six diseases in another British cohort. Reporting: detailed methods; TRIPOD adherence not assessed here. Competing interests: several authors employed by GSK, which co-funded the proteomic measurements (affiliations in the PubMed record).

Potential, not demonstrated: screening metrics (detection rate at fixed false-positive rate) are reported, but no decision-curve analysis, no impact study and no clinical assay. The authors frame the result as the basis for affordable sparse panels for selected diseases, to be validated.

B — guide to which diseases and which proteins; integrable only after validation for the specific use and with targeted assays, not the whole platform. Required data: a targeted protein panel from a laboratory, plus the clinical variables of the benchmark.

PubMed record and the full text of the published article (PubMed Central), read in full on 5 November 2026. Supplementary tables (lists of the 67 and 52 diseases, protein components per disease) not consulted.

For many diseases there are delays in diagnosis due to a lack of objective biomarkers for disease onset. Here, in 41,931 individuals from the United Kingdom Biobank Pharma Proteomics Project, we integrated measurements of ~3,000 plasma proteins with clinical information to derive sparse prediction models for the 10-year incidence of 218 common and rare diseases (81-6,038 cases). We then compared prediction models developed using proteomic data with models developed using either basic clinical information alone or clinical information combined with data from 37 clinical assays. The predictive performance of sparse models including as few as 5 to 20 proteins was superior to the performance of models developed using basic clinical information for 67 pathologically diverse diseases (median delta C-index = 0.07; range = 0.02-0.31). Sparse protein models further outperformed models developed using basic information combined with clinical assay data for 52 diseases, including multiple myeloma, non-Hodgkin lymphoma, motor neuron disease, pulmonary fibrosis and dilated cardiomyopathy. For multiple myeloma, single-cell RNA sequencing from bone marrow in newly diagnosed patients showed that four of the five predictor proteins were expressed specifically in plasma cells, consistent with the strong predictive power of these proteins. External replication of sparse protein models in the EPIC-Norfolk study showed good generalizability for prediction of the six diseases tested. These findings show that sparse plasma protein signatures, including both disease-specific proteins and protein predictors shared across several diseases, offer clinically useful prediction of common and rare diseases.