Multi-Omics: 15 Proteins Sharpen Type 2 Diabetes Risk
- Aug 30
- 4 min read
Clinical risk scores for type 2 diabetes already work well, which sets a high bar. A new biomarker panel has to beat age, BMI, family history and HbA1c before anyone pays to measure it. A multi-omics analysis of 42,840 UK Biobank participants, published in Cardiovascular Diabetology in May 2026, tested plasma proteomics, NMR metabolite profiling and a polygenic risk score against that benchmark over ten years and 1,090 incident cases. Which layer you add matters more than how many.
Key Takeaways
Adding 15 plasma proteins to a clinical type 2 diabetes risk score raised the C-index from 0.862 to 0.884 in independent validation.
Multi-omics integration of proteins, metabolites and a polygenic score reached a C-index of 0.891, only 0.007 above the proteomics-only model.
Proteomics carried nearly all of the added signal, so panel size and assay choice matter more here than the number of omics layers.
How the Multi-Omics Model Was Built and Validated
The design sidesteps the usual overfitting trap. Participants were split by UK Biobank metabolomics release phase into a derivation set of 23,108 and a non-overlapping validation set of 19,732. Proteins came from the Olink Explore 3072 proximity extension assay, filtered from 2,923 targets to 2,085 usable ones, with 15 chosen by LASSO regression with bootstrap resampling; metabolites came from the Nightingale Health NMR platform, 249 measures narrowed to 11. Each layer was stacked onto a recalibrated Cambridge Diabetes Risk Score and scored by Harrell's C-index. Carrasco-Zanini and colleagues (2024) showed this combination working in EPIC-Norfolk, at a fortieth of the scale.
Key Findings
Proteomics dominated the single-layer comparison: the 15-protein panel lifted the C-index by 0.022 (P < 0.001) with a continuous NRI of 42.0%, against 0.008 for the polygenic score and 0.006 for metabolites.
Stacking all three layers helped, barely: the full model reached 0.891, a significant but clinically slim gain of 0.007 over the proteomics-extended version.
The layers were close to independent: polygenic score correlations stayed at or below 0.11, and protein-metabolite correlations all sat under 0.5.
Strongest single associations came from the proteome: IGSF9 reached a hazard ratio of 1.49 per standard deviation (95% CI 1.37 to 1.61), ahead of the polygenic score at 1.36.
Figure 1. Spearman correlation matrix of the selected multi-omics biomarkers and the type 2 diabetes polygenic risk score in the validation set (N = 19,732). Metabolites appear in blue, proteins in orange, the polygenic score in purple. Cross-layer correlations are weak to moderate, with one strong within-layer pair (medium-LDL triglyceride percentage and linoleic acid percentage, r = -0.70). Adapted from Xie et al. (2026), Cardiovascular Diabetology.
Why Protein Profiling Carried the Signal
Tested one at a time, 13 proteins produced a significant C-index gain against only three metabolites. IGSF9, hepatocyte growth factor and cadherin-2 each added at least 0.008 alone, more than the entire NMR panel managed together. Coverage is the likely reason. Targeted NMR quantifies mostly lipoprotein subfractions and fatty acids, while a 2,000-plex proximity extension assay reaches inflammatory and adhesion pathways a lipid-heavy panel never sees.
What This Does Not Yet Prove
A 0.007 C-index gain is real in 19,732 people and close to meaningless at the bedside. The authors say as much, recommending the proteomics-only extension as the translatable option. Both panels were also selected in earlier UK Biobank analyses by this same group, which makes the study an internal replication rather than a clean external test. Samples were mostly non-fasting too, probably handicapping the metabolite panel from the start.
One Sample, Two Omics Layers
Proteins and metabolites here carried largely independent information, with every cross-layer correlation below 0.5, and that independence is exactly the condition under which stacking layers earns its keep. Splitting one biobank aliquot across two platforms is where the advantage usually leaks away, because sample volume and plate-to-plate batch structure start driving variance no correction fully removes. The Omni-MS workflow at Dalton was built around single-injection multi-omics analysis for that reason, keeping proteins, metabolites and lipids on one sample and one batch, which is the starting point we describe in our guide to multi-omics biomarker discovery.
Frequently Asked Questions
What is multi-omics used for in disease risk prediction?
Multi-omics combines two or more molecular layers, such as proteomics, metabolomics and genomics, to estimate who will develop a disease. Each layer captures different biology, so a combined model separates high-risk from low-risk people more sharply than any single layer. Here it reached a C-index of 0.891 for ten-year type 2 diabetes risk.
Does adding proteomics to a clinical risk score improve type 2 diabetes prediction?
Yes, and by a wider margin than genomics or metabolomics. A 15-protein plasma panel raised the C-index of the Cambridge Diabetes Risk Score from 0.862 to 0.884, with 42.0% continuous net reclassification. Genetic and metabolite layers each added less than half that gain.
How many samples does a multi-omics biomarker study need?
Event count drives power, not headcount. This cohort needed 42,840 participants to accumulate 1,090 incident cases over ten years, because diabetes incidence in healthy volunteers is low. Case-control designs in enriched populations reach usable power with a few hundred samples per arm.
Conclusion
Plasma protein profiling adds real predictive value to an established clinical diabetes score, enough to justify prospective testing. What isn't established is that three omics layers beat one in any way a clinician would act on. For biomarker teams, the practical read is to fund the proteomic arm first and treat genomics and metabolomics as mechanistic context.
Related Reading
See how we run these analyses in one lab: Dalton's multi-omics CRO services.
Citation
Xie, R., Herder, C., & Schöttker, B. (2026). Large-scale multi-omics enhance risk prediction for type 2 diabetes. Cardiovascular Diabetology, 25(1), 166. https://doi.org/10.1186/s12933-026-03223-y
Note
This blog post summarizes findings from the above-cited research. Figures are adapted from the original publication. For full details, please refer to the source article.
By Seungjun Yeo, CEO at Dalton Bioanalytics. Specializing in multi-omics mass spectrometry for drug discovery and biomarker research.