Plasma Protein Signatures Predict Rheumatoid Arthritis Years Before Symptoms Appear

Rheumatoid arthritis is usually diagnosed only once joints become painful and swollen, but the immunological shifts that drive the disease appear to be measurable in blood long before the first clinical signs. A new study published in the Journal of Translational Medicine has combined population-scale plasma proteomics with machine learning to build a risk prediction model that identifies people likely to develop rheumatoid arthritis up to a decade or more before diagnosis. Drawing on 47,537 participants from the UK Biobank with a median follow-up of 14.27 years, the researchers profiled 2,920 plasma proteins at baseline using the Olink proximity extension assay and tracked who went on to develop the disease.

Using Cox proportional hazards regression, the team identified 213 plasma proteins significantly associated with the onset of rheumatoid arthritis. Time-to-diagnosis-aligned clustering, which groups participants according to how soon after blood sampling they received a diagnosis, revealed persistent molecular dysregulation across the long preclinical phase. In other words, the protein signatures did not flicker briefly before onset; they showed sustained alterations stretching as far as 15 years ahead of clinical diagnosis, suggesting a prolonged window in which the systemic immune landscape is already shifting.

To probe which of these altered proteins might actually contribute to disease rather than merely mark it, the researchers performed Mendelian randomization, a technique that uses naturally occurring genetic variants as instruments to infer causal relationships. Cis-Mendelian randomization analysis pointed to two proteins with apparently opposing causal roles: higher genetically predicted levels of ICAM3 were associated with increased odds of rheumatoid arthritis (odds ratio 1.283, 95 percent confidence interval 1.185 to 1.390), while higher levels of IL18BP, a natural binding protein for the inflammatory messenger interleukin-18, were associated with reduced odds (odds ratio 0.635, 95 percent confidence interval 0.445 to 0.885). The authors caution that these are inferential estimates, but they flag ICAM3 and IL18BP as candidate drivers meriting experimental follow-up.

The translational payoff came from drug-target matching. Cross-referencing the RA-associated proteins against known pharmacological targets, the team identified 112 targetable genes corresponding to 580 approved drugs, creating a shortlist for potential drug repurposing. This mapping does not establish that any existing medication prevents rheumatoid arthritis, but it narrows the search space for preventive candidates whose safety profiles in other indications are already documented.

The centerpiece of the study is a dual-branch residual DeepSurv model, a deep learning architecture that fuses proteomic measurements with routine clinical features to estimate an individual’s risk of developing RA over time. When validated in an independent held-out cohort of 9,446 participants, the integrated model achieved a mean C-index of 0.786, compared with 0.736 for a clinical-only model and 0.738 for a protein-only model. In the development cohort the pattern was similar, with the integrated model reaching 0.777 against 0.723 and 0.737 respectively. The fusion of data types, rather than either data type alone, was what delivered the performance gain.

Time-dependent performance metrics showed the model was most accurate closest to diagnosis. The area under the curve reached 0.915 at three years before onset, 0.867 at five years, and 0.810 at ten years in the validation cohort. When participants were split into risk groups, those flagged as high risk by the model had a Kaplan-Meier hazard ratio of 3.45 (95 percent confidence interval 2.86 to 4.16) for developing rheumatoid arthritis compared with the low-risk group, a more than threefold difference in observed incidence.

Specificity testing addressed an obvious concern: could the model simply be detecting general inflammation or joint disease of any kind? The answer appears to be no. The model’s C-index for osteoarthritis, a mechanically driven degenerative condition with a different biology, was 0.627, significantly lower than its 0.738 for rheumatoid arthritis. That gap indicates the protein signature the model relies on is genuinely discriminative for rheumatoid arthritis rather than a nonspecific marker of joint complaints.

The clinical context matters. Most patients today are diagnosed before radiographic erosions develop, yet routine clinical and laboratory assessment still fails to capture the systemic immunological alterations that precede clinical onset, leaving management largely reactive. A validated preclinical risk stratification tool could, in principle, allow clinicians to concentrate monitoring and, eventually, preventive trials on the small fraction of the population at highest risk, much as cardiovascular risk scores reshaped the prevention of heart disease.

Several caveats temper the enthusiasm. The proteomic profile was measured only once, at baseline, so the analysis cannot track within-person changes over the preclinical period. Mendelian randomization rests on assumptions about genetic instruments that can be violated, and the odds ratios, while statistically robust, describe modest effects. The model was developed and validated within the UK Biobank, a largely middle-aged British cohort, and external validation in other populations and health systems remains to be demonstrated. The authors themselves frame the work as establishing selectable options for risk stratification, early screening, and drug repurposing rather than as a ready-made screening program.

Even with those qualifications, the study is among the most comprehensive attempts yet to chart the preclinical proteomic landscape of rheumatoid arthritis at population scale. It demonstrates that a single baseline blood draw, interpreted through a multimodal machine learning framework, can carry predictive information about a disease that may not manifest for another ten to fifteen years. If the findings replicate, the combination of prospective proteomics, causal inference, and deep learning could become a template for anticipating other autoimmune conditions whose silent phase has, until now, remained hidden from routine medicine.

Subject of Research: Prospective plasma proteomics and a multimodal deep learning model for preclinical risk prediction of rheumatoid arthritis in the UK Biobank cohort

Article Title: Multimodal fusion prediction model based on prospective plasma proteomics: preclinical risk prediction and key pathway analysis of rheumatoid arthritis

Article References: Wang, T., Wang, L., Zhang, H., Lin, W., Yan, H., Xu, Y., Wang, M., Shen, H., Wen, C., Wang, Z., & Huang, L. (2026). Multimodal fusion prediction model based on prospective plasma proteomics: preclinical risk prediction and key pathway analysis of rheumatoid arthritis. Journal of Translational Medicine. https://doi.org/10.1186/s12967-026-09061-6

Image Credits: AI Generated

DOI: 10.1186/s12967-026-09061-6

Keywords: rheumatoid arthritis, proteomics, UK Biobank, Mendelian randomization, deep learning, predictive modeling, autoimmune disease, drug repurposing, Multimodal, fusion, prediction, model