Using genetic evidence and Bayesian shrinkage methods to discover proteomics biomarkers for the early diagnosis of autoimmune rheumatic diseases

Background

Autoimmune rheumatic diseases (ARDs) are diseases where the body’s immune system mistakenly attacks its own tissues, commonly the joints, causing chronic inflammation, pain, and fatigue.  Early effective treatment can provide substantial long-term benefits by preventing or slowing down progression to irreversible tissue damage and disability in affected individuals. However, diagnosing ARDs early is challenging. Around 30% of all GP consultations present with non-specific musculoskeletal symptoms. Among these, only about 1 in 40 patients has an ARD, therefore, referral to rheumatology is typically delayed until more specific symptoms have developed. This postpones access to treatment, leaving the damaging inflammatory process uncontrolled.

The aim of this project is to develop proteomics-based models for distinguishing individuals with ARDs from those without in the early stages of the disease, and to identify a sparse set of proteins suitable for inclusion in a diagnostic biomarker panel. We will analyse high-throughput proteomic data from the UK Biobank, a large-scale population cohort study that has measured 3,000 circulating proteins in 50,000 participants. Measurements of 5,000 proteins in all 500,000 participants are currently under way and are expected to be available to researchers by Q1 of 2027. Adding proteomic measurements to standard clinical covariates has been shown to improve prediction of incident disease for several conditions¹. For developing useful diagnostic panels for ARDs, a key challenge is how to select the smallest number of proteins that are most predictive, given the relatively low numbers of cases and the high dimensionality of the proteomics data. 

We will address this challenge in three ways: 

1. Exploit recent advances in Bayesian computation to learn sparse predictive models, where most biomarker have no effect, using hierarchical shrinkage priors on the effect sizes². This approach overcomes limitations of older methods such as LASSO – large effects are not shrunk, and cross-validation is not required to learn the penalty parameters. Cross-validation is required only in a final step to evaluate predictive performance on data not seen before. We will combine this approach with projection predictive variable selection, a method that automatically selects the most informative subset of variables from the full model by quantifying the percent of predictive information gained by each additional variable.

2. Exploit our recent findings on putative core genes for rheumatoid arthritis³, systemic lupus erythematosus⁴, and other ARDs, to prioritise proteins that are likely to have a causal role in these diseases.

3. Learn proteomics models to classify prevalent disease cases and controls and use findings from these models to set informative prior distributions in the models for incident disease.