Plasma proteomic signatures for type 2 diabetes mellitus and related traits in the UK Biobank cohort

Abstract

Aims/hypothesis: The plasma proteome holds promise as a diagnostic and prognostic tool that can accurately reflect complex human traits and disease processes. We assessed the ability of plasma proteins to predict type 2 diabetes mellitus (T2DM) and related traits.

Methods: Clinical, genetic, and high-throughput proteomic data from three subcohorts of UK Biobank participants were analyzed for association with dual-energy x-ray absorptiometry (DXA) derived truncal fat (in the adiposity subcohort), estimated maximum oxygen consumption (VO₂max) (in the fitness subcohort), and incident T2DM (in the T2DM subcohort). We used least absolute shrinkage and selection operator (LASSO) regression to assess the relative ability of non-proteomic and proteomic variables to associate with each trait by comparing variance explained (R²) and area under the curve (AUC) statistics between data types. Stability selection with randomized LASSO regression identified the most robustly associated proteins for each trait. The benefit of proteomic signatures (PSs) over QDiabetes, a T2DM clinical risk score, was evaluated through the derivation of delta (Δ) AUC values. We also assessed the incremental gain in model performance metrics using proteomic datasets with varying numbers of proteins. A series of two-sample Mendelian randomization (MR) analyses were conducted to identify potentially causal proteins for adiposity, fitness, and T2DM.

Results: Across all three subcohorts, the mean age was 56.7 years and 54.9% were female. In the T2DM subcohort, 5.8% developed incident T2DM over a median follow-up of 7.6 years. LASSO-derived PSs increased the R² of truncal fat and VO₂max over clinical and genetic factors by 0.074 and 0.057, respectively. We observed a similar improvement in T2DM prediction over the QDiabetes score [Δ AUC: 0.016 (95% CI 0.008, 0.024)] when using a robust PS derived strictly from the T2DM outcome versus a model further augmented with non-overlapping proteins associated with adiposity and fitness. A small number of proteins (29 for truncal adiposity, 18 for VO₂max, and 26 for T2DM) identified by stability selection algorithms offered most of the improvement in prediction of each outcome. Filtered and clustered versions of the full proteomic dataset supplied by the UK Biobank (ranging between 600-1,500 proteins) performed comparably to the full dataset for T2DM prediction. Using MR, we identified 4 proteins as potentially causal for adiposity, 1 as potentially causal for fitness, and 4 as potentially causal for T2DM.

Conclusions/interpretation: Plasma PSs modestly improve the prediction of incident T2DM over that possible with clinical and genetic factors. Further studies are warranted to better elucidate the clinical utility of these signatures in predicting the risk of T2DM over the standard practice of using the QDiabetes score. Candidate causally associated proteins identified through MR deserve further study as potential novel therapeutic targets for T2DM.

Figure 1.. Study design and analysis workflow

Abbreviations: DXA: dual-energy x-ray absorptiometry, T2DM: type 2 diabetes mellitus, T1DM: type I diabetes mellitus, A1c: hemoglobin A1c, VO₂max: maximal oxygen consumption

Figures 2a-b.. Variance explained (R²) of clinical variables, polygenic scores, and plasma proteins in the adiposity and fitness subcohorts

Footnote: (a) Models performed using training and test sets in the adiposity subcohort. (b) Models performed using training and test sets in the fitness subcohort. SS proteins shown in red refer to proteins selected by a randomized LASSO regression model with stability selection algorithm. Abbreviations: LASSO: least absolute shrinkage and selection operator, R²: variance explained, PRSs: polygenic risk scores, SS: stability selection

Figure 3a-c.. Proteins selected by a randomized LASSO with stability selection algorithm for adiposity, fitness, and type 2 diabetes mellitus

Footnote: (a) Proteins selected by RLSS algorithm in the adiposity subcohort. (b) Proteins selected by RLSS algorithm in the fitness subcohort. (c) Proteins selected by RLSS algorithm in the T2DM subcohort. Proteins listed in blue were positively associated with the outcome of interest while those listed in orange were negatively associated. Abbreviations: RLSS: Randomized LASSO regression with stability selection algorithm.

Figures 4a-d.. Area under the curves (AUCs) of clinical variables, polygenic scores, and plasma proteins in the type 2 diabetes mellitus subcohort

Footnote: (a, b) Models performed in training and test sets of the T2DM subcohort using the full proteomic dataset. (c, d) Models performed in training and test sets of the T2DM subcohort using proteins selected by a randomized LASSO regression model with a stability selection algorithm. Colors shown on the receiver operating curves correspond with colors shown on the forest plots. Abbreviations: LASSO: least absolute shrinkage and selection operator, T2DM: type 2 diabetes mellitus, AUC: area under the curve, PRS: polygenic risk scores, SS: stability selection

Figure 5a-c.. Potentially causal proteins for adiposity, fitness, and type 2 diabetes mellitus identified through two-sample Mendelian randomization

Footnote: (a) Forest plot of potentially causal proteins for truncal adiposity based on FDR < 0.5. (b) Forest plot of a potentially causal protein for cardiorespiratory fitness based on FDR < 0.5. (c) Forest plot of potentially causal proteins for T2DM based on FDR < 0.5. Abbreviations: SNP: single nucleotide polymorphism, 95% CI: 95% confidence interval, T2DM: type 2 diabetes mellitus