A Validated, Modifiable Proteomic Score from the EXSCEL Trial Predicts Cardiovascular Events in Diabetes

Figures

• 

  Figure 1. Study overview and analytic workflow. Overview of the EXSCEL proteomics subcohort, proteomic profiling, analytic framework, and external validation strategy. A randomized EXSCEL subcohort with available plasma samples underwent SomaScan 5K proteomic profiling at baseline and 12 months. Baseline proteins were evaluated for association with MACE, and prioritized proteins were incorporated into supervised machine learning models to develop a 140-protein risk score. Additional analyses included treatment responsiveness, biological contextualization analyses, and external validation in the CHS and PROMISE cohorts. MACE included nonfatal myocardial infarction, nonfatal stroke, and cardiovascular death. ASCVD, atherosclerotic cardiovascular disease; CHS, Cardiovascular Health Study; EXSCEL, Exenatide Study of Cardiovascular Event Lowering; MACE, major adverse cardiovascular events; PROMISE, Prospective Multicenter Imaging Study for Evaluation of Chest Pain; RCT, randomized controlled trial; T2DM, type 2 diabetes mellitus.

• 

  Figure 2. Receiver operating characteristic curves for proteomic and clinical prediction models in the EXSCEL proteomics cohort. (A) Receiver operating characteristic (ROC) curves for proteomic machine learning models (LASSO, elastic net, random forest, and gradient-boosted decision trees) for discrimination of MACE. (B) ROC curves for protein-only, clinical-only, and protein plus clinical models. Clinical features were derived from the SCORE2, ADVANCE, and ASCVD Risk Estimator models. All ROC curves were evaluated in the held-out test set of the EXSCEL proteomics cohort.

• 

  Figure 3. Volcano plot demonstrating model-prioritized MACE proteins in EXSCEL proteomics cohort. Multivariate Cox proportional hazards models identified 1,278 individual proteins that were significantly associated with MACE (FDR adjusted P < 0.05) in the EXSCEL proteomics cohort (gray); 140 of these proteins were prioritized in the LASSO model (red). The 2 proteins with the absolute highest LASSO model weights were tetranectin and fibulin-1.

• 

  Figure 4. Differential change in MACE proteomics score with GLP-1 RA treatment in EXSCEL. Linear mixed model adjusted for change in BMI and HbA1c demonstrated differential change in the MACE protein score with GLP-1 RA treatment over 12 months (P = 0.005).

• 

  Figure 5. Kaplan-Meier survival curves comparing tertiles of the MACE protein score in EXSCEL and CHS cohorts. Survival curves demonstrate multivariate Cox proportional hazard models for time to MACE and tertiles of protein score representing high, medium, and low risk of MACE across the 2 cohorts.