Abstract 13312: Statin Signature: Using Proteomics to Detect Pharmacological Fingerprints

Introduction: Lowering low-density lipoprotein cholesterol with statin therapy is a primary strategy for reducing cardiovascular morbidity and mortality. Yet, adherence to and persistence with statin medication is generally low, complicating the ability to evaluate whether treatment is effective.

Hypothesis: Statin treatment may yield significant differences in abundances of proteins detectable in blood samples. Using machine learning and high-throughput proteomics, a blood-based biomarker test can be developed to assess statin medication compliance and demonstrate drug fingerprinting useful for clinical trials.

Methods: Using modified-aptamer proteomics technology, SomaScan® assay v4.0, we assessed ∼5,000 proteins in 8,395 EDTA plasma samples from individuals aged 29-64 at visit 1 of the Fenland study, totaling ∼42 million protein measurements. A total of 305 individuals (3.6%) reported active statin medication use at this study visit. Predicted statin use was modeled with protein measurements using machine learning methods in 70% of Fenland as a training dataset. A hold-out dataset was used to assess model performance. A predictive model using elastic net logistic regression was optimized based on AUC and robustness to assay noise and sample handling conditions to detect a signature of statin usage.

Results: A total of 839 proteins differed significantly between the ‘active statin’ and ‘no statin’ use groups in univariate analysis (FDR <0.01). Eight of the top 50 significant proteins have known mechanistic associations with statin pharmacology, including HMGCS1, PCSK9, and APOB. A six-protein model was developed with an AUC of 0.91 (sensitivity=0.82, specificity=0.88) and 0.90 (sensitivity=0.80, specificity=0.87) on the training and hold-out datasets, respectively, to predict whether a statin signature is present or not.

Conclusions: We successfully developed a blood-based protein-only model that detects mechanistically-relevant protein changes to predict active statin use in adults. The protein model could compliment self-reported medication status in clinical trials and in healthcare.