Statin effects on the lipidome: Predicting statin usage and implications for cardiovascular risk prediction

Abstract

Statin therapy is a highly successful and cost-effective strategy for the prevention and treatment of cardiovascular diseases (CVD). Adjusting for statin usage is crucial when exploring the association of the lipidome with CVD to avoid erroneous conclusions. However, practical challenges arise in real-world scenarios due to the frequent absence of statin usage information. To address this limitation, we demonstrate that statin usage can be accurately predicted using lipidomic data. Using three large population datasets and a longitudinal clinical study, we show that lipidomic-based statin prediction models exhibit high prediction accuracy in external validation. Furthermore, we introduce a re-weighted model, designed to overcome a ubiquitous limitation of prediction models, namely the need for predictor alignment between training and target data. We demonstrated that the re-weighted models achieved comparable prediction accuracy to ad hoc models which use the aligned predictor between training and target data. This innovation holds promise for significantly enhancing the transferability of statin prediction and other 'omics prediction models, especially in situations where predictor alignment is incomplete. Our statin prediction model now allows for the inclusion of statin usage in lipidomic analyses of cohorts even where statin use is not available, improving the interpretability of the resulting analyses.

Keywords: cardiovascular disease; lipidomics; regularized linear models; statins.

Fig. 1

Association of statin usage with LIPID trial lipid species using linear regression. Linear regression analysis between each 12-months lipid species concentration and statin usage was performed while adjusting for age, sex, BMI, hypertension medication, current smoking, blood pressure, baseline lipid concentration, interaction between baseline lipid concentration, statin usage and clinical lipids (total cholesterol, HDL-C, triglycerides) from baseline (A) as well as clinical lipids from 12-months follow-up (B). (C) Boxplots (center line, median; box limits, upper and lower quartiles; vertical line, 1.5x interquartile range; points, outliers) showing natural log-transformed concentrations of top three negatively associated lipids and free cholesterol in baseline and 12-months follow-up. Grey empty circles show species with corrected P > 0.05, dark-green circles show species with corrected P < 0.05, yellow circles show species with corrected p < 1e-20, and purple circles represent top five species with lowest P-values. Whiskers represent 95% confidence intervals. The underlying data can be found in supplemental Table S3. AC, Acylcarnitine; CE, cholesteryl ester; Cer(d), ceramide; Cer(m), Deoxyceramide; COH, free cholesterol; DE, dehydrocholesterol; DG, diacylglycerol; dhCer, dihydroceramide; GM1, GM1 ganglioside; GM3, GM3 ganglioside; HexCer, monohexosylceramide; Hex2Cer, dihexosylceramide; Hex3Cer, trihexosylceramide; LPC, lysophosphatidylcholine; LPC(O), lysoalkylphosphatidylcholine; LPC(P), lysoalkenylphosphatidylcholine; LPE, lysophosphatidylethanolamine; LPE(P), lysoalkenylphosphatidylethanolamine; LPI, lysophosphatidylinositol; NL, neutral loss; PC, phosphatidylcholine; PC(O), alkylphosphatidylcholine; PC (P), alkenylphosphatidylcholine; PE, phosphatidylethanolamine; PE(O), alkylphosphatidylethanolamine; PE(P), alkenylphosphatidylethanolamine; PG, phosphatidylglycerol; PI, phosphatidylinositol; PS, Phosphatidylserine; SM, sphingomyelin; SIM, single ion monitoring; S1P, sphingosine-1-phosphate; SHexCer, Sulfatide; TG, triacylglycerol; TG(O), alkyl-diacylglycerol.

Fig. 2

Effect of statin usage adjustment on the association of prevalent CVD with AusDiab lipid species. Logistic regressions were conducted to assess the association of prevalent CVD with lipid concentration while adjusting for age, sex, BMI, hypertension medication, diabetes history, current smoking, blood pressure (A) and the same covariates with statin usage (B). Grey empty circles show species with corrected P > 0.05, blue circles show species with corrected P < 0.05, yellow circles show species with corrected p < 1e-6. Scatter plot of log odds ratio (C) and minus log-transformed corrected P values (D) of each lipid with and without adjusting statin usage, dot color indicates the group of P values of each lipid as shown in the bottom table. The number of lipids in each P value group were given with lipid showing different p group were highlighted with colors and underline. The dashed line represents the line of identity. CI, confidence interval. AC, Acylcarnitine; CE, cholesteryl ester; Cer(d), ceramide; Cer(m), Deoxyceramide; COH, free cholesterol; DE, dehydrocholesterol; DG, diacylglycerol; dhCer, dihydroceramide; GM1, GM1 ganglioside; GM3, GM3 ganglioside; HexCer, monohexosylceramide; Hex2Cer, dihexosylceramide; Hex3Cer, trihexosylceramide; LPC, lysophosphatidylcholine; LPC(O), lysoalkylphosphatidylcholine; LPC(P), lysoalkenylphosphatidylcholine; LPE, lysophosphatidylethanolamine; LPE(P), lysoalkenylphosphatidylethanolamine; LPI, lysophosphatidylinositol; NL, neutral loss; PC, phosphatidylcholine; PC(O), alkylphosphatidylcholine; PC (P), alkenylphosphatidylcholine; PE, phosphatidylethanolamine; PE(O), alkylphosphatidylethanolamine; PE(P), alkenylphosphatidylethanolamine; PG, phosphatidylglycerol; PI, phosphatidylinositol; PS, Phosphatidylserine; SM, sphingomyelin; SIM, single ion monitoring; Sph, sphingosine; S1P, sphingosine-1-phosphate; SHexCer, Sulfatide; TG, triacylglycerol; TG(O), alkyl-diacylglycerol.

Fig. 3

Development of a re-weighted statin usage prediction model with high accuracy, robusticity and transferability. Ridge models and re-weighted models were built using AusDiab lipidome to predict statin usage in BHS, SAFHS and LIPID. AUC plots show similar prediction accuracy in the ridge models (A) and re-weighted models (B). The re-weighted models show higher AUC than ridge models when randomly removing lipids from training and validation cohorts (C–E). When randomly removing lipids, 10 iterations were performed, and the mean and standard deviation were shown. The dashed line represents the line of identity.

Fig. 4

Comparison of the re-weighted models with re-developed ridge models after removing lipids. The re-weighted models were developed using all AusDiab lipids while adjusting for age and sex, and then used to predict statin usage in BHS (A), SAFHS (B) and LIPID (C) cohorts. The ridge model was re-developed each time after randomly removing lipids. When randomly removing lipids, 10 iterations were performed and the mean and standard deviation were shown. The dashed line represents the line of identity.