Circulating miRNAs as Biomarkers for CYP2B6 Enzyme Activity

Abstract

The CYP2B6 gene is highly polymorphic and its activity shows wide interindividual variability. However, substantial variability in CYP2B6 activity remains unexplained by the known CYP2B6 genetic variations. Circulating, cell-free micro RNAs (miRNAs) may serve as biomarkers of hepatic enzyme activity. CYP2B6 activity in 72 healthy volunteers was determined using the disposition of efavirenz as a probe drug. Circulating miRNA expression was quantified from baseline plasma samples. A linear model consisting of the effects of miRNA expression, genotype-determined metabolizer status, and demographic information was developed to predict CYP2B6 activity. Expression of 2,510 miRNAs were quantified out of which 7 miRNAs, together with the CYP2B6-genotypic metabolizer status and demographics, was shown to be predictive markers for CYP2B6 activity. The reproducibility of the model was evaluated by cross-validation. The average Pearson's correlation (R) between the predicted and observed maximum plasma concentration (C_max ) ratios of efavirenz and its metabolite-8-OH efavirenz using the linear model with all features (7 miRNA + metabolizer status + age + sex + race) was 0.6702. Similar results were also observed using area under the curve (AUC) ratios (Pearson correlation's R = 0.6035). Thus, at least 36% (R² ) of the variability of in vivo CYP2B6 activity was explained using this model. This is a significant improvement over the models using only the genotype-based metabolizer status or the demographic information, which explained only 6% or less of the variability of in vivo CYP2B6 activity. Our results, therefore, demonstrate that circulating plasma miRNAs can be valuable biomarkers for in vivo CYP2B6 activity.

Figure 1.

Workflow of modeling showing data partition, supervision and feature selection followed by model evaluation.

Figure 2:. Results of numerical experiments on supervision cutoffs

Values in the Y-axis are the performance scores (Pearson’s correlation) of the models. X- axis denotes the maximum number of features (N) included in the model. A- E denotes p-value cutoffs (P) of 0.05, 0.01, 0.005, 0.001, and 0.0005. The plots on the left show the performance of the training model, and the model evaluation on unused data is shown on the right. The P of 0.005 and N of 30 resulted in a learning curve with the highest stable score on the training set; and showed optimal performance on the testing set.

Figure 3.. Forest plot of the selected miRNAs with effect sizes on CYP2B6 activity.

Effect sizes (estimated) of the 7 miRNAs included in the model are shown with 95% confidence intervals.

Figure 4.. Performance of prediction model.

Performance is scored as the Pearson’s correlation between predicted and observed values, the results of 200 randomized tests on the full feature model (7 miRNAs + metabolizer status + age + sex + race), metabolizer status-only, demographic only, and both metabolizer status and demographic are shown. (A): Training score -performance of the full feature model and other models on the training set. Average level for full feature model = 0.73 (std. error of mean = 0.004), metabolizer status-only model = 0.29 (std. error of mean = 0.005), demographics only model = 0.24 (std. error of mean = 0.004), and the metabolizer status + demographics model = 0.36 (std. error of mean = 0.005). (B): Testing score -performance of the full feature model and other models on the testing set. Average level for full feature model = 0.67 (std. error of mean = 0.009), metabolizer status-only model = 0.24 (std. error of mean = 0.011), demographics only model = 0.13 (std. error of mean = 0.009), and the metabolizer status + demographics model = 0.21 (std. error of mean = 0.01). Bold line indicates the median, the box spans the lower (1^st) and upper (3^rd) quartile, and the whisker bars extent to the minimum and maximum values. Data points marked by “+” extending beyond the whiskers were the outliers. As shown, (average) training and testing scores are similar, indicating no or negligible overfitting. Additionally, most of the scores fall in the specific range marked by the boxplots, suggesting robust model performance. The small standard errors indicate accurate estimation of the means, i.e. most scores stably reside in a narrow range.

Figure 5.. Model performance showing correlation between predicted and experimental values.

One example (from 200 randomizations) of correlation between CYP2B6 activity predicted by the models and experimental values in the testing data, using an instance from the 200 randomizations that was closest to the average performance. A) Full feature model, B) Metabolizer status only, C) Demographics only, and D) Metabolizer status + Demographics model. Dashed line is the line of identity and the solid black line is the regression line.