Pharmacometabolomic approach to predict QT prolongation in guinea pigs

Abstract

Drug-induced torsades de pointes (TdP), a life-threatening arrhythmia associated with prolongation of the QT interval, has been a significant reason for withdrawal of several medicines from the market. Prolongation of the QT interval is considered as the best biomarker for predicting the torsadogenic risk of a new chemical entity. Because of the difficulty assessing the risk for TdP during drug development, we evaluated the metabolic phenotype for predicting QT prolongation induced by sparfloxacin, and elucidated the metabolic pathway related to the QT prolongation. We performed electrocardiography analysis and liquid chromatography-mass spectroscopy-based metabolic profiling of plasma samples obtained from 15 guinea pigs after administration of sparfloxacin at doses of 33.3, 100, and 300 mg/kg. Principal component analysis and partial least squares modelling were conducted to select the metabolites that substantially contributed to the prediction of QT prolongation. QTc increased significantly with increasing dose (r = 0.93). From the PLS analysis, the key metabolites that showed the highest variable importance in the projection values (>1.5) were selected, identified, and used to determine the metabolic network. In particular, cytidine-5'-diphosphate (CDP), deoxycorticosterone, L-aspartic acid and stearic acid were found to be final metabolomic phenotypes for the prediction of QT prolongation. Metabolomic phenotypes for predicting drug-induced QT prolongation of sparfloxacin were developed and can be applied to cardiac toxicity screening of other drugs. In addition, this integrative pharmacometabolomic approach would serve as a good tool for predicting pharmacodynamic or toxicological effects caused by changes in dose.

Figure 1. Chemical structure of sparfloxacin and its metabolic pathway.

Figure 2. Guinea pig electrocardiograms before (A) and after (B) the administration of sparfloxacin.

Figure 3. Mean plasma concentration and mean increase in QTc (%) over time according to sparfloxacin doses.

(A) Mean plasma concentration of sparfloxacin following a single 1-h intravenous dose of 33.3 mg kg⁻¹, 100 mg kg⁻¹, or 300 mg kg⁻¹. (B) Mean increase in QTc (%) following a single intravenous dose of 33.3 mg kg⁻¹, 100 mg kg⁻¹, or 300 mg kg⁻¹. The percentage QT increase was less in the group dosed with 100 mg kg⁻¹ than that with 33.3 mg kg⁻¹ after 1 h. Bars indicate standard deviations.

Figure 4. PCA and PLS modelling of plasma LC–MS metabolic data for predicting the drug-induced QT prolongation of sparfloxacin.

(A) PCA score plot (t vs. t[2]) obtained from guinea pig plasma samples. Obviously separated clustering of dose groups and the control group was shown by PCA; in addition, dose-dependent metabolomic modification was detected. (B) Loading plot for the above PLS model in which each point represents a metabolic feature detected from plasma LC–MS data and is plotted as its respective coefficient from PLS component 1 vs. its coefficient from PLS component 2. The arrow indicates a positive relationship with the QTc. Metabolite variables with larger coefficient values (positive or negative) have a stronger correlation with the QTc (marked by red boxes; VIP>1.5) and were used to build the PLS model for predicting cardiovascular toxicity. The inset green bar plot shows the correlation coefficients for the key identified metabolites.

Figure 5. PLS model validity.

(A) Plot of predicted QTc vs. actual (measured) QTc from the PLS model using the cross-validation method. Predicted values from the PLS model in which all predicted QTc values show a linear relationship with actual measured QTc values (R² = 0.9884). Colour from blue to red indicates increasing QTc values. RMSEE specifies the root mean square error of the estimation (the fit) for observations in the workset. The values were predicted by exclusion of 1/7^th of the data from the model and predicting the excluded data that are not part of model building. (B) Internal validation of the PLS model by 20 permutation tests to confirm predictability and data overfitting shows that all R² (goodness of fit) and Q² (predictability of model) values from the permuted models (left) are smaller than those of the original model (far right), demonstrating the validity of the PLS model. (C) Internal validation of the PLS model with 100 permutation tests to use stricter validation criteria. (D and E) Plots for normalised intensities of LysoPC (18∶1) (D) and L-aspartic acid (E), which exhibit a negative and positive correlation, respectively, with QTc.

Figure 6. Metabolic network for 15 identified metabolites.

The large nodes in the network represent the key identified metabolites, while the small nodes represent their neighbours in the respective metabolic pathways (see Table 1). Metabolic reactions (arrow) and indirect or possible reactions involving several intermediates between the connected nodes are indicated. This metabolic network revealed six major modules, shown in different colours. All abbreviations used for enzymes or genes and reactions are from KEGG identifiers (http://www.genome.jp/kegg/kegg3.html).

Figure 7. Comparison of the distribution of metabolite intensity levels for control and drug-dosed (low, middle, high) groups.

Box plots indicate the distribution of magnitudes of peak intensity levels of key metabolic phenotype in each group. The box is drawn from the 25^th to 75^th percentiles in the distribution of intensities. The median, or 50^th percentile, is drawn as a black horizontal line inside the box. The whiskers (lines extending from the box) describe the spread of the data within the 10^th and 90^th percentiles.

Figure 8. Scatterplot of the predicted normalised QTc values (QTc_norm) from the equation QTc_norm = 0.537(LA)+0.533(CDP) – 0.431(DC) – 0.640(SA), versus the measured and normalised (QTc_norm) values for the 15 samples.

Using this prediction index with only four metabolite abundances, subjects can be categorised into low, medium, and high QTc groups.