Novel Two-Step Classifier for Torsades de Pointes Risk Stratification from Direct Features

While pre-clinical Torsades de Pointes (TdP) risk classifiers had initially been based on drug-induced block of hERG potassium channels, it is now well established that improved risk prediction can be achieved by considering block of non-hERG ion channels. The current multi-channel TdP classifiers can be categorized into two classes. First, the classifiers that take as input the values of drug-induced block of ion channels (direct features). Second, the classifiers that are built on features extracted from output of the drug-induced multi-channel blockage simulations in the in-silico models (derived features). The classifiers built on derived features have thus far not consistently provided increased prediction accuracies, and hence casts doubt on the value of such approaches given the cost of including biophysical detail. Here, we propose a new two-step method for TdP risk classification, referred to as Multi-Channel Blockage at Early After Depolarization (MCB@EAD). In the first step, we classified the compound that produced insufficient hERG block as non-torsadogenic. In the second step, the role of non-hERG channels to modulate TdP risk are considered by constructing classifiers based on direct or derived features at critical hERG block concentrations that generates EADs in the computational cardiac cell models. MCB@EAD provides comparable or superior TdP risk classification of the drugs from the direct features in tests against published methods. TdP risk for the drugs highly correlated to the propensity to generate EADs in the model. However, the derived features of the biophysical models did not improve the predictive capability for TdP risk assessment.

IC 50,I Kr IC 50,I CaV IC 50,I N a EF T P C T arget           Table S8. Merged Dataset. The table reports the block of seven ion channel currents (I CaV , I Kr , I K1 , I Ks , I to , I N a and I N aL ) at C Drug = EF T P C, the maximum EFTPC value of each of the drugs ("EFTPC" column) , the refrence/source ("Source" column) used to extract the IC 50 , Hill coefficients and EFTPC values for the particular drug, the hERG ratio ("Ratio" column), the drug-trapping parameter values for the available drugs ("vtrap"column) and the different risk definition assigned to them ('target', 'CM', 't1', 't2') columns. target-TdP+:-R1, R2 , R3, CH1, CM1 category or label warning , t1-TdP+: Drugs in CM1 and CM2 category, t2-TdP+: -Drugs in CM1 category, t3-TdP+: -Drugs in CM1 and CM3 category.     Table S9. Merged Dataset. The table reports the block of seven ion channel currents (I CaV , I Kr , I K1 , I Ks , I to , I N a and I N aL ) at C Drug = IC 60,hERG , the maximum EFTPC value of each of the drugs ("EFTPC" column) , the refrence/source ("Source" column) used to extract the IC 50 , Hill coefficient and EFTPC values for the particular drug, the hERG ration ("Ratio" column), the drug-trapping parameter values for the available drugs ("vtrap"column) and the different risk definition assigned to them ('target', 'CM', 't1', 't2') columns. target-TdP+:-R1, R2 , R3, CH1, CM1 category or label warning , t1-TdP+: Drugs in CM1 and CM2 category, t2-TdP+: Drugs in CM1 category, t3-TdP+: Drugs in CM1 and CM3 category. drugs are common to the Dataset 5 and 7. We extract in-vitro assay data for these drugs from Datasets 1, 2, 69 3, 5, 6 and 7. We used the mean value of the block for a particular channel if the information for the drug 70 was present in more than one dataset. No information was available for two (Azimilide and Loratidine) of 71 the 28 drugs in Datasets 1, 2, 3, 5, 6 and 7 and hence a final dataset of 26 drugs was obtained. If the value 72 is not available it was available for particular ion channels in any of the datasets described here, it was set 73 to zero. Drug-trapping parameter for 12 of these drugs reported in Dataset 7 was also also added to one of 74 the columns of the dataset.
EF T P C IC50,I    Table S12. Common drugs: 46 drugs were common across more than one datasets. Here, we list the % block of I CaV , I N a,peak for these drugs at drug concentrations equal to IC 60,hERG to highlight the differences across the datasets. Moreover, we list the risk category each drug was assigned to in the original datasets ("Risk (O)" column) as well the risk category ("Risk") assigned here to analyze the merged dataset. "Prediction" column highlights whether a drug is classified as TdP+ (1) or TdP (0) for the merged dataset using the direct features. We also highlight whether a drug results in EAD or not at C drug = IC 60,hERG in the "EAD" column. Drugs which ended up with different predictions are highlighted in gray. and Table 4 -22.925721

Classification accuracy based on the derived Features taking into account the
Dataset2 5.571355 -0.097846 -

Figure 3
and Table 4 1.58522

Figure 3
and Table 4 1.25502

Figure 3
and

81
TdP classifiers from the direct features built using SVM and Neural networks yielded identical 82 classification accuracy (85%) for the merged dataset as compared to logistic regression models.

83
Classification accuracies obtained using SVM and Neural networks on derived features are shown in 84 Figure S2 and resulted in maximum classification accuracy of 86%, as for the Logistic regression classifier.

85
For the limited data with insignificant nonlinearities the choice of the classifier had not significant impact 86 on the observed accuracies.

87
A B (A,B) TdP risk classification accuracy using the derived Features Figure S2. Heat maps of leave-one-out cross validation scores for logistic regression classifiers from 13 features of the APs and Ca 2+ transients in OHR model simulations using 1) Support Vector machine 2) Feed forward neural network with 13 hidden units. Drug-induced multi-channel block evaluated in the mid, endo and epi cell type at 500, 1000 and 2000 ms pacing rates. TdP risk classification at drug concentrations equal to hERG IC 50 .