Performance of Machine Learning Algorithms for Qualitative and Quantitative Prediction Drug Blockade of hERG1 channel

Abstract

Drug-induced abnormal heart rhythm known as Torsades de Pointes (TdP) is a potential lethal ventricular tachycardia found in many patients. Even newly released anti-arrhythmic drugs, like ivabradine with HCN channel as a primary target, block the hERG potassium current in overlapping concentration interval. Promiscuous drug block to hERG channel may potentially lead to perturbation of the action potential duration (APD) and TdP, especially when with combined with polypharmacy and/or electrolyte disturbances. The example of novel anti-arrhythmic ivabradine illustrates clinically important and ongoing deficit in drug design and warrants for better screening methods. There is an urgent need to develop new approaches for rapid and accurate assessment of how drugs with complex interactions and multiple subcellular targets can predispose or protect from drug-induced TdP. One of the unexpected outcomes of compulsory hERG screening implemented in USA and European Union resulted in large datasets of IC₅₀ values for various molecules entering the market. The abundant data allows now to construct predictive machine-learning (ML) models. Novel ML algorithms and techniques promise better accuracy in determining IC₅₀ values of hERG blockade that is comparable or surpassing that of the earlier QSAR or molecular modeling technique. To test the performance of modern ML techniques, we have developed a computational platform integrating various workflows for quantitative structure activity relationship (QSAR) models using data from the ChEMBL database. To establish predictive powers of ML-based algorithms we computed IC₅₀ values for large dataset of molecules and compared it to automated patch clamp system for a large dataset of hERG blocking and non-blocking drugs, an industry gold standard in studies of cardiotoxicity. The optimal protocol with high sensitivity and predictive power is based on the novel eXtreme gradient boosting (XGBoost) algorithm. The ML-platform with XGBoost displays excellent performance with a coefficient of determination of up to R² ~0.8 for pIC₅₀ values in evaluation datasets, surpassing other metrics and approaches available in literature. Ultimately, the ML-based platform developed in our work is a scalable framework with automation potential to interact with other developing technologies in cardiotoxicity field, including high-throughput electrophysiology measurements delivering large datasets of profiled drugs, rapid synthesis and drug development via progress in synthetic biology.

Keywords: Drug Discovery; Drug-Induced Cardiotoxicity; Gradient-Boosting; Lead Optimization; Machine-Learning; Quantitative Structure Activity Relationship; hERG1 channel.

Figure 1

Machine Learning Platform Flowchart. For feature selection and model tuning only compounds in the training set were used. The compounds in the test sets were saved for the final evaluation of model performance. For model selection 10-fold cross-validation was used using the compounds in the training set. For the final model the complete training set was used.

Figure 2

A) Approximated density of pIC50 values in the Training and Test sets. B) Approximated density of the maximum similarities to compounds in the training set for all test sets. For the training set the similarity to the next most similar compound is shown. The curves are scaled so that the area under the curve is 1.

Figure 3

A) Q2 values from 10-fold cross-validation for the three feature sets FS1 (●), FS2 (▲), FS3 (■). B) Boxplots summarizing Q2 values. C–F) mean Q2 projected on the C) fraction of columns used to build each decision tree, D) the learning rate eta, E) the maximal depth of each tree and F) the size of the sample used for each tree. G) Training (■) and validation (●) RMSE over the size of the training set for different numbers of trees using the final parameters. Here fractions (between 0.2 and 1) of the training set were used to analyse the dependency of the predictive power on the size of the training set.

Figure 4

Model performance for all test sets combined. A) Correlation to experimental data. B) ROC curve using same class criteria. C) Error over the distance to the training set for each compound. Color codes illustrate the MST values (blue: MST > 0.5, red: MST < 0.5). Model performance for the combined test sets (D) and dependences on different similarity thresholds (E and F). G) Location of range violations. Number or range violations indicated by color and size of spheres. Cross-symbols indicate compounds with no range violations. H) and I) model performance for compounds within the recommended applicability domain.

Figure 5

The number of times features have been used in the model: A) the top 20 features and B) the following features. Molecular descriptors (blue) and pharmacophore features (black). Only scores of the top 150 features (of ~400) are shown.

Figure 6

Comparison of the final model with reference models. Only the range between -1 and 1 is shown. The value for the Lasso model for test set 4 (Test4) was - 1.22. The dotted line marks R2 = 0.6.