Arrhythmia risk predictions from molecular simulations of cardiac ion channel-drug interactions

Figures

• 

  Figure 1. Overview of cardiotoxicity prediction pipeline.

• 

  Figure 2. Comparison of docking method scores versus experimental IC50 values. 53 drugs from Kramer et al. (19) compared with the IC50 values from the same study determined from patch-clamp electrophysiology. Both Spearman and Pearson correlation coefficients reported for both the hERG channel (a) and CaV1.2 (b). (c–f) Logistic regression models to predict TdP risk based on hERG and CaV1.2 predicted affinities. Docking scores for each method for 53 drugs from Kramer et al. (19) were used to train the models. Performance of the models was then tested on an independent set of 29 drugs. TdP risk was taken from CredibleMeds where red indicates “Known Risk” and blue indicates “Therapeutic Options” (drugs with no known TdP risk). Optimal threshold is shown as yellow vertical line.

• 

  Figure 3. Logistic regression models trained on experimental (a) and SILCS LGFE (b–f) data from several ion channel structures. (a) Model trained with experimental IC50 data from Kramer et al. (19) for hERG (hERGIC50), NaV1.5 (NavIC50), and CaV1.2 (CavIC50) channels. (b) Model trained with hERG cryo-EM structures with PDB: 5VA2 (hERG5VA2) and 9CHQ (hERG9CHQ), (c) hERG5VA2 and CaV1.2 model based on PDB: 6AGF (Cav6AGF), (d) hERG5VA2, CaV1.26AGF, and NaV1.5 PDB: 6LQA structure (Nav6LQA), (e) hERG5VA2, CaV1.26AGF, and hERG S641A mutant model (hERGS641A), and (f) hERG5VA2, hERG9CHQ, and CaV1.26AGF. Optimal cutoffs for the risk prediction are shown as yellow vertical lines. Figure panel titles indicate data used to train the logistic regression model of the form shown by Eq. (1) in the materials and methods section.

• 

  Figure 4. Feature analysis of random forest classifier model. (a) SHAP values for eight top features of the model Individual feature values in the beeswarm plots are shown by colored dots, ranging from blue (low values) to red (high values), illustrating how variation in each top feature, i.e. physicochemical property or docking score, influences predicted TdP risk. SILCS total binding score is indicated as the ligand grid free energy (LGFE) as compared with binding to a specific fragment methylammonium nitrogen (MAMN), methanol oxygen (MEOO), or benzene carbon (BENC). TPSA indicates topological polar surface area. hERG5 is an abbreviation for hERG_5VA2, hERG9 for hERG_9CHQ, and hERGS for hERG_S641A models. CaV1.2 is the Cav_6AGF model and NaV1.5 is the Nav_6AGF model. (b) Force plots for individual drugs indicating how particular model features influenced the predicted TdP probability. The force plots present a balance of features that are predictive of TdP+ as red bars pointing to the right, whereas features the models associate with TdP– drugs are blue bars pointing to the left. The interface of these red and blue features is the predicted TdP probability for a given drug. The…

• 

  Figure 5. Analysis of important features from random forest classifier model. (a) Interactions of the drug’s basic nitrogen with the MAMN FragMap (cyan mesh) in the hERG channel pore (green ribbons) helps discriminate terfenadine from raltegravir, despite the lower overall hERG LGFE of the former. (b) Inclusion of hydroxyl groups that interact with the MEOO FragMaps (light gray mesh) in NaV1.5 (dark gray ribbons) is associated with safer drugs such as raltegravir. (c) Force plot for quinapril and vitamin D, highlighting contribution of model features to the specific compounds. Predicted TdP probability is shown as the interface between red and blue features. Please see Fig. 4b caption for details. (d) Performance of the model after removing selected features (e.g., –LGFE indicates removing LGFE scores from the model). (e) SHAP analysis highlighting importance of physicochemical properties to the random forest model. Please see Fig. 4a caption for details.