Figures and data
Conceptual framework.
(A) Drugs simultaneously block multiple species of ion channels to differing degrees. The principal ion currents implicated in druginduced Torsades des Pointes are ICaL, IKr, INaL and IKs. (B) Simplified circuit diagram of cardiomyocyte electrophysiology. Drug blockade is simulated by attenuating the ionic conductances (GCaL, GKr, GNaL, GKs). Those parameters are also varied randomly to mimic individual differences in electrophysiology. (C) Simulated action potentials of phenotypically diverse cardiomyocytes. Early after-depolarizations (red) are biomarkers for Torsades des Pointes. Conventional in silico assays simulate the effect of drugs on cardiomyocytes on a case by case basis. Our method inverts the procedure by simulating cardiomyocytes in the absence of drugs and then inferring how drugs would behave.
Benign versus ectopic cardiac phenotypes.
(A) Simulated action potentials for cardiomyocytes with randomly scaled conductance parameters GKr, GCaL, GNaL and GKs. Myocytes that exhibited early after-depolarizations were classified as ectopic (red). Those that did not were classified as benign (gray). (B) Swarm plots of the conductance scalars on a logarithmic scale. Color indicates the classification of the myocyte (benign versus ectopic). (C) Twodimensional slice of parameter space showing the relationship between ectopic and benign phenotypes in GCaL versus GKr. The dashed line is the statistical decision boundary. GNaL and GKs were fixed at unity (e0 = 1). (D) Twodimensional slice showing GNaL versus GKs. In this case GCaL is e0.46 and GKr is e−2.3.
Quantifying drug risk with the axis of arrhythmia.
(A) The axis of arrhythmia runs orthogonally to the decision boundary. As such, it describes the shortest pathway to ectopy for any cardiomyocyte. The action of a drug is represented by 
Susceptibility to a drug in the natural population.
(A) Natural variation in ion channel conductivity is represented by a symmetric Gaussian density function centered at point O. In this example, a tenfold dose of Ibutilide shifts the population by 1.44 units towards the ecoptic region. The proportion of ectopic myocytes in the drugged population is 41.5% (red). (B) The relationship between the drug risk score and ectopy in the natural population. The drug risk score corresponds to position on the axis of arrhythmia. The shaded region is the a priori probability of ectopy along that axis (reproduced from Figure 3B). The Gaussian profile (thin gray line) is the natural population density centered at zero. The proportion of myocytes that are ectopic (heavy black line) is 0.93% at baseline. That proportion rises as the drug shifts the population density towards the decision boundary. See Supplementary Video SV1 for an animated version.
Cases of Ajmaline and Linezolid.
(A) Drug response profiles of ICaL, IKr, INaL and IKs relative to the therapeutic dose. Open circles highlight 25x therapeutic dose. Ajmaline has δCaL = 0.654 and δKr = 0.0986 at 25x dose, whereas Linezolid has δCaL = 0.067 and δKr = 0.437. Data for INaL and IKs were not available, so they were assumed to be unaffected (δNaL = 1 and δKs = 1). (B) The blocking action of Ajmaline and Linezolid at 25x dose. Drug action is defined as α = ln(δ). (C) The corresponding risk scores for Ajmaline (+1.16) and Linezolid (−1.18) at that dose. (E) The drug-induced shifts of the natural population density along the axis of arrhythmia. The proportion of myocytes that are ectopic with 25x dose of Ajmaline is 26% (red), compared to only 0.0095% for Linezolid.
The effect of multi-channel block changes with dosage.
(A) The attenuation of GCaL, GKr, GNaL and GKs by Ajmaline over a range of doses. (B) The action of Ajmaline in parameter space. Gray traces are the pathways of all 109 drugs that we investigated. (C-D) Propafenone. (E-F) Linezolid.
Torsadogenic risk for 109 drugs at 25x dose.
Colors indicate the clinical risk labels from Credible Meds. The drugs are sorted by the score returned by our risk metric (lower axis). The proportion of the natural population that would be susceptible to the drug is shown on the upper axis. Drugs to the right of the scoring threshold (θ = 0.195) were classified as UNSAFE and those to the left of it were classified as SAFE. Misclassified drugs are marked with a triangle and highlighted in bold. In this case, 90.8% of the drugs were correctly classified.
Optimal dosage.
(A) Classification accuracy for drugs assessed at a range of dosages. (A) ROC curve for drugs at 5x dose. (B) ROC curve for drugs at 25x dose. AUROC is area under the ROC curve. TPR is true positive rate. FPR is false positive rate. The false negative rate is 1-TPR.
The baseline scaling factors applied to the ORD11 model.
Animation of Figure 4 illustrating the relationship between drug risk and population risk. A. Drug-induced shift of the the natural population along the axis of arrhythmia. The size of the shift corresponds to the risk score of the drug. B. The a priori probability of ectopy along the axis of arrhythmia (shaded) in conjunction with the natural population density (light gray). C. The probability of ectopy for the natural population (black line) against the a priori probability of ectopy (shaded).
The drug response data in CSV format. The data was reconstructed from the IC50 values published in Supplementary Table S2 of Llopis-Lorente et al (2020). The Hill coefficient was assumed to be h = 1 in all cases. Ion channels with missing values were assumed to be unaffected by the drug (δ = 1). The clinical risk labels were transcribed from Table 1 of Llopis-Lorente et al (2020)
