Cardiac Safety Risk Classification

Section 01

Executive summary

Torsade de pointes is a rare but potentially fatal ventricular arrhythmia, and the risk that a drug causes it has historically been judged largely on how strongly it blocks a single cardiac potassium channel. That single-channel view both misses dangerous drugs and wrongly condemns safe ones. The Comprehensive In Vitro Proarrhythmia Assay (CiPA) initiative was established to replace it with a multi-channel assessment, and defined a reference set of drugs with agreed high, intermediate and low risk labels.

This tool takes measured ion channel inhibition data and free clinical exposure for a drug and returns a three-category proarrhythmic risk classification. Across the 27 CiPA reference drugs, it classified 26 correctly (96.3%), including all 8 high-risk and all 11 low-risk drugs, with sensitivity 1.000 and specificity 0.947 for high-risk detection.

The single misclassification, pimozide, was called high risk when its reference label is intermediate. The error is in the conservative direction: the tool over-warned rather than under-warned, and no high-risk drug was missed.

Section 02

What goes in and what comes out

Input

Per drug: measured half-maximal inhibitory concentrations for the cardiac ion currents (I_Kr, I_CaL, I_NaL, I_Na, I_Ks, I_K1, I_to) and the free therapeutic plasma concentration. Channel values in this validation come from the Crumb 2016 dataset. This tool classifies risk from measured in vitro data; it does not predict channel potency from chemical structure.

Output

A single risk category of high, intermediate or low, together with the intermediate quantities the decision rests on: fractional block of each current at one and four times clinical exposure, combined inward current block, I_Kr block, and net charge shift at four times exposure.

Reproducibility

The classification is deterministic. The same channel data and exposure always produce the same category, with no training data, random seed or ensemble averaging anywhere in the calculation.

The classification logic and thresholds are proprietary and are not disclosed here. All inputs are public, and intermediate quantities can be made available for independent audit under confidentiality.

Section 03

Results

Rows are the agreed CiPA label, columns are the tool’s call. The one off-diagonal entry is pimozide.

Why multi-channel matters

The high-risk drugs do not look alike on any single channel. Bepridil reaches 126% combined inward current block with a net charge shift of −279, while dofetilide shows no measurable inward block and a shift of +55. Both are high risk, reached by different routes. A classifier watching only I_Kr would treat terfenadine, at 4% block at clinical exposure, as unremarkable.

Section 04

Published comparison

The CiPA reference drugs are a well-used benchmark, so this result can be placed against the published record directly.

The CiPA in silico framework’s own metric is qNet, the net charge carried by the cardiac currents across one beat, averaged across one to four times clinical exposure and converted to a risk category by ordinal logistic regression. In an independent evaluation, qNet classified 24 of 28 CiPA drugs correctly and outperformed alternative metrics including action potential duration and the EAD threshold. Its developers reported high accuracy for torsade risk prediction and judged it potentially fit for regulatory use within the CiPA paradigm.

Later work has sought further gains from machine learning on the same simulation outputs. One neural network approach built on an alternative ventricular cell model reported category accuracies of 0.979 for high risk, 0.791 for intermediate risk and 0.937 for low risk. The intermediate category is consistently the hardest in published work, which is where the 7 of 8 result here sits.

Performance on this reference set is also known to fall when methods are carried to wider drug sets. In one larger evaluation, a classifier combining several proarrhythmic indices reached 94.5% on 109 compounds, and the authors noted that a metric scoring perfectly on the 12 CiPA training drugs performed less well once the set was broadened. That is the right caution to attach to any result on 27 drugs, including this one.

References: Li Z et al. Assessment of an in silico mechanistic model for proarrhythmia risk prediction under the CiPA initiative. Clin Pharmacol Ther 2019;105(2):466–475. Dutta S et al. Optimization of an in silico cardiac cell model for proarrhythmia risk assessment. Front Physiol 2017;8:616. Passini E et al. Human in silico drug trials demonstrate higher accuracy than animal models in predicting clinical pro-arrhythmic cardiotoxicity. Front Pharmacol 2019;10:1054. Llopis-Lorente J et al. In silico classifiers for the assessment of drug proarrhythmicity. J Chem Inf Model 2020;60(10):5172–5187. Crumb WJ et al. An evaluation of 30 clinical drugs against the comprehensive in vitro proarrhythmia assay (CiPA) proposed ion channel panel. J Pharmacol Toxicol Methods 2016;81:251–262.

Section 05

Limitations

Reference set size

27 drugs across three categories means 8 to 11 drugs per category. Confidence intervals around every figure here are wide, and published experience shows accuracy on this set tends to overstate accuracy on broader drug sets. An independent held-out set would strengthen this result and is welcomed.

Dependence on input data quality

Classification rests entirely on the measured channel values and the free exposure estimate supplied. Channel potencies vary between laboratories and assay conditions, and free fraction estimates vary between sources. The tool inherits that variability; it does not reduce it.

Intermediate category

The single error falls in the intermediate category, which is the hardest of the three in published work as well. Intermediate drugs sit near the thresholds by definition, and calls in that band should be treated as less certain than high or low calls.

Channels not used in the decision

Measured values for I_K1, I_Ks and I_to are reported per drug but do not currently enter the classification. Coverage for these currents across the reference set is also incomplete.

Scope of claim

This is research validation evidence. The tool is not cleared or qualified for clinical or regulatory use, and no result from it should direct the care of an individual patient.