Application of Convolutional Neural Networks Using Action Potential Shape for In-Silico Proarrhythmic Risk Assessment

Abstract

This study proposes a convolutional neural network (CNN) model using action potential (AP) shapes as input for proarrhythmic risk assessment, considering the hypothesis that machine-learning features automatically extracted from AP shapes contain more meaningful information than do manually extracted indicators. We used 28 drugs listed in the comprehensive in vitro proarrhythmia assay (CiPA), consisting of eight high-risk, eleven intermediate-risk, and nine low-risk torsadogenic drugs. We performed drug simulations to generate AP shapes using experimental drug data, obtaining 2000 AP shapes per drug. The proposed CNN model was trained to classify the TdP risk into three levels, high-, intermediate-, and low-risk, based on in silico AP shapes generated using 12 drugs. We then evaluated the performance of the proposed model for 16 drugs. The classification accuracy of the proposed CNN model was excellent for high- and low-risk drugs, with AUCs of 0.914 and 0.951, respectively. The model performance for intermediate-risk drugs was good, at 0.814. Our proposed model can accurately assess the TdP risks of drugs from in silico AP shapes, reflecting the pharmacokinetics of ionic currents. We need to secure more drugs for future studies to improve the TdP-risk-assessment robustness.

Keywords: action potential shape; convolutional neural network; drug screening.

Figure 1

Schematic of in silico cell model for drug simulation; the ventricular cell model used in this study is the ORd model optimized to assess drug effects as described by Dutta et al. I_Na, fast Na⁺ current; I_NaL, L-type Na⁺ current; I_to, transient outward K⁺ current; I_Kr, rapid-delayed rectifier K⁺ current; I_Ks, slow-delayed rectifier K⁺ current; I_K1, inward rectifier K⁺ current; I_{NaCa, i}, 80% of Na⁺-Ca²⁺ exchange current; I_{NaCa, ss}, 20% of Na⁺-Ca²⁺ exchange current; I_NaK, Na⁺-K⁺ exchange current; I_CaK, Ca²⁺-K⁺ exchange current; I_CaNa, Ca²⁺-Na⁺ exchange current; I_CaL, L-type Ca²⁺ current; I_Nab, background Na⁺ current; I_Cab, background Ca²⁺ current; I_Kb, background K⁺ current; I_pCa, Ca²⁺ pump current; J_up, Ca²⁺ upstroke flux from myocyte into network sarcoplasmic reticulum (NSR); J_rel, Ca²⁺ flux through ryanodine receptor inside junctional sarcoplasmic reticulum (JSR); J_{diff, Na}, Na⁺ diffusion flux between subspace and myoplasm; J_{diff, Ca}, Ca²⁺ diffusion flux between subspace and myoplasm; J_{diff, K}, K⁺ diffusion flux between subspace and myoplasm; PLB, phospholamban; CSQN, calsequestrin; CaMK, Ca²⁺-calmodulin-dependent protein kinase II; BSR, anionic SR binding sites for Ca²⁺; BSL, anionic sarcolemmal binding sites for Ca²⁺.

Figure 2

Model structure of proposed CNN model using AP shapes as input: 1D CNN, one-dimensional convolutional neural network layer; BN, batch-normalization layer; AP, action potential. Each AP shape has 1000 data points with a 2 ms time resolution (cycle length of 2000 ms). After passing through three CNN groups, machine-learning features of 214 were extracted and fed into the artificial neural network layers with five nodes.

Figure 3

AP traces for 12 training drugs according to the Cmax variation: (a) quinidine, (b) sotalol, (c) dofetilide, (d) bepridil, (e) cisapride, (f) terfenadine, (g) chlorpromazine, (h) ondansetron, (i) verapamil, (j) ranolazine, (k) diltiazem, and (l) mexiletine. The solid black lines denote the median value among 500 AP traces in each Cmax condition.

Figure 4

AP traces for 16 test drugs according to Cmax variation: (a) disopyramide, (b) ibutilide, (c) vandetanib, (d) azimilide, (e) clarithromycin, (f) clozapine, (g) domperidone, (h) droperidol, (i) pimozide, (j) risperidone, (k) astemizole, (l) metoprolol, (m) nifedipine, (n) nitrendipine, (o) tamoxifen, and (p) loratadine. The solid black lines denote the median value among 500 AP traces in each Cmax condition.