bioRxiv preprint doi: https://doi.org/10.1101/2020.10.26.352740; this version posted October 27, 2020. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license. Sex-specific classification of drug-induced Torsade de Pointes susceptibility using cardiac simula- tions and machine learning Alex Fogli Iseppe1, Haibo Ni1, Sicheng Zhu1, Xianwei Zhang1, Raffaele Coppini3, Pei-Chi Yang2, Uma Srivatsa4, Colleen E. Clancy1.2, Andrew G. Edwards1, Stefano Morotti1, Eleonora Grandi1* 1 Department of Pharmacology, University of California, Davis, CA, USA. 2 Department of Physiology and Membrane Biology, University of California, Davis, CA, USA. 3 Department of Neuroscience, Psychology, Drug Sciences and Child Health (NeuroFarBa), University of Florence, Italy. 4 Department of Internal Medicine, University of California, Davis, CA, USA. *Correspondence [email protected] Running Title: Sex-specific classifiers of torsadogenic risk Key words: Cardiac safety pharmacology, computational modeling, machine learning classification Abstract Torsade de Pointes (TdP), a rare but lethal ventricular arrhythmia, is a potential cardiac side effect of drugs. To assess TdP risk, safety regulatory guidelines require to quantify the effects of new therapeutic com- pounds on hERG channel block in vitro and QT interval prolongation in vivo. Unfortunately, these have proven to be poor predictors of torsadogenic risk, and are likely to have prevented safe compounds from reaching the clinical phase. While this has stimulated numerous efforts to define new paradigms for cardiac safety, none of the recently developed strategies accounts for patient conditions. In particular, despite being a well-established independent risk factor for TdP, female sex is vastly underrepresented in both basic re- search and clinical studies, and thus current TdP metrics are likely biased toward the male sex. Here, we apply statistical learning to synthetic data, generated by simulating drug effects on cardiac myocyte models capturing male and female electrophysiology, to develop new sex-specific classification frameworks for TdP risk. We show that 1) TdP classifiers require different features in females vs. males; 2) male-based classifiers perform more poorly when applied to female data; 3) female-based classifier performances are largely unaffected by acute effects of hormones (i.e., during various phases of the menstrual cycle). Notably, when predicting TdP risk of intermediate drugs on female simulated data, male-biased predictive models consistently underestimate TdP risk in women. Therefore, we conclude that pipelines for preclinical cardi- otoxicity risk assessment should consider sex as a key variable to avoid potentially life-threatening conse- quences for the female population. 1. Introduction have caused withdrawal from the market of sev- During drug development, promising therapeutic eral drugs, including antihistamines, antidepres- compounds are tested to evaluate their potential sants, chemotherapeutics, pain medications, that risk of inducing Torsade de Pointes (TdP), a spe- had been associated with TdP proclivity in pa- cific form of polymorphic ventricular tachycar- tients (e.g., Cisapride and Astemizole)2,3. The dia that can precipitate ventricular fibrillation most simple mechanistic explanation of torsado- and cause sudden cardiac death1. While TdP is a genicity involves a reduction of the rapid delayed very rare adverse event, amounting to less than rectifier potassium current (IKr), carried by the one case out of 100,000 exposures for some non- human Ether-à-go-go-Related Gene (hERG) antiarrhythmic drugs,2 cardiac safety concerns channel, which importantly contributes to 1 bioRxiv preprint doi: https://doi.org/10.1101/2020.10.26.352740; this version posted October 27, 2020. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license. cardiac action potential (AP) repolarization4,5. sex: it is well established that women are more Pharmacological block of the hERG channel, susceptible to Torsade than men when treated which is a very promiscuous target interacting with QT-prolonging drugs,20–22 suggesting that with cardiac and non-cardiac drugs, produces AP TdP risk classifiers could benefit from inclusion duration (APD) and QT interval prolongation, of this variable. However, female sex is highly and leads to an increased susceptibility to pro-ar- underrepresented in both basic research23 and rhythmic events. Based on this evidence, current clinical studies24 involved in the drug develop- safety regulatory guidelines require the measure- ment process, with important consequences on ment of hERG channel block in vitro and QT in- the identification of accurate TdP predictors. In terval prolongation in vivo to estimate TdP vitro studies tend to use mostly male animals23, risk6,7. Since their adoption, these guidelines raising concerns on the generalizability of find- have successfully avoided that cardiotoxic drugs ings to the whole population. This sex bias prop- could endanger the welfare of people. However, agates onto the mathematical models of cardiac it has also become apparent that these biomarkers cells25,26, which are parameterized based on are poor predictors of torsadogenic risk, and have male-dominated datasets. The issue of underesti- in all probability prevented safe treatments from mating potential health risks for women is then reaching the market8,9. aggravated by the fact that female sex is also un- In response to this problem, recent efforts derrepresented in clinical cohorts,24 making have led to several proposed new paradigms for training of classifiers harder due to the lack of re- the prediction of TdP. One notable example is the liable ground truth data. Comprehensive In Vitro Proarrhythmia Assay Yang and Clancy have recapitulated in silico (CiPA) initiative, an international multi-group male and female ventricular human cardiac elec- initiative by regulatory, industry, and academic trophysiology by incorporating experimentally partners including the US Food and Drug Admin- determined sex- and hormone-specific differ- istration10. This paradigm relies upon the idea of ences in gene and protein expression into virtual combining in vitro studies to measure the drugs male and female myocytes27,28. In this paper we effects on each of the different types of ion chan- combined simulations of these mathematical nels and in silico models of cardiac myocyte elec- models with machine learning to generate sex- trophysiology to understand how these effects specific TdP risk classifiers. We simulated the ef- combine to influence cardiac function, thus cre- fects of 59 training drugs under different condi- ating a novel tool for TdP risk assessment of new tions using in silico models of human ventricular drugs11–13. Mathematical models of cardiac elec- myocytes with sex-specific parameterizations. trophysiology, in fact, make it possible to simu- We fed the resulting high-dimensional datasets late with precision extreme conditions, e.g., high of simulated biomarkers to machine learning al- drug concentrations, and to obtain insights pre- gorithms to generate male and female classifiers cluded to animal experiments. Thus, computa- of torsadogenic risk. Finally, we evaluated the ef- tional approaches have become essential compo- fects of using sex-specific models for risk predic- nents of numerous strategies to predict torsado- tion on a separated set of 36 drugs, which are genic risk14–19. In addition, simulated measure- deemed at intermediate risk of TdP. Our results ments extracted from the biophysical model sim- show that TdP classifiers trained on sex-specific ulations can also be fed to machine learning (ML) datasets identify distinct and not interchangeable pipelines, as demonstrated by the Sobie group15, sets of optimal features, suggesting potential dif- with the potential to bring out mechanistic in- ferent drivers of drug-induced arrhythmias, and sights buried in the data that could be otherwise that the use of sex-biased predictive models un- ignored. derestimates the torsadogenic risk of drugs with To our knowledge, however, no simulation- intermediate risk of TdP in females, which could based approach has considered any risk factor for potentially lead to life-threatening consequences Torsade in their predictive pipelines. An em- for women. blematic example is represented by the female 2 bioRxiv preprint doi: https://doi.org/10.1101/2020.10.26.352740; this version posted October 27, 2020. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license. Figure 1: Workflow for the creation and testing of the sex-specific classifiers for torsadogenic risk. 2. Materials and Methods clinical source of torsadogenic risk categoriza- 2.1 Models and simulations tion that takes in account the sex variable. In or- We used the male and female human epicardial der to assign a unique binary label to each drug, ventricular cardiomyocyte models developed by we took advantage of the TdP risk classification the Clancy lab28,