Selection of safe artemisinin derivatives using a machine learning-based cardiotoxicity platform and in vitro and in vivo validation

Abstract

The majority of drug candidates fails the approval phase due to unwanted toxicities and side effects. Establishment of an effective toxicity prediction platform is of utmost importance, to increase the efficiency of the drug discovery process. For this purpose, we developed a toxicity prediction platform with machine-learning strategies. Cardiotoxicity prediction was performed by establishing a model with five parameters (arrhythmia, cardiac failure, heart block, hypertension, myocardial infarction) and additional toxicity predictions such as hepatotoxicity, reproductive toxicity, mutagenicity, and tumorigenicity are performed by using Data Warrior and Pro-Tox-II software. As a case study, we selected artemisinin derivatives to evaluate the platform and to provide a list of safe artemisinin derivatives. Artemisinin from Artemisia annua was described first as an anti-malarial compound and later its anticancer properties were discovered. Here, random forest feature selection algorithm was used for the establishment of cardiotoxicity models. High AUC scores above 0.830 were achieved for all five cardiotoxicity indications. Using a chemical library of 374 artemisinin derivatives as a case study, 7 compounds (deoxydihydro-artemisinin, 3-hydroxy-deoxy-dihydroartemisinin, 3-desoxy-dihydroartemisinin, dihydroartemisinin-furano acetate-d3, deoxyartemisinin, artemisinin G, artemisinin B) passed the toxicity filtering process for hepatotoxicity, mutagenicity, tumorigenicity, and reproductive toxicity in addition to cardiotoxicity. Experimental validation with the cardiomyocyte cell line AC16 supported the findings from the in silico cardiotoxicity model predictions. Transcriptomic profiling of AC16 cells upon artemisinin B treatment revealed a similar gene expression profile as that of the control compound, dexrazoxane. In vivo experiments with a Zebrafish model further substantiated the in silico and in vitro data, as only slight cardiotoxicity in picomolar range was observed. In conclusion, our machine-learning approach combined with in vitro and in vivo experimentation represents a suitable method to predict cardiotoxicity of drug candidates.