Prediction of all forms of drug-induced cardiotoxicity by combined transcriptome analysis and machine learning

[Speaker] Alfonso Bueno-Orovio:1
[Co-author] Polina Mamoshina:1, Blanca Rodriguez:1
1:Computer Science, University of Oxford, UK

Background: Cardiotoxicity is one of the major drug safety concerns leading to discontinuation of the drug development process and market withdrawal. Current preclinical methods for drug cardiotoxicity largely rely on hERG ion channel inhibition or prolongation of the QT interval. Despite their recognized role in safety pharmacology, they mostly focus on drug-induced electrophysiological changes, and frequently fail to predict other types of cardiotoxicity, such as contractile and structural heart toxicity related to a long-term drug exposure.
Methods: To develop a method that could predict those cardiotoxic-related changes and potential safety of drugs, we analysed the gene expression profiles of 30 safe and 23 cardiotoxic drugs. For this purpose, we used the open-data FAERS and SIDER databases of side effects, complemented with literature data mining on compounds with cardiac side effect and matching safe compounds. Drugs were then linked to their gene expression profiles in the open-access DToxS and Connectivity map databases. Differential gene expression and pathway analysis were performed to compare signatures of cardiotoxic and safe drugs, and to preprocess the data to train multiple architectures of machine learning classifiers.
Results: The cardiotoxic drug signatures identified from transcriptome analysis were diverse, reflecting the heterogeneity of cardiotoxic mechanisms. Frequent cardiotoxic signatures were however identified between the analysed antineoplastic, cardiovascular, anaesthetic, anti-edema, anti-diarrheal, antibiotic, anti-acid, anti-viral and anti-depressant agents. These included common signalling pathways related to pro-apoptotic, anti-apoptotic, fibrotic, cardiomyocyte proliferation and function, thrombotic and ion channel function. Our analysis further allows the identification of additional mechanisms of drug cardiotoxicity, such as inhibition of ion channel trafficking by the tyrosine kinase inhibitor imatinib. Machine learning classification based on transcriptome analysis achieved global sensitivity of 90% and specificity of 82% in discrimination between cardiotoxic and non-cardiotoxic drugs for all types of drug-induced cardiotoxicity.
Conclusion: The proposed transcriptome and machine learning analysis allows identification of drug-induced cardiotoxic signatures at the level of all the most important signalling pathways of cardiac function, achieving state-of-the-art classification for all forms of drug-induced cardiotoxicity. These techniques could therefore complement current preclinical and clinical practice for drug safety testing beyond the identification of drug-induced pro-arrhythmic cardiotoxicity.
Advanced Search