A framework for drug safety assessment based on target affinity, drug exposure and plasma protein binding – dissecting FDA drug-induced cardiotoxicity data from a translational pharmacology perspective

Daniel Röshammar ¹, Simona Catozzi ¹, Fianne Sips ², Niccolò Totis ¹, Mario Torchia ¹, Luca Emili ¹, Vincenzo Carbone ¹

1 InSilicoTrials Technologies S.p.A. (Trieste, Italy), 2 InSilicoTrials Technologies B.V. (Amsterdam, The Netherlands)

Introduction: Cardiac safety assessment is an integrated part of drug discovery and development. Drug candidates with risk of adversely affecting cardiac and hemodynamic function should be stopped early on in the process, unless the expected benefit-risk ratio still is considered favorable for the patient. The drug-induced cardiotoxicity rank (DICTrank) data is based on 1318 drugs associated with cardiac safety concerns according to FDA labelling. The drugs are classified in four categories depending on the concern (most, less, ambiguous, no) and according to severity level (mild, moderate, severe) [1]. There are increasingly many attempts to explore factors causing the toxicity, using machine-learning [2-3]. However, it appears that potential predictors of event may be further informed by better considering the underlying pharmacology.

Objectives: To develop a conceptual framework for predicting cardiovascular toxicity based on in vitro drug affinities to pharmacological on- and off-targets of risk, expected therapeutic drug exposure and human plasma protein binding.
Methods: The DICTrank data was downloaded from the FDA [4]. The data was expanded with drug target affinity for the plausible receptor, ion channel or enzyme causing the event, clinical drug exposures and fraction unbound drug in plasma, based on publicly available information. An AI-assisted literature search was performed using Biomni (Phylo, https://www.phylo.com) based on sources such as ChEMBL, PubChem, DrugBank etc. Retrieved data was used for visualizing the most frequent categories of drug classes and targets involved in the various reported safety events. The underlying concentration-response relationships for target engagement were plotted for each drug (with Emax-functions) and stratified by the different identified pharmacological classes drugs, highlighting the free maximum plasma drug concentration.

Results: For DICTrank drugs of any safety concern used in our analysis (n=941), most could mechanistically be classified as binding to 25 targets involved in regulation of cardiac and hemodynamic function (adreno-, muscarinic-, dopamine-, serotonin-, GABA-, AT-1, opioid-, VEGFR-, EGFR- and sex hormone receptors, hERG-, Ca-, and Na-channels, cox-1 and cox-2, ACE, acetylcholinesterase, topoisomerase, monoamine transporters and PDE). For drugs classified as no concern (n=343), most were not categorized in any of these risk classes. An exception was the antibacterial drug trimethoprim, classified to be of no concern in DICTrank, but categorized as an hERG-channel inhibitor in our analysis. However, the target engagement was only predicted at 30%. In contrast, antibiotics not interacting with hERG-channels or other risk targets, were in general associated with no or less DICTrank concern (of less severity). Approximately 20% of all most safety concerns appear to be related to hERG and QT-prolongation effects.
Concentration-response relationships revealed differences in potency and free drug exposure. This explains some of the variability in severity of the event observed across drugs hitting the same target. The results can be used for benchmarking when screening new compounds, enabling early dropping of candidates with high cardiovascular risk. Through the use of vitro data, animal study designs and dose selection may be further optimized for more detailed safety assessment of the lead compounds carried forward.

Conclusion: The DICTrank database provides an opportunity for exploring and understanding a variety of reported cardiotoxic mechanisms of existing drugs. This can facilitate the safety assessment of new drugs hitting similar targets. Machine-learning based methods have great potential to discriminate between molecular properties potentially related to the adverse drug reactions. However, the predictors need to be explainable. They should be based on mechanistic understanding and established principles of pharmacology for useful translation to the clinic. For any drug, the target affinity to receptors, ion channels and enzymes of risk should be considered in combination with therapeutic exposure and human plasma protein binding.

References:
[1] Qu Y, Li T, Liu Z, Li D, Tong W. DICTrank: The largest reference list of 1318 human drugs ranked by risk of drug. Drug Discovery Today 28:11, 2023
[2] Mukherjee S, Swanson K, Walther P, Shivnaraine RV, Leitz J, Pang PD, Zou J, Wu JC. ADMET-AI enables interpretable predictions of drug-induced cardiotoxicity. Circulation 2025; 151;285-287.
[3] Seal S, Spjuth O, Hosseini-Gerami L, Garcia-Ortegon M, Singh S, Bender A, Carpenter AE. Insights into drug cardiotoxicity from biological and chemical data: the first public classifiers for FDA drug-induced cardiotoxicity rank. J. Chem. Inf. Model 2024; 64: 1172-1186
[4] https://www.fda.gov/science-research/bioinformatics-tools/drug-induced-cardiotoxicity-rank-dictrank-dataset)

Reference: PAGE 34 (2026) Abstr 12164 [www.page-meeting.org/?abstract=12164]

Poster: Drug/Disease Modelling - Safety