Abstract
Drug-induced cardiotoxicity (DICT) is a major concern in drug development, accounting for 10-14% of postmarket withdrawals. In this study, we explored the capabilities of various chemical and biological data to predict cardiotoxicity, using the recently released Drug-Induced Cardiotoxicity Rank (DICTrank) dataset from the United States FDA. We analyzed a diverse set of data sources, including physicochemical properties, annotated mechanisms of action (MOA), Cell Painting, Gene Expression, and more, to identify indications of cardiotoxicity. We found that such data, including protein targets, especially those related to ion channels (such as hERG), physicochemical properties (such as electrotopological state) as well as peak concentration in plasma offer strong predictive ability as well as valuable insights into DICT. We also found compounds annotated with particular mechanisms of action, such as cyclooxygenase inhibition, could distinguish between most-concern and no-concern DICT compounds. Cell Painting features related to ER stress discern the most-concern cardiotoxic compounds from non-toxic compounds. While models based on physicochemical properties currently provide substantial predictive accuracy (AUCPR = 0.93), this study also underscores the potential benefits of incorporating more comprehensive biological data in future DICT predictive models. With the availability of -omics data in the future, using biological data promises enhanced predictability and delivers deeper mechanistic insights, paving the way for safer therapeutic drug development. All models and data used in this study are publicly released at https://broad.io/DICTrank_Predictor.
Introduction
Drug-induced cardiotoxicity (DICT) remains a formidable challenge in pharmaceutical development, often acting as a "silent killer" that bypasses early safety screens only to cause patient harm and costly postmarket withdrawals. Historically, cardiotoxicity screening has leaned heavily on proxy assays like the hERG potassium channel test. While valuable, these assays fail to capture the multi-faceted nature of cardiac damage, which can involve diverse biological pathways and systemic interactions.
The release of the FDA’s DICTrank dataset marks a significant milestone. Categorizing over 1,300 drugs based on clinical evidence and FDA labeling, DICTrank provides a robust "ground truth" for developing more sophisticated predictive models. This paper presents the first major effort to build public machine learning classifiers using this benchmark, evaluating the relative strengths of chemical and biological data in identifying cardiac risk.
Methodology and Framework
The researchers developed a comprehensive framework, the DICTrank Predictor, which integrates multiple data modalities to classify drugs into concern levels.
The study utilized Random Forest classifiers trained across 11 distinct feature spaces. These spaces were broadly categorized into:
- Chemical Descriptors: Including Morgan fingerprints for structural patterns and Mordred descriptors for 2D physicochemical properties (e.g., lipophilicity, molecular weight).
- Biological and Pharmacokinetic Data: Annotated mechanisms of action (MOA), predicted protein targets via CellScape, and peak plasma concentration (Cmax).
- High-Dimensional Omics: Transcriptomic data from L1000 gene expression and morphological profiles from Cell Painting assays.
To address the inherent sparsity of biological data, the team employed a v-NN (variant Nearest Neighbor) imputation strategy, allowing them to estimate biological profiles for compounds based on their structural similarity to well-characterized drugs.
Data Integration and Insights
The study drew from a vast array of public resources, including SIDER for side-effect enrichment, the Drug Repurposing Hub for MOA annotations, and the LINCS project for omics data.
A key finding was the surprising efficacy of "simple" chemical descriptors. Models using Mordred descriptors achieved the highest predictive accuracy, with an AUCPR of 0.93 and an AUC-ROC of 0.84. This suggests that fundamental physical properties are often sufficient to flag high-risk compounds during early discovery.
However, the value of biological data lies in its ability to provide mechanistic context—the "why" behind the toxicity. For instance:
- Cell Painting features revealed that drugs in the "Most-Concern" category often induce significant Endoplasmic Reticulum (ER) stress.
- Target Analysis identified biomarkers beyond hERG, such as α-l-fucosidase I and P-selectin, which are linked to cardiac hypertrophy and inflammatory responses.
- MOA Patterns confirmed that cyclooxygenase (COX) inhibitors and certain tyrosine kinase inhibitors are disproportionately associated with high cardiotoxicity, likely due to their impact on vascular resistance and cellular homeostasis.
Results and Model Performance
The ensemble approach, which combined the top six feature spaces, performed on par with the best chemical models (AUCPR = 0.92). This ensemble provides a more holistic view, balancing structural flags with biological signals.
The researchers also applied their best-performing model to 82 drugs labeled as "Ambiguous" in the original FDA dataset. The model flagged 43 of these as potentially cardiotoxic, providing a starting point for further clinical review. Furthermore, the inclusion of Cmax data highlighted the importance of dosage; many compounds with toxic potential are clinically safe because their therapeutic concentrations remain well below the threshold for cardiac damage.
Conclusion
While chemical descriptors currently lead in predictive performance due to their broad availability, this study demonstrates that biological "omics" data offer indispensable insights into the mechanisms of drug-induced heart damage. As datasets like JUMP-CP expand the availability of high-quality morphological and transcriptomic profiles, the predictive power of biological models is expected to grow. For now, the authors recommend a dual-track approach: using chemical models for rapid, large-scale screening and biological data for detailed mechanistic validation of lead candidates.
All models, code, and processed datasets have been made publicly available to foster further innovation in cardiovascular safety.
Summary updated