TL;DR
This paper presents a novel AI-driven approach to identify potential hERG channel blockers within the DrugBank database, a critical step in preventing drug-induced cardiotoxicity. By integrating advanced molecular embedding techniques with machine learning models, the researchers successfully screened 8641 compounds, flagging 227 as potential blockers, including 92 FDA-approved drugs. This work offers a crucial tool for early-stage drug safety assessment and highlights potential risks associated with existing medications.
What Problem Does This Paper Address?
The hERG potassium channel plays a vital role in cardiac function, and its blockade by certain drugs can lead to serious, even fatal, cardiac arrhythmias like Torsade de Pointes (TdP). Consequently, many drugs have been withdrawn from the market due to hERG-related cardiotoxicity. Early identification of potential hERG blockers is therefore paramount in drug discovery and development to ensure patient safety. While experimental methods exist, they are often time-consuming and expensive, necessitating the development of rapid and cost-effective in silico screening tools. This paper addresses the need for accurate and robust computational models to screen large drug databases like DrugBank, which contains a vast collection of approved, experimental, and investigational drugs.
Methods
The core of this work lies in a sophisticated machine learning pipeline that combines several advanced techniques:
- Molecular Embeddings: The researchers employed a multi-faceted approach to represent molecules numerically:
- Sequence-based Embeddings: Two Natural Language Processing (NLP)-inspired methods, Transformer and Autoencoder, were used to capture molecular information from SMILES strings. These methods treat molecular sequences like text, learning patterns and relationships.
- 3D Structure-based Embeddings: Topological Laplacians and Algebraic Graphs were utilized to extract geometric and topological features from the 3D structures of molecules. Topological Laplacians, in particular, capture persistent topological invariants and shape information.
- Machine Learning Algorithms: These molecular embeddings were then fed into two powerful machine learning algorithms:
- Gradient Boosting Decision Tree (GBDT): An ensemble method known for its robustness and ability to handle complex datasets.
- Deep Neural Network (DNN): A multi-layered neural network capable of learning intricate patterns from high-dimensional data.
- Consensus Model: To enhance predictive power and robustness, the predictions from individual models were combined using a consensus approach (averaging probability scores).
- Classification and Regression: A binary classification model was first developed to identify potential hERG blockers versus non-blockers. Subsequently, a regression model was built to quantitatively predict the binding affinity of these identified blockers.
Datasets
The study utilized several publicly available hERG blockade datasets for training and validating their models:
- Braga et al. (52)
- C. Zhang et al. (53)
- Li et al. (54)
- Cai et al. (55)
- Doddareddy et al. (24)
- X. Zhang et al. (28)
- Ogura et al. (27) - This was noted as the largest hERG blockade dataset available.
The primary target for screening was the DrugBank database (version 5.1.9), containing 8641 compounds.
Key Insights and Ideas
- Synergy of 2D and 3D Information: The paper strongly suggests that combining sequence-based (2D) and structure-based (3D) molecular representations leads to superior predictive performance compared to using either alone. This highlights the importance of capturing both the chemical sequence and the spatial arrangement of atoms.
- Topological Laplacians as Powerful Descriptors: The use of topological Laplacians for 3D embedding proved particularly effective, capturing intricate geometric and topological features that are crucial for understanding molecular interactions.
- Consensus Models Enhance Robustness: Averaging predictions from multiple models (a consensus approach) significantly improved the overall accuracy and reliability of the predictions, mitigating the weaknesses of individual models.
- DrugBank as a Rich Source for Safety Screening: The successful application to the DrugBank database demonstrates its utility for identifying potential safety liabilities in both approved and investigational drugs, offering a proactive approach to drug safety.
- Potential for Hidden Cardiotoxicity: The identification of 92 FDA-approved drugs as potential hERG blockers raises concerns about existing medications and underscores the need for continuous safety monitoring and re-evaluation.
Key Results
- Classification Accuracy: The consensus model achieved high accuracy across various benchmark datasets, with performance comparable to or exceeding state-of-the-art methods. For instance, on the Ogura et al. dataset, the model achieved an accuracy of 0.981.
- DrugBank Screening: The model identified 227 out of 8641 DrugBank compounds as potential hERG blockers.
- Of these, 92 are FDA-approved drugs.
- The remaining 135 are experimental or off-market drugs.
- Binding Affinity Prediction: The regression model provided quantitative binding affinity predictions for these potential blockers.
- Consistency: The classification and regression models showed high consistency, with predicted IC50 values for 85 out of 92 potential blockers being less than 10 µM.
- Identification of High-Risk Drugs: The study highlighted several drugs, including Ibutilide and Dofetilide, as potent predicted blockers, with findings validated by existing literature and clinical observations. Some of these drugs have been withdrawn due to cardiotoxicity.
Limitations and Future Work
- Experimental Validation Needed: While the models demonstrate high predictive power, the authors emphasize that experimental validation is crucial to confirm the predicted hERG liability and cardiotoxicity of the identified compounds.
- Similarity Scores for Low-Confidence Predictions: For compounds with relatively low similarity scores to the training data, further investigation is recommended to determine their hERG side effects.
- Broader Application: The authors suggest that their approach can be extended to screen other large drug databases or to investigate other drug-related toxicities.
- Further Refinement of Embeddings: While topological Laplacians and Transformer embeddings showed strong performance, ongoing research into novel molecular embedding techniques could further enhance predictive capabilities.
Summary updated