[Paper Review] ToxTree: descriptor-based machine learning models for both hERG and Nav1.5 cardiotoxicity liability predictions

Issar Arab, Khaled Barakat|arXiv (Cornell University)|Dec 27, 2021

Computational Drug Discovery Methods4 citations

TL;DR

ToxTree introduces two robust 2D descriptor-based machine learning models—ToxTree-hERG and ToxTree-Nav1.5—for predicting cardiotoxicity liability via hERG and Nav1.5 channel blockade. Trained on curated datasets of 8,380 and 1,550 compounds, respectively, the models achieve state-of-the-art performance, with the Nav1.5 model reaching Q4 = 74.9% and Q2 = 86.7% on an external test set.

ABSTRACT

Drug-mediated blockade of the voltage-gated potassium channel(hERG) and the voltage-gated sodium channel (Nav1.5) can lead to severe cardiovascular complications. This rising concern has been reflected in the drug development arena, as the frequent emergence of cardiotoxicity from many approved drugs led to either discontinuing their use or, in some cases, their withdrawal from the market. Predicting potential hERG and Nav1.5 blockers at the outset of the drug discovery process can resolve this problem and can, therefore, decrease the time and expensive cost of developing safe drugs. One fast and cost-effective approach is to use in silico predictive methods to weed out potential hERG and Nav1.5 blockers at the early stages of drug development. Here, we introduce two robust 2D descriptor-based QSAR predictive models for both hERG and Nav1.5 liability predictions. The machine learning models were trained for both regression, predicting the potency value of a drug, and multiclass classification at three different potency cut-offs (i.e. 1$μ$M, 10$μ$M, and 30$μ$M), where ToxTree-hERG Classifier, a pipeline of Random Forest models, was trained on a large curated dataset of 8380 unique molecular compounds. Whereas ToxTree-Nav1.5 Classifier, a pipeline of kernelized SVM models, was trained on a large manually curated set of 1550 unique compounds retrieved from both ChEMBL and PubChem publicly available bioactivity databases. The proposed hERG inducer outperformed most metrics of the state-of-the-art published model and other existing tools. Additionally, we are introducing the first Nav1.5 liability predictive model achieving a Q4 = 74.9% and a binary classification of Q2 = 86.7% with MCC = 71.2% evaluated on an external test set of 173 unique compounds. The curated datasets used in this project are made publicly available to the research community.

Motivation & Objective

• To address the high failure rate in drug development due to hERG and Nav1.5-mediated cardiotoxicity.
• To develop fast, cost-effective in silico models for early prediction of cardiotoxicity liability.
• To create publicly available, high-quality curated datasets for hERG and Nav1.5 bioactivity data.
• To improve prediction accuracy over existing tools using advanced machine learning pipelines.

Proposed method

• ToxTree-hERG uses a Random Forest pipeline trained on 8,380 unique compounds from public bioactivity databases.
• ToxTree-Nav1.5 employs a kernelized SVM pipeline trained on 1,550 manually curated Nav1.5 compounds from ChEMBL and PubChem.
• Both models use 2D molecular descriptors as input features to predict potency and classification outcomes.
• The models are evaluated using external test sets with standard metrics including Q2, Q4, and Matthews Correlation Coefficient (MCC).
• Regression and multiclass classification are performed at three potency thresholds: 1 µM, 10 µM, and 30 µM.
• The datasets used for training are made publicly available to support reproducibility and further research.

Experimental results

Research questions

• RQ1 Can descriptor-based machine learning models accurately predict hERG and Nav1.5 cardiotoxicity liability in early drug discovery?
• RQ2 How do ToxTree-hERG and ToxTree-Nav1.5 models compare to state-of-the-art tools in predictive performance?
• RQ3 What level of accuracy can be achieved using curated, publicly available datasets for Nav1.5 and hERG bioactivity?
• RQ4 Can a single unified pipeline effectively model both hERG and Nav1.5 liabilities using 2D molecular descriptors?
• RQ5 What is the performance of the Nav1.5 model on an external test set using multiclass and binary classification metrics?

Key findings

• The ToxTree-hERG model outperforms most state-of-the-art models in predictive performance for hERG liability.
• The ToxTree-Nav1.5 model achieves a Q4 of 74.9% on an external test set of 173 unique compounds.
• The Nav1.5 model reaches a binary classification Q2 of 86.7% with a Matthews Correlation Coefficient (MCC) of 71.2%.
• The curated datasets used in the study are publicly available to support community use and future model development.
• The models demonstrate strong generalization on external test sets, indicating robustness for early-stage drug safety screening.
• The integration of 2D descriptors with ensemble and kernelized learning methods enables high-accuracy prediction of cardiotoxicity.