TOWARDS MORE RELIABLE DRUG TOXICITY PREDICTION: AN ENSEMBLE APPROACH

Abstract

The development of a single pharmaceutical drug is a time- and resource-consuming process with a high likelihood of rejection. In recent years, the cost-effectiveness of a single drug has decreased drastically, as the criteria for passing has become more rigorous. A huge fraction of attrition rates is caused by the toxicity of chemical compounds. Recent findings in Machine Learning (ML) have revolutionized the drug toxicity prediction field, developing many model architectures and data representations. The faced challenges are different ways of representing the molecules’ chemical structure, as well as many different toxicity types. This study proposes a novel drug toxicity prediction framework. It uses several classification models, based on different data representations and different ways of combining their features. The evaluation of six different datasets with different toxicity types shows that choosing majority voting across all models can improve the ROC AUC score and accuracy. Using a single classification model to combine these datasets demonstrates that it is possible to achieve 84% accuracy on data with various toxicity types. The findings of this research provide insights into the application of ML in pharmaceutical research. Improving current methods of toxicity assessment can have a positive effect on the efficiency and cost-effectiveness of drug development.