Abdulk084/HybridTox2D

In recent times, toxicological classification of chemical compounds is considered to be a grand challenge for pharma-ceutical and environment regulators. Advancement in machine learning techniques enabled efficient toxicity predic-tion pipelines. Random forests (RF), support vector machines (SVM) and deep neural networks (DNN) are often ap-plied to model the toxic effects of chemical compounds. However, complexity-accuracy tradeoff still needs to be ac-counted in order to improve the efficiency and commercial deployment of these methods. In this study, we implement a hybrid framework consists of a shallow neural network and a decision classifier for toxicity prediction of chemicals that interrupt nuclear receptor (NR) and stress response (SR) signaling pathways. A model based on proposed hybrid framework is trained on Tox21 data using 2D chemical descriptors that are less multifarious in nature and easy to calcu-late. Our method achieved the highest accuracy of 0.847 AUC (area under the curve) using a shallow neural network with only one hidden layer consisted of 10 neurons. Furthermore, our hybrid model enabled us to elucidate the inter-pretation of most important descriptors responsible for NR and SR toxicity.

Jupyter NotebookMIT