Predicting Molecule Toxicity Using Deep Learning

Abstract

Knowledge about certain features of molecules is crucial in many fields such as drug design. Toxicity of a compound gives information on the degree to which it can be harmful to a living organism. Experimental evaluation of a molecule’s toxicity requires specialised staff and equipment and generates high costs. Deep learning is a tool used in many fields and can also be helpful in the case of predicting the molecule’s toxicity. For the dataset we used a free database called ‘SMILES Toxicity’ which is available on kaggle. This dataset consists of 6698 non-toxic molecules and 964 toxic ones. Its imbalance creates the problem called ‘overfitting’. In this article, we describe several methods for overcoming the overfitting problem. We developed two models with well-balanced true positives to false positives ratio. With the first model, we achieved 96% accuracy on the training set, 89% on the validation set and 65% on the test set. The second model scored the following values of accuracy: 80% on the training set, 78% on the validation set and 77% on the test set.