A machine learning model to predict small molecules with antischistosomal activity

Caused by parasitic worms that live in fresh water, schistosomasis is a disease that affects 251 million people world wide. This study proposes a computer-aided drug design system to predict small molecules that can inhibit schistosomal activity, and thus could be used as a treatment for this disease. We introduce a regression model that estimates the IC50 value for each molecule and a classification model that predicts whether a certain molecule is active or not against the Schistosoma mansoni parasite. We acquire a set of active and nonactive molecules from the ChemBL database and generate descriptors of those molecules using the rdkit library. The resulting features are preprocessed and fed into machine learning models to perform regression and classification tasks. In both cases, the artificial neural network scored the highest accuracy while tested on a set of 754 molecules using cross-validation techniques. Finally, we train a genetic algorithm using the proposed regression model within its fitness function and retrieve 3 molecules that have the best fitness values and have not been screened before.