Quantitative Structure–Activity Relationship (QSAR) modeling is a common method for predicting the bioactivity and physicochemical properties of chemical compounds. Artificial Intelligence algorithms are being tested and applied to build these QSAR models. The challenge lies in finding optimal descriptors for molecules and building reliable predictive models that can be used, for instance, in drug discovery. This paper investigates the usage of a deep neural network for machine learning and the prediction of bioactivity together with an approach for feature selection using a genetic algorithm for numerical experiments based on a Python implementation making use of several libraries. Our results indicate that approximately half of the features can be evaluated without deterioration of prediction quality, which may allow for smaller computation times and improved interpretability of the results.