Many bacteria are acquiring more resistance to usual treatments worldwide, to the point that the possible advent of pathogens resistant to the entire current arsenal is a true concern. Therefore, there is an urgent need for finding new effective antibacterial drugs. Associated to data mining methods, in silico ligand-based drug design techniques may extract the most relevant molecular features and eventually lead to the discovery of innovative potent antibacterial molecules. In this work, we use feature selection techniques to build molecular filters with demonstrated ability to discriminate between antibacterial and non-antibacterial small molecules. A very large number of molecular properties translated into molecular descrip-tors, being simultaneously diverse and redundant, were processed using various feature selection techniques. It is shown that this approach was efficient in decreasing the models complexity by identifying most relevant features for antibac-terial activity. For reducing the number of considered descriptors, we have trained multiple machine learning algorithms until resulting models performance in virtual screening could not be optimized further. We also discuss the interest of using log-linear analysis to improve our data-driven process and to increase the chance to predict efficiently new antibacterials.