For a data set of 661 organic chemicals including many drug-like compounds, discrimination between antibacterial and non-antibacterial activity was modeled using hydrophobicity in terms of the logarithmic octanol/water partition coefficient (log K_ow) and AM1-level molecular descriptors encoding geometric, electrostatic, nucleophilic and electrophilic characteristics of the compounds. Linear discriminant analysis (LDA) and binary logistic regression (BLR) achieved an overall classification rate of around 90%, using two to three variables selected from log K_ow, charged-weighted negative surface area (PNSA-3), positive surface area of heavy atoms (PPSA-1Z), and maximum donor delocalizability (D^E_max). Model validation was performed using complementary subsets for training and prediction as well as by training the total set with 50% of the activity data allocated wrongly in several arbitrarily selected ways. The discussion includes a comparative analysis of force-field and AM1 geometries as well as of the 3D variation of AM1-level molecular descriptors. Surprisingly, 3D geometry variations have only little impact on the discriminatory performance of the models.