Benzodiazepines belong to a large family of drugs, being used as hypnotics, anxiolytics, tranquillizers, anticonvulsants, in pre-medication and intravenous sedation. Several quantitative structure-toxicity (lethal oral dose for mouse) relationship (QSTR) models for 54 benzodiazepine derivatives have been developed. The molecular structure of these compounds was energetically optimized by molecular mechanics calculations. To the lowest energy conformations thus obtained, quantum chemical calculations (RM1 approach) were applied to finally optimize the structures. Several structural descriptors, volumes, molecular surface area, hydrophobicities and quantum chemical descriptors were calculated from the minimized structures. Multiple linear regression (MLR) combined with genetic algorithm for variable selection, artificial neural networks (ANNs), support vector machines (SVMs) and partial least squares (PLS) have been employed. Few satisfactory MLR models with predictive power were obtained. Nonlinear modelling methods of ANNs and SVMs gave somewhat better models than those obtained by MLR using same set of descriptors. Additional information on the factors which influence the benzodiazepine toxicity was given by PLS. The obtained models can be used for a rough evaluation of benzodiazepine toxicity.