The toxicity data of 28 nitroaromatic compounds (nitrobenzenes and, for comparison, benzene and toluene) related to a 50% lethal dose concentration for rats (LD50) were used to develop quantitative structure-activity relationships (QSARs).A genetic algorithm and multiple regression analysis were applied to select the descriptors and to generate the correlation models. The obtained equations consist of one to three descriptors. A number of molecular descriptors was obtained from HF/6-31G(d) and DFT (B3LYP/6-311+G(d, p)) level calculations. The calculated molecular geometry and electronic properties were evaluated by comparison with the available experimental data (where applicable). All parameters obtained at the B3LYP/6-311+G(d, p) level and the topological descriptors derived from this geometry were found to be reliable, except for dipole moment, due to the large uncertainty of its estimation. Satisfactory relationships were observed for the one-parameter structure-toxicity models between topological (X5Av, Ms) and quantum-chemical (ELUMO) descriptors. For better predictability two- and three-parameter QSAR analyses were performed. These analyses resulted in much better equations with correlation coefficient values r = 0.872-0.924. These models have been obtained with a set of topological, fragment and quantum-chemical descriptors (Ms, PCR, PCD, BELe1, C-026 and ELUMO). The toxicity of nitroaromatic compounds appears to be governed by a number of factors, such as the number of nitrogroups, the electrotopological state, the presence of certain fragments and the electrophilicity/reactivity parameter (ELUMO). Nitrobenzenes exhibited electrophilic reactivity (as was shown by correlation of the toxicity with the energy of the lowest unoccupied orbital, ELUMO). The toxicity LD50 parameter for rats has been utilized for the first time for QSAR analysis of nitrobenzenes. The predictive ability of the models is determined by a cross-validation "leave-one-out" method.