Abstract This study investigates the ability of neural networks to estimate by a quantitative structure–activity relationship the adsorbability onto activated carbon of a wide range of organics in water. Adsorption data from several sources and molecular connectivity indexes constituted a training data set of 368 aliphatic and aromatic compounds. The neural network architecture was optimized to obtain a three-layer neural network, composed of four input neurons, three hidden neurons and one output neuron. A good correlation was obtained ( r 2 =0.875), the statistical results being better and less time-consuming than those obtained by Blum et al. (1994) [Blum D. J. W., Suffet I. H. and Duguet J. P. (1994) Quantitative structure–activity relationship using molecular connectivity for the activated carbon adsorption of organic chemicals in water. Water Res . 28(3), 687-699] with the same data set using a regression analysis. The predictive ability of the neural network was tested with 16 new compounds. The results were satisfactory but slightly lower than those obtained by the regression analysis. The weight–partitioning method was applied to assess the relative importance of the input variables on the adsorbability.