The efficiency of TiO2 photocatalysis to degrade three estrogen hormones, that is, 17 alpha-ethynylestradiol (EE2), estrone (E1), and 17 beta-estradiol (E2), in environmentally relevant samples was investigated. Radiation at a photon flux of 1.7 x 10(-7) einstein/s was provided by a solar simulator and experiments were conducted at various TiO2 loadings (25-1500 mg/L), estrogen concentrations (85-300 mu g/L in individual solutions and 400 mu g/L in mixture), and water matrices (ultrapure water, drinking water, and secondary treated wastewater). Changes in estrogen concentration were followed by high performance liquid chromatography and estrogenicity by the yeast estrogen screening assay. Aeroxide P25, a commercially available mixture of 75:25 anatase/rutile, was considerably more active than carbon-doped and undoped anatase titania, with degradation increasing with increasing catalyst loading and treatment time. The organic and inorganic constituents typically found in wastewater and drinking water impeded degradation presumably due to the scavenging of oxidizing species. For example, the time needed for complete 100 mu g/L EE2 degradation in pure water was an order of magnitude lower than that in wastewater. The three estrogens exhibited comparable reactivity, with E1 being slightly more reactive than the rest. Degradation in multicomponent mixtures was slower than in individual solutions, thus implying estrogen competition for oxidizing species. Although the mixture of three in wastewater could be degraded fully after 120 min, overall estrogenicity was reduced by just about 30%, highlighting the role of the complex water matrix. Several oxidation products of EE2 were identified by means of LC-MS/MS and a reaction network for the photocatalytic degradation of EE2 is suggested. An artificial neural network comprising five input variables (reaction time, TiO2 and EE2 concentration, organic content, and conductivity of the water matrix), eight neurons and an output variable (conversion) was optimized, tested, and validated for EE2 degradation. The network, based on tangent sigmoid and linear transfer functions for the hidden and input/output layers, respectively, and the Levenberg-Marquardt back-propagation training algorithm, can successfully predict EE2 degradation.