We present a new feature selection algorithm for structure-activityand structure-property correlation based on particle swarms. Particle swarmsexplore the search space through a population of individuals, which adaptby returning stochastically towards previously successful regions, influencedby the success of their neighbors. This method, which was originally intendedfor searching multidimensional continuous spaces, is adapted to the problemof feature selection by viewing the location vectors of the particles asprobabilities, and employing roulette wheel selection to construct candidatesubsets. The algorithm is applied in the construction of parsimonious QSARmodels based on feed-forward neural networks, and tested on three classicaldata sets from the QSAR literature. It is shown that the method comparesfavorably with simulated annealing, and is able to identify a better andmore diverse set of solutions given the same amount of simulation time.