Stacking (Wolpert (1992), Breiman (1996)) is known to be a successful way of linearly combining several models. We modify the usual stacking methodology when the response is binary and predictions highly correlated,by combining predictions with PLS-Discriminant Analysis instead of ordinary least squares. For small data sets we develop a strategy based on repeated split samples in order to select relevant variables and ensure the robustness of the nal model. Five base (or level-0) classiers are combined in order to get an improved rule which is applied to a classical benchmark of UCI Machine Learning Repository. Our methodology is then applied to the prediction of dangerousness of 165 chemicals used in the cosmetic industry, described by 35 in vitro and in silico characteristics, since faced to safety constraints, one cannot rely on a single prediction method, especially when the sample sizeis low.