Solubility is one of the most important properties of drug candidates for achieving the targeted plasma concentrations following oral dosing. Furthermore, the formulations adopted in the in vivo preclinical studies, for both oral and intravenous administrations, are usually solutions. To formulate compounds sparingly soluble in water, pharmaceutically acceptable cosolvents or surfactants are typically employed to increase solubility. Compounds poorly soluble also in these systems will likely show severe formulation issues. In such cases, relatively high amount of compounds, rarely available in the early preclinical phases, are needed to identify the most appropriate dosing vehicles. Hence, the purpose of this study was to build two computational models which, on the basis of the molecular structure, are able to predict the compound solubility in two vehicle systems (40% PEG400/water and 10% Tween80/water) used in our company as screening tools for anticipating potential formulation issues. The two models were developed using the solubility data obtained from the analysis of approximately 2000 chemically diverse compounds. The structural diversity and the drug-like space covered by these molecules were investigated using the ChemGPS methodology. The compounds were classified (high/low preformulation risk) based on the experimental solubility value range. A combination of descriptors (i.e. logD at two different pH, E-state indices and other 2D structural descriptors) was correlated to these classes using partial least squares discriminant (PLSD) analysis. The overall accuracy of each PLSD model applied to independent sets of compounds was approximately 78%. The accuracy reached when the models were used in combination to identify molecules with low preformulation risk in both systems was 83%. The models appeared a valuable tool for predicting the preformulation risk of drug candidates and consequently for identifying the most appropriate dosing vehicles to be further investigated before the first in vivo preclinical studies. Since only a small number of 2D descriptors is need to evaluate the preformulation risk classes, the models resulted easy to use and characterized by high throughput.