The untargeted profiling of steroids constitutes a growing research field because of their importance as biomarkers of endocrine disruption. New technologies in analytical chemistry, such as ultra highpressure liquid chromatography coupled with mass spectrometry (MS), offer the possibility of a fast and sensitive analysis. Nevertheless, difficulties regarding steroid identification are encountered when considering isotopomeric steroids. Thus, the use of retention times is of great help for the unambiguous identification of steroids. In this context, starting from the linear solvent strength (LSS) theory, quantitative structure retention relationship (QSRR) models, based on a dataset composed of 91 endogenous steroids and VolSurf þ descriptors combined with a new dedicated molecular fingerprint, were developed to predict retention times of steroid structures in any gradient mode conditions. Satisfactory performance was obtained during nested cross-validation with a predictive ability (Q2) of 0.92. The generalisation ability of the model was further confirmed by an average error of 4.4% in external prediction. This allowed the list of candidates associated with identical monoisotopic masses to be strongly reduced, facilitating definitive steroid identification.