Toward structure-based predictive tools for the selection of chiral stationary phases for the chromatographic separation of enantiomers

- Data mining the Chirbase database of >200,000 reported chromatographic enantiomer separations affords QSAR models for the prediction of which chiral stationary phase will work for a given compound.
- QSAR models for only 4 CSPs afforded good predictivity: DNB-Leu, Crownpak, Whelk-O, Chirobiotic T.
- Fair to good predictivity for successful separations was obtained.
- Excellent predicitivity for unsuccessful separations was obtained.
- More comprehensive inclusion of 'negative results' for separations that work poorly or not at all may be helpful in obtaining improved QSAR models for CSP selection.
- Precompetitive collaboration may be helpful in data sharing for generation of improved QSAR models for CSP selection.

ChirBase, a database for the chromatographic separation of enantiomers containing more than 200,000 records compiled from the literature, was used to develop quantitative structure activity models for the prediction of which chiral stationary phase will work for the separation of a given molecule. Constructuion of QSAR models for the enantioseparation of nineteen chiral stationary phases was attempted using only analyte structural information, leading to the producton of self-consistent models in four cases. These models were tested by predicting which in-house racemic compounds would and would not be resolved on the different columns. Some degree of success was observed, but the sparseness of data within ChirBase, which contains enantioseparations for only a subset of molecules on a subset of columns under a variety of conditions may limit the creation of effective models. Augmented data sets gleaned from automated chromatographic method development systems deployed in academic and industrial research laboratories or the use of models that take other factors such as solvent composition, temperature, etc. into account could potentially be useful for the development of more robust models.