Determining molecular properties with differential mobility spectrometry and machine learning

The fast and accurate determination of molecular properties is highly desirable for many facets of chemical research, particularly in drug discovery where pre-clinical assays play an important role in paring down large sets of drug candidates. Here, we present the use of supervised machine learning to treat differential mobility spectrometry - mass spectrometry data for ten topological classes of drug candidates. We demonstrate that the gas-phase clustering behavior probed in our experiments can be used to predict the candidates' condensed phase molecular properties, such as cell permeability, solubility, polar surface area, and water/octanol distribution coefficient. All of these measurements are performed in minutes and require mere nanograms of each drug examined. Moreover, by tuning gas temperature within the differential mobility spectrometer, one can fine tune the extent of ion-solvent clustering to separate subtly different molecular geometries and to discriminate molecules of very similar physicochemical properties.