Stereochemically-aware bioactivity descriptors for uncharacterized chemical compounds

We recently presented a set of deep neural networks to generate bioactivity descriptors associated to small molecules (i.e. Signaturizers), capturing their effects at increasing levels of biological complexity (i.e. from protein targets to clinical outcomes). However, such models were trained on 2D representations of molecules and are thus unable to capture key differences in the activity of stereoisomers. Now, we systematically assess the relationship between stereoisomerism and bioactivity on over 1M compounds, finding that a very significant fraction (~40%) of spatial isomer pairs show, to some extent, distinct bioactivities. We then used these data to train a second generation of Signaturizers, which are now stereochemically-aware, and provide an even more faithful description of complex small molecule bioactivity properties.