Local structure-induced atomic features (LEAFs) are extracted through comprehensive statistical analysis of the geometries of inorganic crystal structures. As a result, LEAFs afford modelling of materials solely as compositions while incorporating structural insights. LEAFs outperform other sets of elemental features in structure-focused tasks, such as crystal structure prediction, and can suggest likely atomic substitutions and doping elements for material design.