Parts of ligands in complex crystal structures that are involved in well-defined protein-ligand interactions are extracted and encoded as ensembles of atom-centered fragments, termed atom-centered interacting fragments (AIFs), which implicitly capture three-dimensional interaction information. AIF reference sets are utilized for feature count-based ranking of databases to search for molecules having similar activity. AIF calculations are reported for eight enzyme targets with multiple crystallographic inhibitor complexes and, in addition, for a complex of a G protein coupled receptor with an antagonist. The AIF approach further increases compound recall of structural key-based fingerprints. Moreover, AIF combinations are found to be specific markers of different classes of active compounds that lead to an early enrichment of inhibitors and selective antagonists in similarity search calculations.