SiSi3D: Using Significantly Similar Subgraphs for Similarity Search in 3D
Tobias Girschick, Technical University Munich, Germany
Our approach - SiSi3D (significant similarities in 3D) - addresses the
task of similarity search for 3D molecular graphs, where atoms are
annotated with coordinates in three-dimensional Euclidean space and
atomic descriptors. The similarity of two molecules is determined by the
occurrence of significantly similar subgraphs, more precisely sets of
annotated paths spanning the molecule. SiSi3D not only finds similar
molecules, it also provides significance estimates of how unlikely it is
to obtain such similarity based on a statistical background model.
Our
experiments on two drug classes (sartans and statins) as well as on
several PubChem BioAssays show that SiSi3D outperforms standard
feature-vector (PubChem fingerprint) and maximum common subgraph (MCS)
based methods for identifying structures with similar modes of action in
a database of unrelated structures.
(presenting author: Tobias Girschick)