Molecular similarity: Methods and performance

Molecular similarity has long been a hot topic, which has been evaluated and compared by various approaches and plays a significant role in protein-ligand and protein-protein interactions recognition. There are currently many types of molecular similarity evaluation methods with their own advantages and disadvantages. Molecular fingerprints are the most common methods for molecular similarity evaluation which only concern about rapid 2D common substructure retrieval but lack the ability to encode the information about 3D conformers. 3D molecular descriptor based methods bear the advantages of representing the structure information of a conformer, but the descriptors are not guaranteed to describe the molecules precisely. Molecular alignment based methods try to superimpose two molecules and evaluate the similarity using the optimal poses which are generally more precise than the molecular descriptor but require a time-consuming optimization process. Pharmacophore based methods only focus on the chemical features about a molecule and are not capable of dealing with the molecular shape similarity. In order to evaluate the performance of molecular similarity based screening, many kinds of metrics are available, e.g., visual representation, quantitative measurements and scaffold hopping ability measurements. Further applications of molecular similarity include construction of molecule interaction network or generation of diverse compounds library. Some important aspects of molecular similarity are reviewed in this paper, including the implementations of various similarity evaluation methods, summaries and characteristics of a number of molecular databases, different performance evaluation metrics, etc.