Mini-fingerprints for virtual screening: Design principles and generation of novel prototypes based on information theory

Abstract

Binary fingerprint representations of molecular structure and properties are convenient computational tools for similarity searching in compound databases and virtual screening (VS). We are investigating the design of relatively simple fingerprints for the identification of molecules having similar biological activity and recognition of remote similarity relationships. Since our designs are considerably shorter than other fingerprints used in VS, we have previously termed them "mini-fingerprints" (MFPs). A key aspect of the design strategy is the identification of suitable molecular descriptors. Whereas our initial fingerprint designs have relied on descriptor combinations that performed well in compound classification according to biological activity, second generation MFPs encode combinations of descriptors with high information content in large compound databases and high frequency of occurrence in drug-like molecules. Thus, the design of these new fingerprints does not depend on the analysis of specific classes of bioactive compounds, but rather on descriptor information content in large compound databases. Systematic evaluation of fingerprint performance in VS test calculations demonstrates that these new prototypes perform better than previously generated MFPs. The analysis described herein provides an example for the development of search tools for VS.