Combination of Similarity Rankings Using Data Fusion

■ INTRODUCTION Similarity searching is one of the most common techniques for ligand-based virtual screening and involves scanning a chemical database to identify those molecules that are most similar to a user-defined reference structure using some quantitative measure of intermolecular structural similarity. The similar property principle states that molecules that are structurally similar are also likely to exhibit similar properties, and hence ranking a database in order of decreasing similarity with a bioactive reference structure is expected to highlight database structures that have high a priori probabilities of exhibiting the same activity. Any similarity measure has three principal components: the representation that is used to describe each of the structures that are to be considered; the weighting scheme that is used to assign weights to different parts of the structure representation that reflect their relative degrees of importance; and the similarity coefficient that is used to quantify the degree of resemblance between two suitably weighted representations. Multiple approaches have been described for each of these three components, resulting in a potentially vast number of similarity measures that could be used for virtual screening. This multiplicity has resulted in many studies that seek to determine which measures are most effective using some quantitative measure of screening performance. While certain types of measures are known to provide a reasonable level of retrieval effectiveness (e.g., those making use of Pipeline Pilot ECFP_4 fingerprints, of occurrence-based frequency weighting, or of the Tanimoto similarity coefficient), there is a general recognition that there is no single similarity measure that will provide optimal screening in all circumstances. This situation has been well summarized by Sheridan and Kearsley, when they note that “we have come to regard looking for ‘the best’ way of searching chemical databases as a futile exercise. In both retrospective and prospective studies, different methods select different subsets of actives for the same biological activity and the same method might work better on some activities than others”. Given that there is no single, consistently effective, similarity searching method that can be used to rank a database in decreasing similarity order, it has been suggested that multiple searches should be carried out. The results of these individual searches are then combined (or merged or fused) into a single ranking that is the final output presented to the user for subsequent compound selection and biological testing. Such combination approaches have been used not only in similarity searching and other types of ligand-based virtual screening, where they are normally referred to as data fusion, but also in structure-based virtual screening, where they are normally referred to as consensus scoring. There is hence much interest in combining these two approaches to virtual screening. This perspective discusses the use of data fusion in similaritybased virtual screening; other similarity-related applications of data fusion include the analysis of molecular diversity and of structure−activity landscapes inter alia. A previous review provided an overview of data fusion methods in ligand-based virtual screening up to 2005. However, the technique has now been so widely adopted that it is difficult to provide a comprehensive review, with a Google Scholar search in late 2012 for “Data Fusion” AND “Virtual Screening” identifying over 350 post-2005 items. Accordingly, after a description of the basic approach and the various ways in which it can be implemented in the next section, the review focuses on two specific aspects of data fusion: the various fusion rules that have been described in the literature for combining rankings; and work at Sheffield that seeks to provide a rationale for why data fusion methods work in practice. The focus here is the combination of similarity rankings but many of the methods described here are equally applicable to the combination of the rankings that result from the use of, e.g., machine learning techniques for screening chemical databases.