Enhancing the Effectiveness of Ligand‐Based Virtual Screening Using Data Fusion

Abstract

Data fusion is being increasingly used to combine the outputs of different types of sensor. This paper reviews the application of the approach to ligand-based virtual screening, where the sensors to be combined are functions that score molecules in a database on their likelihood of exhibiting some required biological activity. Much of the literature to date involves the combination of multiple similarity searches, although there is also increasing interest in the combination of multiple machine learning techniques. Both approaches are reviewed here, focusing on the extent to which fusion can improve the effectiveness of searching when compared with a single screening mechanism, and on the reasons that have been suggested for the observed performance enhancement.

Effectiveness of 2D Fingerprints for Scaffold Hopping

“…An obvious approach is hence to combine the outputs of different scaffold-hopping methods that exploit both types of information. This can be done by applying data fusion to ranked search outputs [28] or, if appropriate training data are available, by using belief theory [29]. Thus, Muchmore et al have described the use of the latter approach on Abbott internal data, and demonstrated that effective scaffold-hopping can be achieved by combining 2D and 3D similarity measures (specifically fingerprints encoding ECFP6 circular substructures, analogous to, but larger than, the ECFP4 ones used here, and the ROCS shape-similarity software from OpenEye Scientific [107]).…”

Analysis and use of fragment-occurrence data in similarity-based virtual screening

“…We also intend to study the use of data fusion to combine rankings produced using different occurrence-based weighting schemes [58], and to explore the use of weighting schemes that take account of the frequency with which a fragment occurs in an entire database of molecules (rather than its frequency in a single molecule as here) [24]. Table 3.…”

Chemoinformatics at the University of Sheffield 2002–2014

“…[29] Thus far, its main application in chemoinformatics has been for ligand-based or structure-based virtual screening (where, in the latter case, it is often referred to as consensus scoring). [30][31] The basic idea underlying data fusion is that the use of multiple sources of evidence will result in better decision making than if only a single source of evidence is available. Thus, in the context of ligand-based virtual screening, the use of multiple screening methods is expected to increase the extent to which active molecules are clustered at the top of a database ranking, with our work on data fusion focussing on the combination of multiple similarity measures.…”

“…Work in Sheffield and elsewhere [30][31] has shown clearly that fusion-based screening is often comparable with, or superior to, the best of the individual similarity searches that are being combined and that screening effectiveness is much more consistent from search to search than when just a single similarity method is available. Is there an underlying theoretical rationale for this empirical finding?…”