Combination of Similarity Rankings Using Data Fusion

The use of data fusion in similarity-based virtual screening is studied. Data fusion is the name given to a body of techniques that combine multiple sources of data into a single source, with the expectation that the resulting fused source will be more informative than will the individual input sources. The scores that are merged by the fusion rule can be of two types, either the structure's actual similarity, as computed using some particular similarity measure; or the rank of the structure when all of the N computed similarities are ranked in decreasing order of the scores for the chosen similarity measure. It first found application in similarity fusion, where a single reference structure is searched using different similarity measures; it has since been extended to encompass multiple reference structures. Both approaches can benefit from the availability of training data linking similarity scores and probabilities of activity, but unsupervised fusion rules are available that enables effective searches to be carried out even in the absence of such data.