Combination Rules for Group Fusion in Similarity-Based Virtual Screening

This paper evaluates the screening effectiveness of 15 parameter-free, similarity-based and rank-based rules for group fusion, where one combines the outputs of similarity searches from multiple reference structures using ECFC_4 fingerprints and a Bayesian inference network. Searches of the MDDR and WOMBAT databases show that group fusion is most effective when as many reference structures as possible are used, when only a small proportion of each ranked similarity list is submitted to the final fusion rule, and when a fusion rule based on reciprocal rank positions is used to combine the individual search outputs. An analysis of the reciprocal rank rule suggests that its effectiveness derives from the close relationship that exists between the reciprocal rank of a database structure and its probability of activity.