Similarity-based virtual screening using bayesian inference network: enhanced search using 2D fingerprints and multiple reference structures

Abstract

It has been known that different reference structure retrieve different sets of structures. Recent works in similarity searching have suggested that significant improvements in retrieval effectiveness can be achieved by combining results from different reference structures. One of an important characteristic of the Bayesian inference network (BIN) model is that permits the combining of multiple reference structures. In this paper we introduce a formal inference net model to directly combine the contributions of multiple reference structures, and propose a novel approach to the combination of information from various reference structures. The inference net model of similarity, which was designed from this point of view, treats similarity searching as an evidential reasoning process where multiple sources of evidence about target structure are combined to estimate similarity scores. In this paper, we have compared BIN with other similarity searching methods when multiple bioactive reference structures are available. Six different 2D fingerprints were used in combination with data fusion (DF) and nearest neighbor (NN) approaches as search tools and also as descriptors for BIN. Our empirical results show that the BIN consistently outperformed all conventional approaches such as DF and NN, regardless of the fingerprints that were tested. The superiority of BIN over conventional approaches is ascribed to the fact that BIN understands the content of the descriptors of the structures and references and used this understanding to infer the direct relationship between structures and references.