Group fusion

Figure 1. Comparison of the average recall at 5% obtained with group fusion (GF) and similarity searching (SS) using the ECFP 4 descriptors and 30 MDDR activity classes chosen as being of low, medium and high diversity. The upper part of the figure shows the recall at 5% obtained with GF vs. recall at 5% obtained with SS, while the lower part shows the diversity (as measured by the mean pair-wise similarity) vs. the ratio of the recalls at 5% obtained with GF and with SS.

Williams, C. (2006) Reverse fingerprinting, similarity searching by group fusion and fingerprint bit importance. Mol. Div., 10, 311-332. 

The combination of virtual screening results with several reference ligands and/or several search methods is called data fusion. Whittle et al. [49, 50] refer to the combination of similarity searches for one reference ligand with different descriptors as similarity fusion, whereas the combination of the results for a set of reference ligands with one method is called group fusion. Many reports on different methods and results of data fusion approaches can be found in the literature [41, 46, 51-53]. 

Data fusion can be performed in a number of ways. Most often used are the MIN, MAX, and SUM rules. According to these rules, the minimum or maximum score value or the sum of the scores is taken from a list of different similarity values for a compound that result from different virtual screening runs. It was also recommended to consider only the ranks of the compound in the corresponding nearest-neighbor lists [51]. Hert et al. [41] observed the superiority of the SUM rule when fusion is performed on ranks, whereas the MAX rule gives better results if the fusion is performed on scores, which is in line with earlier results from Schuffenhauer et al. [52]. Whittle et al. [50] observed a superiority of the MAX rule over the SUM rule for group fusion. 

There are two major approaches, nominally called similarity fusion and group fusion. Application of these procedures is exemplified by similarity searches carried out in LEVS studies [58, 59]. In the case of similarity fusion, a set of similarity values is computed with respect to a single active reference (query) molecule, tj p, using a number of different similarity measures, y ), where A"= 1,2,...,... 

A variant of group fusion called turbo similarity is employed when a single reference structure is available [138]. It is an iterative procedure that takes a subset of the retrieved compounds with high similarity to the active reference compound and uses these hits, whether or not they are active, as reference compounds in the next iteration—thus, the correspondence of this procedure with group fusion. The process, which can be continued if desired, is also reminiscent of document retrieval methods that take the hits from a given query as queries for subsequent retrievals [51]. 

Chen B, Mueller C, Willett P. Combination rules for group fusion in similarity-based virtual screening. Mol Inf 2010 29 533-541. 

Determine corresporxiing rankings arxi compute group fusion values... 

Fig. 1.6 Data tables illustrating group fusion of similarity and rank values a and b depict the procedure for group fusion of similarity values, c, d, and e depict the corresponding procedure for...

The prevalence of similarity cliffs noted above also provides a rationale, albeit a tentative one, as to why group fusion (Sect. 1.2.3.2) performs as well does. Numerous analyses by Willett and his colleagues show that it appears to work best with diverse rather than highly similar reference sets [92, 94, 106]. Their conclusion... 

Group fusion, on the other hand, employs multiple reference actives and, as noted above, performs best when the reference compounds are as diverse as possible. Hence, the dispersion of active compounds is explicitly accounted for by the method, although the available reference set may not, in many cases, provide sufficient coverage of all of the regions of CS that contain active compounds with respect to the given assay, and some actives will undoubtedly be missed. 

Because group fusion uses either the MAX rule for similarities or the RRF rule for rankings, compounds located close to the reference compounds are given preference over more distant, less similar compounds, a situation that accords well with the SPP since compounds located close to known actives are more likely to also be active than are less similar compounds. Thus, the performance of group fusion can be rationalized by the significant presence of similarity cliffs in activity landscapes. 

This, however, raises a new issue, namely, how ate multiple active reference compounds handled in the LEVS process There are several approaches to this problem. One way is through the use of group fusion described in Sect. 1.2.3.2, which is ideally suited to deal with this problem since multiple active reference compounds are an inherent feature of the method. And, as discussed in Sects. 1.2.3.2 and 1.2.4, group fusion exhibits excellent performance as a means for identifying new actives. Interestingly, group fusion based on the fusion maximum similarity or minimum distance values is essentially identical to an approach called hst-based searching [76, 78, 86],... 

Interestingly, the procedure appears to be a combination of group fusion (i.e., list-based searching) and similarity fusion. The reasons, the first two of which are associated with group and similarity fusion, are as follows (1) multiple active reference compounds are used, (2) the most similar (closest) compounds to each active reference compound are retained, and (3) multiple similarity measures are applied. 

---