Computational library design selection methods

In the context of general screening, computational library design involves the selection of a subset of compounds that are optimally diverse and representative of available classes of compounds, leading to a non-redundant chemical library for biological testing. Methods reported in the literature include... 

Cluster-based and dissimilarity-based methods for compound selection were first discussed in the Eighties but it is only in the last few years that the area has attracted substantial attention as a result of the need to provide a rational basis for the design of combinatorial libraries. The four previous sections have provided an overview of the main types of selection method that are already available, with further approaches continuing to appear in the literature. Given this array of possible techniques, it is appropriate to consider ways in which the various methods can be evaluated, both in absolute terms and when compared with each other. A method can be evaluated in terms of its efficiency, /.< ., the computational costs associated with its use, and its effectiveness, /.< ., the extent to which it achieves its aims. As we shall see, it is not immediately obvious how effectiveness should be quantified and we shall thus consider the question of efficiency first, focusing upon the normal algorithmic criteria of CPU time and storage requirements. 

The chapter begins with a discussion of similarity and diversity measures and how they can be applied in a virtual screening context. The various computational filters in use are also discussed. The rest of the chapter is concerned with different approaches to combinatorial library design, beginning with reagent-based methods followed by product-based approaches of cherry picking and combinatorial subset selection. Finally, approaches to designing libraries optimized on multiple properties simultaneously are discussed. 

Product-based selection is much more computationally demanding than reagent-based selection. Typically, it requires the computational enumeration of the full virtual combinatorial library and calculation of the descriptors for all possible products, prior to the application of a subset selection method. Consider a three-component reaction with 100 reagents available at each substituent position and assume that the aim is to build a 10 x 10 x 10 combinatorial library. In reagent-based selection, this requires the calculation of descriptors for 300 compounds (100 + 100 + 100). In product-based design, however, the full library of 1 million compounds (100 x 100 x 100) must be enumerated and descriptors must be calculated for each product molecule. 

Multipoint pharmacophore approaches to library design are computationally demanding because they require the exploration of conformational space for each product molecule prior to calculating the pharmacophore keys. Thus, Beno and Mason [19] report that the four-point pharmacophore method is currently limited to selecting subsets of a few hundred compounds. 

---