|
Statistical Molecular Design, part 5 |
|
|
|
|
Written by Lennart Eriksson, Per M Andersson, Erik Johansson and Torbjörn Lundstedt
|
We have in the four previous tutorials reviewed the concept of statistical molecular design (SMD) and shown some illustrations of where and when this technology may be used, see previous tutorials [1 - 4]. The objective of this fifth editorial is to consider ways to extend the basic SMD-approach, so that also more demanding problems may be addressed. In so doing, we will also come across some of the latest developments in this field.
In the early days of SMD -- when this approach was called multivariate design -the design protocol used was usually a member of the full factorial or fractional factorial  design families. As described in [1], a factorial design defines a regular cube or hypercube, which are useful geometrical structures for design-making among controllable factors. However, in the context of SMD, which involves design-making among molecular property factors, configurations spanned by the selected experiments (i.e., compounds) rarely correspond to such regular geometries. Hence, full and fractional factorial designs are sometimes felt a bit too stiff and intractable for QSAR.An interesting alternative is provided by the D-optimality approach, and this was also partly hinted at in the first editorial [1]. A D-optimal design is more flexible than a full or fractional factorial design and may therefore better adapt to severely constrained QSAR problems . We will in this editorial shed some more light on the D-optimality approach and discuss when and where it is applicable. D-optimal design is a key ingredient in recent extensions of SMD, notably cluster-based design, and D-optimal onion designs. One objective of this editorial is to outline the features of these extensions. Additionally, we will also consider the principles of hierarchical SMD .
|
Tutorial