GRID/PCA

For an interpretation of the GRID/PCA model in structural terms, contour plots for individual probes were derived from the loadings of the first and second PC,... 

In the last few years, a number of publications have demonstrated that the GRID/PCA or GRID/CPCA methods can be successfully applied to characterize the structural differences between protein binding sites, and to identify differences in the protein-ligand interactions as well as the regions on the target enzymes which mediate highly selective interactions [4—17]. 

In this chapter, we will first review the basic principles of GRID/PCA and GRID/CPCA, highlighting some important technical aspects necessary for their successful application. Some space will be given to the discussion of the differences between PGA and the (hierarchical) GPGA methods, before their application to various selectivity problems will be summarized. 

However, in many cases the proteins considered in the analysis are very similar and show only little structural variation. In these cases a superimposition of structurally conserved regions with standard protein homology modeling tools has yielded satisfactory protein superimpositions suited for the GRID/PCA or GRID/ GPGA analysis [8,9,13,17]. 

Figure 3.1. Procedure for building the X matrix in GRID/PCA. The analysis of the interaction energies of the m probes with the 2 target proteins produces 2xm three-dimensional matrices. These are unfolded to obtain 2 x m one-dimensional vectors from which the two-dimensional X matrix is built.

If most of the variation of the original data can be described by the first few PCs, a much simpler data structure exists. In GRID/PCA, typically about 60-75% of the variance can be explained with the first two PCs this allows an interpretation based on the 2D plots of the leading components. 

Figure 3.2. Typical PC 1 vs. PC 2 score plot obtained with GRID/PCA (tl vs. t2) (T. Fox, unpublished work). The points in the plot represent the objects of the X matrix the interactions of a particular probe with one of the two targets, in this example thrombin (thr) andt psin (try).

Whereas the GRID/PCA approach has been successfully used in a number of selectivity problems (see below), this procedure has several shortcomings ... 

We note that on the block level, an objective function is used to obtain the scores which is different from the standard PCA the principal components should also reproduce the values obtained in the overall PCA level. As a consequence, the percentage of the variance explained by PC 1 and PC 2 varies between 20 and 30% compared to the much higher values in the GRID/PCA approach. Also, within the blocks the scores are not necessarily ordered in decreasing importance. However, the separation of the objects and the interpretability of a model should be considered as more important criteria for the quality of a model than the percentage of the variance explained by PC 1 and PC 2. 

The first study to investigate selectivity profiles using MIFs in connection with a chemometric analysis was published about 10 years ago [5]. There, the behavior of all 64 possible DNA triplets with respect to 31 GRID probes was studied with GRID/PCA. 

Starting from the X-ray structures of one bacterial and one human DHFR, the GRID/PCA analysis shows that PG 1 distinguishes between the two target proteins, clustering the objects into two groups, while PG 2 ranks the probes. 

In two recent studies, Braiuca et al. applied GRID/PCA to the investigation of substrate selectivity of different forms of penicillin acylase (PA), an important enzyme in the 9-lactam antibiotics industry [12, 17]. Several microbiological sources of PA exist, the enzymes differing in selectivity, activity, or stability. The authors used GRID-MIFs to explain the differences in PA from different sources, E. coli (PA-EC), P. rettgeri (PA-PR), and A. faecalis (PA-AF). GRID/PCA was employed to focus on the important parts in the active site and to reduce the noise in the untreated MIFs. 

An important aspect of this analysis was that the authors decided to build up to four different sub-models for different probe types (e.g. donor, acceptor, hydro-phobic, and halogen probes) to circumvent the known problem of underestimating hydrophobic interactions in the GRID/PCA approach. 

In an earlier GRID/PCA study. Matter and Schwab [4] presented a detailed comparison of MMP3 and MMPS. There, the main selectivity difference was attributed to differences in the ST pocket the identified contour regions were in the vicinity of amino acid differences between the two enzymes. This analysis was supported by parallel CoMFA and CoMSIA analyses, which produced a consistent picture explaining the experimental affinity and selectivity of a series of MMP3 and MMPS inhibitors. 

L. Gardossi, P. Linda, An innovative application of the flexible GRID/PCA computational method Study of differences in selectivity between PGAs from Escherichia coli and a Providentia rettgeri mutant, Biotechnol. Prog. 2004, 20,1025-1031. 

---