In silico or virtual affinity fingerprints

Encouraged by Terrapin's and NCI's success using in vitro affinity fingerprints, we were the first who developed and reported virtual affinity fingerprints [5].

The in vitro assays were replaced by computational docking of the ligands into binding pockets of protein structures solved by X-ray crystallography. Consequently, a molecule's virtual affinity fingerprint is defined as a vector of its calculated docking scores with respect to the reference panel. Our first approach - subsequently termed DOCKSIM - employed the docking program DOCK (Version 3.5) [12] which at that time only allowed rigid docking of the ligand test set. A reference panel of eight uncorrelated protein structures was selected from the Brookhaven Protein Data Bank [13]. In contrast to Terrapin's work, we did not aim at a quantitative affinity prediction but rather used the pairwise euclidean distances of the affinity fingerprints as a similarity measure in order to classify compounds correctly (see Reference 5 and Methods section). Recently, we described a further development of the DOCKSIM approach, called Flexsim-X [14]. The main enhancements compared to DOCKSIM can be summarized as follows:

- Instead of rigid docking we were able to apply flexible docking using the program FlexX (Version 1.65) [15].

- The composition and size of the reference panel of protein binding pockets were optimized by systematic as well as genetic algorithm-based (GA) procedures. This resulted in a remarkable performance increase compared to arbitrary selection of the panel.

Finally, Ghuloum et al. [16] from MetaXen proposed an approach called molecular hashkeys, which uses surface-based comparisons of target molecules with a reference panel comprising small, drug-like molecules instead of proteins. Despite this difference compared to all the other methods described so far, the molecular hashkey approach nevertheless clearly lies at the heart of the original affinity fingerprint idea.

Ghuloum et al. used their program to predict two molecular properties:

Tablel. Characteristics of DOCKSIM, Flexsim-X and Flexsim-S

DOCKSIM Flexsim-X Flexsim-S

Underlying program

Reference panel composition

Reference panel selection

Ligand treatment

8 binding pockets from PDB

Arbitrarily Rigid docking

41 binding pockets from PDB

GA-based optimization out of 100 pockets

Flexible docking

44 small molecules from MDDR

GA-based optimization out of 100 molecules

Rigid fitting

By applying a weighted K nearest neighbor (KNN) classification, they were able to predict the octanol-water partition coefficient (logP) of a set of almost 1000 compounds by an accuracy similar to the ClogP program [17]. In addition, they successfully trained a neural network in order to get a predictive model for intestinal absorption.

Interestingly, by simply adding more members to the reference panel they observed a plateau, i.e. no further increase in predictiveness, similar to our experience described in our Flexsim-X paper [14]. This supports our hypothesis that on one hand a careful selection and optimization of the panel is crucial for the success of the affinity fingerprint methods, but on the other hand a kind of natural limit seems to exist for the panel size.

Independently from MetaXen's work, we have recently developed a similar approach which we termed Flexsim-S. It involves superimposing ligand molecules onto a set of small reference molecules employing the FlexS [ 18] program (Version 1.32). As in our docking approaches DOCKSIM and Flex-sim-X, a virtual affinity fingerprint is constructed as a linear vector of scores. In Flexsim-S, however, the docking scores are substituted by measuring the alignment quality of two small molecules. During development of Flexsim-X we learned that an optimization both in size and composition of the reference panel is crucial in order to improve the classification power of the virtual affinity fingerprints. Consequently, we used the same GA-based optimization procedure [19] as described for Flexsim-X.

Table 1 summarizes some characteristic features of our three different virtual affinity fingerprint approaches DOCKSIM, Flexsim-X and Flexsim-S.

A detailed description of Flexsim-S will be given in the Methods section below.

0 0

Post a comment