Conclusion and outlook

We were able to demonstrate that virtual affinity fingerprints can be used for virtual screening of molecular databases in order to classify compounds by their activity classes.

In this context, we believe that a careful selection and optimization of the reference panel of either protein binding pockets (DOCKSIM and Flexsim-X) or small molecules (Flexsim-S) is an important step to improve the predictive performance. In addition, scoring functions and treatment of the ligands (rigid vs flexible) might play an important role in this respect. As improved scoring functions and enhanced docking algorithms become available, we will check their influence on our results.

By comparing the hit lists obtained by either virtual affinity fingerprints or some popular 2D descriptors we showed that there is high redundancy amongst the latter approaches whereas the 3D fingerprints can be regarded as complementary to both the 2D methods as well as amongst themselves. Consequently, for a general database search for compounds similar to a given lead structure, a mix taking into account the results of all methods would be highly desirable.

Therefore we are particularly interested in data fusion techniques as proposed by Willett's group [26] to merge search results from different descriptor types.

Search Query

Search Query

DOCKSIM Hit

FlBLSlm-X Hit

Flaxaim-S Hit

o. 0. kAOH

Figure 3. Example of hits found with DOCKSIM, Flexsim-X and Flexsim-S. All the hits are members of the respective 10 nearest neighbor lists of the query structure. None of them, however, are found amongst the nearest neighbor hit lists of any ofthe 2D descriptor methods. All compounds shown are described as PAF antagonists in the MDDR database.

Figure 3. Example of hits found with DOCKSIM, Flexsim-X and Flexsim-S. All the hits are members of the respective 10 nearest neighbor lists of the query structure. None of them, however, are found amongst the nearest neighbor hit lists of any ofthe 2D descriptor methods. All compounds shown are described as PAF antagonists in the MDDR database.

Most important, affinity fingerprints are capable to reflect biological similarities of molecules beyond their structural classes. This can be highly desirable, e.g. to search a corporate database for compounds similar to an early screening hit or even to a competitor's compound.

In order to illustrate this 'island-hopping' situation, three examples are given in Figure 3 (one for each of our virtual affinity methods). All of the 'hits' are found amongst the top 10 nearest neighbors of the query molecule and are correctly classified as PAF antagonists. On the other hand, none of these compounds is part of the top 10 hit lists determined by any of the 2D finger print methods.

0 0

Post a comment