|
|
|
|
|
|
| Joost N. Kok, Leiden University |
|
| Joost N. Kok has worked at the Free University in Amsterdam, at the Centre for Mathematics and Computer Science in Amsterdam, at Utrecht University, and at the Abo Akademi University (Finland). Since 1995 he has been a professor in fundamental computer science at Leiden University. Currently, he is the head of the Algorithms and Program Methodology (ALP) group and Director of Computer Science Education. He is also docent at the computer science department of the Abo Akademi University, advisor at the Centre for Mathematics and Computer Science, member of the editorial board of Fundamenta Informaticae and editor of the Series on Natural Computing of Springer Verlag. He is also the Artificial Intelligence theme editor of Encyclopedia Of Life Support Systems, Unesco, associate editor of the Natural Computing Journal of Kluwer, editor of Theoretical Computer Science, section C, and editor of the Journal of Universal Computer Science. The research within the ALP group is concentrated around the topics Bioinformatics, Coordination, Optimization and Data Mining, bridges the gap between the theory section and the applied sections within the Leiden Institute of Advanced Computer Science, and has cooperation with a large number of institutions and companies. Currently, there are 15 PhD students within ALP.
|
|
Finding Discriminative Substructures Using Elaborate Chemical Representation
Joost N. Kok, Leiden Institute of Advanced Computer Science, Leiden University, The Netherlands
Substructure mining algorithms are important drug discovery tools as they can detect substructures that are discriminative of physico-chemical and biological properties. Current methods, however, miss potentially important substructures as they consider only limited chemical information.
We illustrate this point by a case study. The aims of our study were:
- to increase the chemical information considered by computational substructure mining methods;
- to test our graph-based algorithm GASTON designed to find all substructures that satisfy preset criteria from a chemical dataset;
- to automatically extract a compact set of informative, nonredundant and discriminative substructures (e.g., for mutagenicity).
An elaborate representation method was composed that uses atomic hierarchies. In this way, one atom can be represented with a wildcard while additional specific chemical information can be appended as extra nodes. For example, a wildcard label (e.g. [N,O]) can be supplemented with specifiers for atom type (N), number of connected hydrogens (H2), charge (unused), ring size (unused). A mutagenicity dataset of 4069 compounds was pre-processed in this way.
Given this dataset, GASTON computed substructures of any size, any complexity (i.e. linear, tree-shaped and cyclic) and of any degree of chemical detail. In order to obtain all substructures that are directly related to mutagenicity (p < 10-10), each substructure needed to occur in at least 40 mutagens. This resulted in a total collection of over 300,000 substructures.
The most discriminative substructure was determined and compounds that contained this substructure were split from the mutagenicity dataset. This process was repeated until no new substructure could be detected that was strongly associated with mutagenicity (p < 10-10). Each of these split datasets was then searched for detoxifying substructures (at least 80% nonmutagens and p < 10-10).
(Joint work with Jeroen Kazius, Siegfried Nijssen, Thomas Bäck, Ad P. IJzerman)
|
|
|
|
|
|
|
|
|
