Predicting Compound Selectivity by Self‐Organizing Maps: Cross‐Activities of Metabotropic Glutamate Receptor Antagonists

The self-organizing map (SOM) principle was introduced by Kohonen in 1982, and has been applied to a variety of tasks in chemistry and chemical biology ever since. In this study, we used the SOM algorithm for mapping known ligands according to a topological pharmacophore descriptor (CATS) and predicting potential cross-activities. Our aim was to see whether 1) the descriptor is able to discriminate antagonists of metabotropic glutamate receptors (mGluR) 1 and 5, and 2) the SOM could be used for predicting potential additional binding behaviors of the ligands. First, an mGluR reference collection containing 338 compounds was compiled including published and Merz in-house structures of noncompetitive group I mGluR antagonists. The collection comprises two subsets: allosteric mGlu1 receptor antagonists (213 compounds), and allosteric mGlu5 receptor antagonists (125 compounds). These molecules cover a broad range of binding activities (Ki values between 1 nm and 10 mm) and represent different chemical classes. This mGluR reference library was complemented by the molecules from the COBRA database (v. 3.12; 5376 molecules) containing a broad set of known drugs, leads, and lead candidates affecting a large number of different drug targets. Subsequently, the molecules were converted to a vector representation giving the scaled occurrence frequencies of topological potential pharmacophore point pairs (CATS2D method). 7] In this study, intramolecular distances from zero to nine bonds were considered, resulting in a 150-dimensional vector representation of each molecular compound. The complete COBRA database was subjected to clustering and mapping onto a two-dimensional grid by the SOM approach. The SOM provides a nonlinear two-dimensional projection of the 150-dimensional data space (“chemical space”), where local neighborhood is conserved. This means, that molecules that are located close to each other on the map are also close in the original high-dimensional space. For SOM training we applied a slightly modified version of the Kohonen algorithm as described previously. 8] As a result, all molecules from COBRA were distributed into 225 (15A15) clusters ACHTUNGTRENNUNG(“neurons”