Mining High-Throughput Screening Data of Combinatorial Libraries:  Development of a Filter to Distinguish Hits from Nonhits

Kohonen neural networks generate projections of large data sets defined in high-dimensional space. The resulting self-organizing maps can be used in many applications in the drug discovery process, such as to analyze combinatorial libraries for their similarity or diversity and to select descriptors for structure−activity relationships. The ability to investigate thousands of compounds in parallel also allows one to conduct a study based on single-dose experiments of high-throughput screening campaigns, which are known to have a greater uncertainty than IC50 or K i values. This is demonstrated here for a data set of 5513 compounds from one combinatorial library. Furthermore, a method was developed that uses self-organizing maps not only as an indicator of structure−activity relationships, but as the basis of a classification system allowing predictive modeling of combinatorial libraries.