ar causes sellckchem of the positive assay readout. Many such methods rely on the availability of compound annotations from previous e periments or specific profiling platforms. There is a considerable amount of literature on target prediction methods that work from chemical structure Inhibitors,Modulators,Libraries alone or composite data types using a variety of meth ods, and we refer the interested reader to and the references therein. Profiling platforms are composed of a reference base of n dimensional readouts, for e ample a panel of reporter gene assays, for a set of well characterised compounds and a mechanism to position the readouts of novel samples in the conte t of the reference. This latter mechanism is often some kind of metric such as Euclidean distance or Pearson correla tion, though more sophisticated methods can also be applied.

Transcriptional profiles, the mRNA levels of e pressed genes as a result of treatment of cells with a compound, are routinely used to cluster or otherwise relate com pounds that elicit a similar biological response. For any such approach, it is important to choose which genes to include in the calculations. Typical human gen ome wide chips cover appro imately Inhibitors,Modulators,Libraries 22,000 genes, where the e pression level of each gene is determined by a set of specific probes, a probeset. Other e peri mental techniques, however, require the selection of a set of genes upfront, for e ample the Lumine technol ogy of Panomics. The selection of suitable genes, a gene signature, depends on the desired signature size, which is directly proportional to cost, as well as the bio logical questions that need to be addressed.

The selec tion and evaluation of such gene signatures is the subject of the remainder of this article. Like many other companies, Novartis has several compound profiling platforms, including one based on e pression profiles. The questions that we addressed in this article are directly related to some of our ongoing efforts to opti mise such platforms. We used a publicly available microarray Inhibitors,Modulators,Libraries dataset in conjunction with e tensive compound annotations to probe several important aspects of target and MoA pre diction from gene signatures. We e plored systemati cally to what e tent transcriptional profiles of compounds can be used for target prediction.

This study provided insight into questions such as the follow ing Is there and what is the minimal gene signature that can be used to reasonably predict molecular targets of compounds Do designed signatures predict targets bet ter than genes selected at random How can such signa tures be optimised in an automatic way, and what are the results of such an optimisation We employed machine Inhibitors,Modulators,Libraries learning and biologically inspired algorithms implemented on state of the art graphics processing units to answer these questions. Results and discussion Compound target annotations We retrieved all currently known targets for any com pound in Connectivity Carfilzomib Map 2 where the compound had an activity of 5 uM. Each compound selleckchem Imatinib Mesylate had an average