Figure 3

The graphic algorithm of system procedures. The whole system is implemented by Matlab 6.5. Notice that the RBF classifier is not designed for predicting novel TFs for yeast or human species. The main purpose of the RBF classifier is to group the kinds of categories that TFs belong to. As Figure shows, we input TFs into the system and the trained RBF classifier then can decide on the category according to the microarray expression values and transcription factor binding sites sequences of the TFs. This step also indicates that what kinds of RNN architecture are used in the following steps. The GA uses a standard random mutation and a standard binary representation with one point crossover. One TF may not only regulate one target gene, but may regulate several target genes simultaneously. To acquire the "good" combinations of target genes that are regulated by one or more common TFs, we select appropriate GA mutation and crossover operators to alter the chromosomes. One chromosome of the GA represents a number of genes taken from the full set of genes and is used by RNN to check how "good" the expression values of this combination of genes affected by particular transcription regulators are. On the other hand, the GA consists of populations of such chromosomes, and each chromosome is evaluated by the RNN for its fitted value to the given TFs. The choice of RNN architecture is according to the labels assigned by the RBF classifier. Furthermore, the final returned RNN output error (RMSE) is treated as a fitted value for some particular combinations of target genes. The stopping criterion includes not only the fitted value fit for some criteria but also the determination of the RNN selecting steps. In other words, the GA never stops until all appropriate RNN architectures are executed. In that case, each TF can choose suitable RNN architecture more than once, and find out a dissimilar set of target genes. After all TFs are run by this system procedures and output regulatory modules, the GA is then complete.