2.1. Data
Drosophila melanogaster (fruit fly) data have been employed in our integrative analysis, with several types of data retrieved from publicly available databases. These include time series data from two platforms (retrieved from the Gene Expression Omnibus (GEO) database [31]), a set of knock-out (KO) microarray experiments, position-specific weight matrices (PSWMs; [32]), known cis-regulatory modules and Gene Ontology (GO) annotations. For model validation, a set of previously known transcriptional interactions extracted from the DROID database (Drosophila Interactions Database; [33]) has been used. A subnetwork of 27 genes involved in embryo development, listed in Table 1, has been analysed. These have been chosen by starting from the main genes involved in segmentation [17], followed by the addition of several genes known to interact with this main set (based on the FlyBase interactions browser [34]).
microarrays-04-00255-t001_Table 1 Table 1  Set of 27 genes selected for network analysis for the Drosophila melanogaster dataset.   Dual-channel (DC) dataset. This time-course dataset analyses gene expression during Fly embryo development, using dual-channel microarrays (GEO Accession GSE14086 [35]). The dataset contains seven time points sampled at 1- and 2-h intervals, up to 10 h after egg laying. Three biological replicates are available, resulting in three time series in total.
Single-channel (SC) dataset. The single-channel dataset [36], measured with Affymetrix arrays, contains gene expression measurements for 12 time points during Drosophila melanogaster embryo development. Samples have been taken every hour up to 12 and a half hours after egg laying. Three biological replicates are included. Both the SC and DC datasets were normalised using cross-platform normalisation [37], which was shown previously to be a valid option for time series data integration [29].
Previously known transcriptional interactions (DROID dataset). For validation purposes, a set of known transcriptional interactions has been retrieved from DROID (Drosophila Interactions Database; [33]), Version 2010_10. This consists of 16 pair-wise interactions between transcription factors and their target genes, for the 27-gene network under analysis. This gold standard was used due to the fact that these interactions are confirmed experimentally. Although it is not the complete network for the 27 genes and it does not include PPIs, it does help to indicate the quality of our models in terms of underlining transcriptional regulation, which is the interest of this study. The exact interactions are included as Supplementary Material.
KO dataset. Five KO microarray datasets have been retrieved form the GEO database. These contain knock-out experiments for 8 genes and the corresponding wild-type measurements. The accession numbers for the datasets are GSE23346 ([38], Affymetrix Drosophila Genome 2.0 Array, 6 samples), GSE9889 ([39], Affymetrix Drosophila Genome Array, 20 samples), GSE7772 ([40], Affymetrix Drosophila Genome Array, 4 samples), GSE3854 ([41], Affymetrix Drosophila Genome Array, 54 samples) and GSE14086 ([35], dual-channel array, 63 samples). For these, the log-ratios between knock-out and wild-type expression values have been used within our framework.
Binding site affinities (BSAs). A set of PSWMs for 11 transcription factors have been retrieved from [42]. These matrices have been computed using DNA foot-printing data from [43]. In order to compute BSAs using PSWMs, the promoter sequence for each gene is required. For the Drosophila genome, the RedFly database [44] provides a set of known cis-regulatory modules, which have been used here for this purpose. Cis-regulatory modules for 16 genes have been retrieved, while for the other genes, the upstream 2-Kbp sequence has been used to assess BSA. Using both information types, BSAs were computed for use in our algorithm.
GO annotations. GO [45] is a database of genes, which have been annotated to have a specific function or to be involved in specific processes. These annotations come from various sources and have been determined using technologies ranging from those in wet-lab experiments to computational methods. The database is a valuable source of meta-information that can be used in different ways. Here, we have used the GO platform to identify which of the gene products involved in the network analysed have been previously shown to display transcription factor activity.
Correlations (CORR). All gene expression data related to the genes of interest were combined, and the Pearson product moment correlation coefficient was computed between gene pairs. These data were fed into the EGIA framework.