The 3T3-L1 cell line treated with a differentiation cocktail was used as a model to study gene expression profiles during adipogenesis. Three independent time series differentiation experiments were performed. RNA was isolated at the preconfluent stage (reference) and at eight time points after confluence (0, 6, 12, 24, 48 and 72 hours, and 7 and 14 days). Gene expression levels relative to the preconfluent state were determined using custom-designed microarrays with spotted polymerase chain reaction (PCR) products. The microarray developed here contains 27,648 spots with mouse cDNA clones representing 16,016 different genes (UniGene clusters). These include 15,000 developmental clones (the NIA cDNA clone set from the US National Institute of Aging of the National Institutes of Health NIH), 11,000 clones from different brain regions in the mouse (Brain Molecular Anatomy Project [BMAP]), and 627 clones for genes which were selected using the TIGR Mouse Gene Index, Build 5.0 19.

All hybridizations were repeated with reversed dye assignment. The data were filtered, normalized, and averaged over biological replicates. Data processing and normalization are described in detail under Materials and methods (see below). Signals at all time points could be detected from 14,368 elements. From these microarray data, we identified 5205 ESTs that exhibited significant differential expression between time points and had a complete profile (P < 0.05, one-way analysis of variance [ANOVA]). Because ANOVA filters out ESTs with flat expression profiles, we used a fold change criteria to select the ESTs for further analysis. We focused on 780 ESTs that had a complete profile over all time points, and that were more than twofold upregulated or downregulated in at least four of those time points. These stringent criteria were necessary to select a subset of the ESTs for in-depth sequence analysis and for examination of the dynamics of the molecular processes. The overlap between the ANOVA and twofold filtered ESTs was 414. All of the data, together with annotations and other files used in the analyses, are available as Additional data files and on our website 20. The analyses described in the following text were conducted in the set of 780 ESTs.

Four lines of evidence support the quality of our data and its consistency with existing knowledge of fat cell biology. First, our array data are consistent with reverse transcriptase (RT)-PCR analysis. We compared the microarray data with quantitative RT-PCR for six different genes (Pparg [number 592, cluster 6], Lpl [number 14, cluster 6], Myc [number 224, cluster 11], Dcn [number 137, cluster 7], Ccna2 [number 26, cluster 5/8], and Klf9 [number 6, cluster 9]) at different time points (Additional data file 9 and on our website 20). A high degree of correlation was found (r² = 0.87), confirming the validity of the microarray data.

Second, statistical analyses of the independent experiments showed that the reproducibility of the generated data is very high. The Pearson correlation coefficient between the replicates was between 0.73 and 0.97 at different time points. The mean coefficient of variation across all genes at each time point was between 0.11 and 0.27. The row data and the details of the statistical analyses can be found in Additional data file 10 and on our website 20.

Third, comparison between our data and the Gene Atlas V2 mouse data for adipose tissue 21 shows that the consistency of the two data sets increases with differentiation state (Additional data files 11 and 12, and our on website 20). Therefore, this analysis supports the relevance of the chosen cell model to in vivo adipogenesis. Among the 382 transcriptionally modulated genes common in both data sets, 67% are regulated in the same direction at time point zero (confluent pre-adipocyte cell culture). At the final stage of differentiation, the correlation increases up to 72%. If the Gene Atlas expression data are restricted to strongly regulated genes (at least twofold and fourfold change respectively), then the consistency in mature adipocytes rises to 82% (135 genes) and 93% (42 genes), respectively. Out of all 60 tissues in the Gene Atlas V2 mouse, the adipose tissue describes the differentiated state of the 3T3-L1 cells best. Brown fat tissue is the second best match to the differentiated adipocytes (69% of the 382 genes), followed by adrenal gland (66%), kidney (65%), and heart (64%). At each time point in which cell cycle genes were not repressed (12 hours and 24 hours), all tissues had similar correlation to the data set (44-55% for 382 genes).

The fourth line of evidence supporting the quality of our data is that there is clear correspondence between our data and a previously published data set 8. For a group of 153 genes shared among the two studies, the same upregulation or downregulation was found for 72-89% (depending on time point) of all genes (see Additional data file 13 and our website 20). The highest identity (89%) was found for the stage terminally differentiated 3T3-L1 cells, for which the profile is less dependent on the precise extraction time. If the comparison is restricted to expression values that are strongly regulated in both experiments (at least twofold change at day 14, 96 ESTs), then the coincidence at every time point is greater than 90%. Comparisons with this 8 and two additional data sets 912, and the data pre-processing steps are given in Additional data files 13, 14, 15 and on our website 20. Note that, because of the differences in the used microarray platforms, availability of the data, normalization methods, and annotations, only 96 genes are shared between all four studies. Of the 780 ESTs monitored in our work, 326 were not detected in the previous studies 8912: 106 RefSeqs (with prefix NM), 43 automatically generated RefSeqs (with prefix XM), and 173 ESTs (Additional data file 16).

To examine the relationships between coexpression and gene functions, we first clustered 780 ESTs that were twofold differentially expressed into 12 temporally distinct patterns, containing between 23 and 143 ESTs (Figure 1). ESTs in four of the clusters are mostly upregulated during adipogenesis, whereas genes in the other eight clusters are mostly downregulated.

We then categorized ESTs with available RefSeq annotation and Gene Ontology (GO) term (486 out of 780) for molecular function, cellular component, and biological process (Figure 2). Genes in clusters 5 and 8 are downregulated through the whole differentiation process and upregulated at 12/24 hours. Many of the proteins encoded by these genes are involved in cell cycle processes and were residing in the nucleus (Figure 2). Re-entry into the cell cycle of growth arrested pre-adipocytes is known as the clonal expansion phase and considered to be a prerequisite for terminal differentiation in 3T3-L1 adipocytes 22. Genes grouped in cluster 2 are highly expressed from 6 hours (onset of clonal expansion) to 3 days (start of the appearance of adipocyte morphology) but are only modestly expressed at the terminal adipocyte differentiation stage. These include a number of genes that encode signaling molecules. Genes increasingly expressed toward the terminal differentiation stage are in clusters 4, 6, and 7, although from different starting values. Some genes in cluster 6 are known players in lipid metabolism and mitochondrial fatty acid metabolism, whereas some genes can be associated with cholesterol biosynthesis and related to extracellular space or matrix in clusters 4 and 7, respectively.

Previous studies suggest that protein production for 10% or more of all human and mouse genes are regulated by microRNAs (miRNAs) 2324. miRNAs are short, noncoding, single-strand RNA species that are found in a wide variety of organisms. miRNAs cause the translational repression or cleavage of target messages 25. Some miRNAs may behave like small interfering RNAs. It appears that the extent of base pairing between the small RNA and the mRNA determines the balance between cleavage and degradation 26. Rules for matches between miRNA and target messages have been deduced from a range of experiments 24 and applied to the prediction and discovery of mammalian miRNA targets 2327. Moreover, it was shown that human miRNA-143 is involved in adipocyte differentiation 28.

Here we conducted an analysis to determine which of the 780 ESTs differentially expressed during adipocyte differentiation were potential targets of miRNAs and whether there is an over-representation of miRNA targets of coexpressed ESTs clustered in 12 distinct expression patterns. From the 780 ESTs, the 3'-untranslated region (UTR) could be derived for 539. Of these, 518 had at least one exact antisense match for the seven-nucleotide miRNA seed (base 2-8 at the 5' end) from the 234 miRNA sequences (18-24 base pairs [bp]; Additional data file 14). From 395 ESTs with a unique 3'-UTR, 282 (71%) had at least one match over-represented compared with the whole 3'-UTR sequence set (21,396; P < 0.05, by one-sided Fisher's exact test). The distribution of statistically over-represented miRNA motifs in 3'-UTRs across the clusters was variable, with genes grouped in cluster 9 (including many transcriptional regulators) having the most statistically over-represented miRNA motifs and genes in cluster 5 having no detectable motifs (Additional data file 18). The results of the analysis of cluster 9 are given in Figure 3. One of the genes with the most significantly over-represented miRNA motifs in the 3'-UTR is related to the ras family (Figure 3). It was previously shown that human oncogene RAS is regulated by let-7 miRNA 29. Further potential miRNA target genes from all clusters are given in Additional data files 9, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30.

In order to functionally characterize the molecular components underlying adipogenesis in detail, comprehensive bioinformatics analyses of 780 differentially expressed ESTs were performed. A total of 659 protein sequences could be derived, and these were subjected to in-depth sequence analytic procedures. The protein sequences have been annotated de novo using 40 academic prediction tools integrated in the ANNOTATOR sequence analysis system. The structure and function was annotated on a sequence segment/domain-wise basis. After extensive literature search and curation using the sequence architecture, 345 gene products were mapped onto known pathways, possible cellular roles, and subcellular localizations (Figure 4) using the PathwayExplorer web service 30 as well as manual literature and domain-based assignment. The results of the sequence analyses and additional information is available in the supplementary material available on our website 20 and Additional data files 6, 7, 8.

This molecular atlas of fat cell development provides the first global view of the underlying biomolecular networks and represents a unique resource for deriving testable hypotheses for future studies on individual genes. Below we demonstrate the usefulness of the atlas by highlighting the following: established regulators of fat cell development, recently discovered fat cell gene products, and candidate transcription factors expressed during adipogenesis. The numbering of the genes is given according to the de novo functional annotation (Additional data file 7).

In addition to the metabolic networks, the molecular atlas also provides a bird's eye view of other molecular processes, including signaling, the cell cycle, remodeling of the extracellular matrix, and cytoskeletal changes. Changes that occur during adipogenesis (phenotypically seen as rounding of densely packed cells) have aspects in common with other tissue differentiation processes such as endothelial angiogenesis (protease, collagen, and noncollagen molecule secretion) 56 and specific features. Here we show that phenotypic changes that occur in maturing adipocytes are paralleled by expression of the respective genes.

We next used the molecular atlas to derive novel biological insights from the global view of molecular processes. We analyzed transcriptionally regulated genes that are members of 36 different metabolic pathways. Within each pathway, we considered whether these transcriptionally regulated genes occupy key positions, such as a position at the pathway start, which is the typical rate-limiting step where the amount of enzyme is critical 76, or at some other point of regulation. We found that such key positions are occupied by transcriptionally regulated targets in 27 pathways (an overview is provided in Table 1). Those pathways that are strongly transcriptionally regulated at key points are illustrated in Figure 5 at the time points 0, 24 and 48 hours, and 14 days. For additional time points and images with more detailed information from all investigated pathways, see our website 20 and Additional data files Additional data file 31 and Additional data file 32.

In the following discussion we present the evidence for transcriptional regulation at key points for five selected metabolic pathways. Further information on other pathways can be found in Additional data files 31, 32, 33, 34, 35, 36, 37, 38 and on our website 20.

To determine whether coexpressed genes are also coregulated, we analyzed the available promoter sequences of the 780 ESTs. Promoter sequences could be retrieved for 357 genes. Most ESTs are sequenced from the 3' end, and hence it is easier to retrieve the 3'-UTR. Retrieval of promoters is more difficult than retrieval of the 3'-UTR because of experimental problems in extracting full-length cDNAs (and hence transcription start sites) and insufficient computational methods for identifying beginning of the 5'-UTR. We analyzed the occurrences of the binding sites of all transcription factors in vertebrates from the TRANSFAC database. Based on statistical analyses, among transcription factors with binding site motifs described in TRANSFAC 100 those listed in Table 2 are the most promising candidates for further functional studies on transcriptional regulation.

One example of a functional transcription factor binding site is SREBP-1 in cluster 4. A comparison among clusters showed that cluster 4 has significantly more genes with a SREBP-1 (SRE and E-box motifs 101) binding site than all other clusters (P = 0.0484, Fisher's exact test; Table 3). Similarly, a putative SREBP-1 regulatory region is significantly more frequent in the promoters of the genes in cluster 4 compared with all unique sequences in the PromoSer database (P < 0.0289; PromoSer contains 22,549 promoters of 12,493 unique sequences). For a subset of the genes in cluster 4 with predicted SREBP-1 binding site (most genes of the cholesterol biosynthesis pathway), transcriptional regulation with SREBP-1 has been experimentally proven 102.

Surprisingly, binding sites for the key regulators of adipogenesis, namely PPAR and C/EBP, are not significantly over-represented in any of the promoters of the coexpressed genes. We generated a novel matrix for PPAR using 22 experimentally verified binding sites from the literature and analyzed the promoters of the coexpressed genes and all PromoSer promoters. Again, using this matrix the PPAR binding sites were not significantly over-represented.

Finally, we considered whether coexpressed genes also colocalize on the chromosomes. In a broad genomic interval (5 megabases [Mb]) on each mouse chromosome we mapped the ESTs from each cluster. Unexpectedly, our data do not support the observation of the highly significant correlation in the expression and genomic positioning of the genes. A typical example of mapped ESTs to chromosome 10 is illustrated in Figure 6, showing that expression levels of colocalized ESTs are divergent because only two mapped ESTs are members of the same cluster.

Additionally, we analyzed the genomic position of 5,205 ESTs that exhibited significant differentially expression between time points (P < 0.05; one-way ANOVA). These ESTs were grouped in 12 clusters, and we then searched for regions with three or more members in a genomic interval of 500 kilobases (kb). On average, 7 ± 5% of the ESTs from one cluster were colocalized. Comprehensive results of this analysis are accessible within the supplementary website 20 and Additional data files 42, 43, 44, 45.

In summary, these data do not provide evidence that colocalized genes in the genomic sequence are subject to the same transcriptional regulation (coexpression), as indicated by examples for different processes in other studies 103.

The microarray developed here contains 27,648 spots with mouse cDNA clones representing 16,016 different genes (UniGene clusters). These include developmental clones (the 15 K NIA cDNA clone set from National Institute of Aging, US National Institutes of Health) and the 11 K clones from different brain regions in the mouse (Brain Molecular Anatomy Project [BMAP]). Moreover, 627 clones for adipose-related genes were selected using the TIGR Mouse Gene Index Build 5.0 19. These cDNA clones were obtained from the IMAGE consortium (Research Genetics, Huntsville, AL, USA). The inserts of the NIA and BMAP clones were sequence verified (insert size about 1-1.5 kb). All PCR products were purified using size exclusion vacuum filter plates (Millipore, Billerica, MA, USA) and spotted onto amino-silanated glass slides (UltraGAPS II; Corning, Corning, NY, USA) in a 4 × 12 print tip group pattern. As spotting buffer 50% dimethyl sulfoxide was used. Negative controls (genomic DNA, genes from Arabidopsis thaliana, and dimethyl sulfoxide) and positive controls (Cot1-DNA and salmon sperm DNA) were included in each of the 48 blocks. Samples were bound to the slides by ultraviolet cross-linking at 200 mJ in a Stratalinker (Stratagene, La Jolla, CA, USA).

3T3-L1 cells (American Type Culture Collection number CL-173) were grown in 100 mm diameter dishes in Dulbecco's modified Eagle's medium supplemented with 10% fetal bovine serum, 100 units/ml penicillin, 100 μg/ml streptomycin, and 2 mmol/l L-glutamine in an atmosphere of 5% carbon dioxide at 37°C. Two days after reaching confluence (day 0), cells were induced to differentiate with a two-day incubation of a hormone cocktail 109110 (100 μmol/l 3-iso-butyl-1-methylxanthine, 0.25 μmol/l dexamethasone, 1 μg/ml insulin, 0.16 μmol/l pantothenic acid, and 3.2 μmol/l biotin) added to the standard medium described above. After 48 hours (day 2), cells were cultured in the standard medium in the presence of 1 μg/ml insulin, 0.16 μmol/l pantothenic acid, and 3.2 μmol/l biotin until day 14. Nutrition media were changed every second day.

Three independent cell culture experiments were performed. Cells were harvested and total RNA was isolated at the preconfluent stage and at eight time points (0, 6, 12 and 24 hours, and 3, 4, 7 and 14 days) with TRIzol reagent (Invitrogen-Life Technologies; Carlsbad, CA, USA) 111. For each independent experiment, RNA was pooled from three different culture dishes for each time point and from 24 dishes at the preconfluent stage used as reference. The quality of the RNA was checked using Agilent 2100 Bioanalyzer RNA assays (Agilent Technologies, Palo Alto, CA, USA) by inspection of the 28S and 18S ribosomal RNA intensity peaks.

The labeling and hybridization procedures used were based on those developed at the Institute for Genomic Research 112 and detailed protocols can be viewed on the supplementary website 20. Briefly, 20 μg total RNA from each time point was reverse transcribed in cDNA and indirectly labeled with Cy5 and 20 μg RNA from the preconfluent stage (reference) was indirectly labeled with Cy3, respectively. This procedure was repeated with reversed dye assignment. Slides were prehybridized with 1% bovine serum albumen. Then, 10 μg mouse Cot1 DNA and 10 μg poly(A) DNA was added to the labeled cDNA samples and pair-wise cohybridized onto the slide for 20 hours at 42°C. Following washing, slides were scanned with a GenePix 4000B microarray scanner (Axon Instruments, Sunnyvale, CA, USA) at 10 μm resolution. Identical photo multiplier voltage settings were used in the scanning of the corresponding dye-swapped hybridized slides. The resulting TIFF images were analyzed with GenePix Pro 4.1 software (Axon Instruments).

Data were filtered for low intensity, inhomogeneity, and saturated spots. To obtain expression values for the saturated spots, slides were scanned a second time with lower photomultiplier tube settings and reanalyzed. All spots of both channels were background corrected (by subtraction of the local background). Different sources of systematic (sample, array, dye, and gene effects) and random errors can be associated with microarray experiments 113. Nonbiological variation must be removed from the measurement values and the random error can be minimized by normalization 114115. In the present study, gene-wise dye swap normalization was applied. Genes exhibiting substantial differences in intensity ratios between technical replicates were excluded from further analysis based on a two standard deviation cutoff. The resulting ratios were log₂ transformed and averaged over three independent experiments. The expression profiles were not rescaled in order to identify genes with high expression values. All experimental parameters, images, and raw and transformed data were uploaded to the microarray database MARS 116 and submitted via MAGE-ML export to a public repository (ArrayExpress 117, accession numbers A-MARS-1 and E-MARS-2). Differentially expressed genes were first identified using one-way ANOVA (P < 0.05). They were then subjected to a more stringent criterion; specifically, we considered only those genes with a complete temporal profile that were more than twofold upregulated or downregulated at a minimum of four time points. The twofold cutoff for differentially expressed genes was estimated by applying the significance analyses of microarrays method 118 to the biological replicates and assuming false discovery rate of 5%. In order to capture the dynamics of various processes, only ESTs differentially expressed in at least half of the time points were selected. Data preprocessing was performed with ArrayNorm 119.

Microarray expression results were confirmed with RT-PCR. cDNA was synthesized from 2.5 μg total RNA in 20 μl using random hexamers and SuperScript III reverse transcriptase (Invitrogen, Carlsbad, CA, USA). The design of LUX™ primers for Pparg, Lpl, Myc, Dec, Ccna2, and Klf9 was done using the Invitrogen web service (for sequences, see Additional data file 9 and our website 20). Quantitative RT-PCR analyses for these genes were performed starting with 50 ng reverse transcribed total RNA, with 0.5× Platinum Quantitative PCR SuperMix-UDG (Invitrogen, Carlsbad, CA, USA), with a ROX reference dye, and with a 200 nmol/l concentration of both LUX™ labeled sense and antisense primers (Invitrogen, Carlsbad, CA, USA) in a 25 μl reaction on an ABI PRISM 7000 sequence detection system (Applied Biosystems, Foster City, CA, USA). To measure PCR efficiency, serial dilutions of reverse transcribed RNA (0.24 pg to 23.8 ng) were amplified. Ribosomal 18S RNA amplifications were used to account for variability in the initial quantities of cDNA. The relative quantification for any given gene with respect to the calibrator (preconfluent stage) was determined using the ΔΔC_t method and compared with the normalized ratios resulting from microarray experiments.

Common unsupervised clustering algorithms 120 were used for clustering expression profiling of 780 selected ESTs, according to the log ratios from all time points. Using hierarchical clustering the boundaries of the clusters were not clearly separable and required arbitrary determination of the branching point of the tree, whereas the results of the clustering using self-organized maps led to clusters with highly divergent number of ESTs (between 3 and 242). We have therefore used the k means algorithm 121 and Euclidean distance. The number of clusters was varied from k = 1 to k = 20, and predictive power was analyzed with the figure of merit 122. Subsequently, k = 12 was found to be optimal. To evaluate the results of the k means clustering, principal component analysis 123 was applied and exhibited low intracluster distances and high intercluster dissimilarities. GO terms and GO numbers for molecular function, biological process, and cellular components were derived from the Gene Ontology database (Gene Ontology Consortium) using the GenPept/RefSeq accession numbers for annotated proteins encoded by selected genes (ESTs). All cluster analyses and visualizations were performed using Genesis 124.

For each of the 780 selected EST sequences, we attempted to find the corresponding protein sequence. Megablast 125 searches (word length w = 70, percentage identity p = 95%) against nucleotide databases (in the succession of RefSeq 126127, FANTOM 128, UniGene 129, nr GenBank, and TIGR Mouse Gene Index 19 until a gene hit was found) were carried out. For the ESTs still remaining without gene assignment, new Megablast searches were conducted with the largest compilation of RefSeq (including the provisional and automatically generated records 126127). If an EST remained unassigned, then the whole procedure was repeated with blastn 130. In addition, a blastn search against the ENSEMBL mouse genome 131 was performed, and ESTs with long stretches (>100 base pairs) of unspecified nucleotides (N) were excluded.

All protein sequences were annotated de novo with academic prediction tools that are integrated into ANNOTATOR, a novel protein sequence analysis system 132: compositional bias (SAPS 133, Xnu, Cast 134, GlobPlot 1.2 135); low complexity regions (SEG 136); known sequence domains (Pfam 137, Smart 138, Prosite and Prosite pattern 139 with HMMER, RPS-BLAST 140, IMPALA 141, PROSITE-Profile 139); transmembrane domains (HMMTOP 2.0 142, TOPPRED 143, DAS-TMfilter 144, SAPS 133); secondary structures (impCOIL 145, Predator 146, SSCP 147148); targeting signals (SIGCLEAVE 149, SignalP-3.0 150, PTS1 151); post-translational modifications (big-PI 152, NMT 153, Prenylation); a series of small sequence motifs (ELM, Prosite patterns 139, BioMotif-IMPlibrary); and homology searches with NCBI blast 130. Further information was retrieved from the databases of Mouse Genome Informatics 154 and LocusLink 126.

The promoters were retrieved from PromoSer database 155 through the gene accession number. PromoSer contains 22,549 promoters for 12,493 unique genes. Nucleotides from 2,000 upstream and 100 downstream of the transcription start site were obtained. With an implementation of the MatInspector algorithm 156, the Transfac matrices 100 were checked for binding sites in the promoter regions with a threshold for matrix similarity of 0.85. We counted the number of those gene sequences that were found to carry a predicted transcription factor binding site. As a reference set all unique genes of the PromoSer were reanalyzed. A one-sided χ² test and a one-sided Fisher's exact test (to improve the statistics for view counts) were performed with the statistical tool R 157 to determine the clusters with a higher affinity for a transcription factor.

All available 3'-UTR sequences (21,396) for mouse genes were derived with EnsMart 158, using Ensembl gene build for the NCBI m33 mouse assembly. 3'-UTRs for unique genes represented by the 780 selected ESTs were extracted using Ensembl transcript ID. A total of 234 mouse miRNA sequences were derived from the Rfam database 159. The 3'-UTR sequences were searched for antisense matches to the designated seed region of each miRNA (bases 1-8, 2-8, 1-9, and 2-9 starting from the 5' end). Significantly over-represented miRNA motifs in each cluster in comparison with the remaining motifs in the whole 3'-UTR sequence set were determined using the one-sided Fisher's exact test (significance level: P < 0.05) and miRNA targets of all clusters were analyzed for significantly over-represented miRNAs.

RefSeq sequences for 780 selected ESTs, shown to be more than two times upregulated or downregulated in a minimum of four time points during adipocyte differentiation and clustered according their expression profiles, were mapped onto the chromosomes from the NCBI Mus musculus genome (build 33) using ChromoMapper 2.1.0 software 160 based on MegaBlast with the following parameters: 99% identity cutoff, word size 32, and E-value (0.001). Colocalized sequences of all selected ESTs and from each of the 12 clusters within a 5 Mb genomic interval were identified. Within the 5Mb genomic intervals of each chromosome with the highest density of mapped ESTs, relative gene expression levels (log2 ratios) of these ESTs at different time points were related to the genomic localization.