Up to now there are no databases for genome-wide multi-protein interactions in mammals. Thus, we focused our study on 11 permanent and 21 transient human complexes of different sizes that are well characterized in the literature (Table 1, and see Materials and methods). Since transient complexes display dynamic properties, we analyzed microarray data from several temporal series describing a dynamic cellular condition. For this, we selected from the Stanford Microarray Database 24 three independent datasets analyzing the cell-cycle of HeLa cells synchronized either with double thymidine (Thy-Thy) or thymidine-nocodazole (Thy-Noc). In particular, only data from the first full cell-cycle (14 hours long) after synchronization were considered.

To extract putative protein-protein interactions from gene expression data, we first evaluated the Pearson's correlation for each pair of genes in the above described HeLa datasets. To assess if the number of highly correlated components had been obtained by chance, results were compared with the global behavior of the dataset by a standard hypergeometric test (Materials and methods).

Among the 32 analyzed protein complexes, 23 showed a p value lower than 0.05, including 5 in Thy-Thy dataset 2 (Thy-Thy2; Additional data file 1a), 10 in Thy-Thy dataset 3 (Thy-Thy3; Additional data file 1b) and 8 in the Thy-Noc dataset (Table 2). Among them (in particular in the very low p value range), a dominance of permanent with respect to transient protein complexes was observed. As an example, proteasome and small ribosomal subunit (SRS), which are well known stable complexes, were both characterized by very low p values in at least two datasets. However, we also found several complexes in which the number of highly correlated genes was clearly not statistically significant (that is, with a p value ≥ 0.7). In particular, this occurred in 15, 11 and 12 complexes in the Thy-Thy2, Thy-Thy3, and Thy-Noc datasets, respectively, including both permanent and transient complexes. RNA polymerase III is an example of a permanent complex without a significant p value in all three datasets.

As previously observed, the Pearson-based method was unable to detect significant correlations (that is, with a p value not ≥ 0.7) for almost half of the tested complexes. To improve the level of detection, we set up an alternative approach, which we call the 'expression peaks method'. Gene expression was analyzed every one (for Thy-Thy datasets) or two (for the Thy-Noc dataset) hours by computing the variation of mRNA levels between consecutive time points. A threshold was then defined on computed differences, which represents the value above which we considered the increase of expression between two consecutive time points a peak of expression. Next, we placed all computed expression values in a binary 1-0 system where 1 represents an expression peak. By calculating the expression peaks for each gene along the cell-cycle in each dataset, we found that a high percentage of genes participating in the same complex peaked synchronously at least in one temporal interval (Table 3 for the Thy-Noc dataset, and Additional data file 3 for the Thy-Thy datasets). Since there was more than one peak of expression per gene, we established the peak of expression of each complex as the time interval in which the genes of the complex peaked synchronously with the best p value (see below). To exclude that the number of synchronously peaking genes had been obtained by chance, we performed the same analysis on the Pearson's case described above by using a hypergeometric test. Among the 32 protein complexes analyzed, 14 in Thy-Thy2, 13 in Thy-Thy3 and 13 in Thy-Noc showed a p value lower than 0.05 in at least one time interval along the cell-cycle. As stable complexes we detected the mitochondrial large ribosomal subunit (MLRS), SRS, the proteasome and RNA polymerase II. Interestingly, low p values appeared for a large number of transient protein complexes in all three datasets; dynactin, exocyst, the nucleosome, the replication complex (RFC) and the skp1-cull-F-box complex (SCF) are transient complexes with a significant p value in two out of three datasets (Table 2 for the Thy-Noc dataset, and Additional data file 1 for the Thy-Thy datasets). Another remarkable difference with respect to the Pearson-based method is that we never found complexes with a p value ≥ 0.7.

To assess the quality of the expression peaks method in finding co-regulated genes that encode interacting proteins, we estimated false discovery rates (FDRs; see Materials and methods, and Additional data file 2). We plotted the FDRs for the Pearson correlation coefficient and the expression peaks methods as a function of the Bonferroni corrected p value (Figure 1). The results from the Thy-Noc dataset (Figure 1) indicate an additional benefit of the expression peaks method. Clearly, for each p value, the expression peaks method displayed a smaller FDR than the Pearson method. In particular, the p value that corresponds to a 10% FDR for the expression peaks method (p = 0.1, that is, -log₁₀(p value) = 1) corresponds to a 30% FDR for the Pearson's method.

Furthermore, we also compared the sensitivity of the Pearson and expression peaks methods (Figure 2). At a fixed FDR, the number of identified real complexes using either of the two methods was assessed. For the Thy-Thy datasets, with a low FDR, the Pearson's coefficient had a higher sensitivity in detecting high co-regulation among components of the same complex, while the expression peaks method clearly performed better across the different FDR ranges for the Thy-Noc dataset and at high FDRs for the Thy-Thy datasets.

The Pearson's coefficient analysis and the expression peaks method were also used to study protein complexes in additional time series datasets analyzing non-synchronized HeLa cells subjected to several stresses 25. Similar sensitivity for both synchronized and non-synchronized cells were obtained with the former method, while, as expected, the latter was more powerful in analyzing synchronized cells. Figure 3 and Additional data file 4 show the sensitivity of both methods for non-synchronized cells.

To improve the ability of the expression peaks method to identify new putative interactors of given genes, our approach was combined with an extensive GO annotation analysis. We developed a web based tool named PEGO (Peaks Expression and Gene Ontology) 26 to provide public access to such an analysis. PEGO selects two groups of genes; the first contains all the genes that have the same expression peak pattern as the input genes while the second includes all the genes with the same GO categories as the input. The user can then intersect the two sets of genes to identify the putative interacting proteins in their input dataset. Moreover, the tool allows the output data to be restricted, such as selecting preferred GO annotation terms or isolating a given time point in the array experiment 1925. In Additional data files 5-7 results are shown that were obtained by querying PEGO with a list of genes from a subset of the analyzed human complexes.

To test the predictive capability of our approach in detecting novel protein interactions, PAK1 was selected as candidate for study from the Thy-Noc dataset. PAK1 is a serine/threonine kinase implicated in the control of a number of cellular activities, including regulation of adhesive and trafficking processes, apoptosis, cell-cycle, and cytoskeletal dynamics 2728. We queried PEGO for PAK1 by using its ID [Entrez-Gene:5058], Organelle organization and biogenesis [GO:0006996] as the Biological process term and Cytoskeleton [GO:0005856] or Cytoplasm [GO:0005737] as the Cellular component term. According to this analysis, PAK1 was associated with three peaks. The highest percentage of genes with the same PAK1 GO annotation (Organelle organization and biogenesis) peaked in the time interval 14 h-12 h. Among them, 106 genes also displayed Cytoskeleton or Cytoplasm GO annotation (Additional data file 8); 5 of these genes are known interactors of PAK1 2930313233, 8 are similar to actin or actin-binding proteins, 4 are tubulins or tubulin-related proteins, 28 are proteins that localize also to the nucleus and 2 are involved in endocytosis. All these data largely match the known roles of PAK1, including the F-actin binding activity 28, the regulation of microtubule dynamics 34 and the involvement in cellular trafficking 283536.

Using the described approach, α-tubulin and EEA1 were selected as new interacting partners of PAK1 to be experimentally validated in living mammalian cells. Using immunoprecipitation assays, we detected the physical interaction between endogenous PAK1 and α-tubulin in HeLa cells (Figure 4, and Additional data file 9).

It is known that both PAK1 and EEA1 are involved in growth factor stimulated 3637 macropinocytosis 38 and that PAK1 localizes to ruffling F-actin areas where macropinosomes form 283940. Therefore, to investigate the interaction between PAK1 and EEA1, murine embryo fibroblasts (MEFs) were stimulated with platelet-derived growth factor (PDGF) to produce F-actin ruffles 4142. Because there are no suitable antibodies for PAK-1 immunofluorescence, the MEFs were transfected with PAK-green fluorescent protein (GFP). Figure 5a-c shows the colocalization of PAK-GFP with endogenous EEA1 in vesicle-like structures located in ruffling areas. A similar pattern was observed also in MEFs transfected with PAK1-mRFP (data not shown) to exclude any non-specific effect the fluorescent tag may have on the colocalization.

To further demonstrate the direct interaction between PAK1 and EEA1, we screened a phage displayed peptide library with the Cdc42/Rac interactive binding (CRIB) domain of PAK1 fused to the glutathione S-transferase (GST), in the presence of glutathione-derivatized sepharose beads. An increase in phage binding over the negative control (GST/glutathione beads) was observed after three rounds of selection. DNA sequencing revealed the presence of a peptide insert corresponding to amino acids 271-280 of EEA1 (Figure 5d). The specificity of this peptide was confirmed by ELISA, where its binding affinity was tested on GST-CRIB purified protein compared to GST protein alone. Figure 5e shows that the selected peptide had a specific affinity for GST-CRIB, supporting the physical association between PAK1 and EEA1.

In this work, we propose a new method to identify protein-protein interactions using gene expression data. The rationale behind our approach is the idea that a common transcriptional program drives the formation of both transient and permanent protein complexes in mammalian cells. It suggests that a selected gene expression dataset may contain useful information for de novo identification of protein interactions.

Because the decay rates of individual mRNAs range from 15 minutes to 24 hours 43, we focused our analysis on gene co-regulation in a single time interval to reduce noise. To asses the performance of our method with respect to the standard Pearson-based one we tested both of them with a set of 32 known complexes. To avoid problems due to multiple testing, we evaluated FDRs by comparing our results with those of thousands of randomly chosen sets of genes.

The main result of our analysis is that the study of synchronous peaks of expression can successfully complement the standard Pearson-based analysis of expression data. While Pearson-based methods are more effective in the identification of permanent interactions, our method is particularly suited for transient interactions. This observation suggests that it is best to use a combination of Pearson's and expression peak analyses for computational evaluation of protein complexes.

The higher sensitivity of the expression peaks method for transient complexes seems to be connected with its ability to detect quantitatively modest but functionally important changes in gene expression, which would otherwise be missed, especially in non-synchronized cell populations. With Pearson's analysis, a high statistical significance is obtained only for a small subset of complexes. In contrast, the expression peaks method gives statistically significant results for a greater number of complexes, although with higher p values than the Pearson's method.

Another important observation is that the expression peaks method performs better for well synchronized datasets (that is, the Thy-Noc treatment). On the basis of our methodological assumptions (that is, the half-life of mRNA 43), this is not surprising as it restricts the application of this method to highly selected datasets. However, the current technical efforts to improve cell synchronization will extend the reliability of the expression peaks method to a larger number of gene expression datasets.

The analysis of co-regulation during only a single time point increases the sensitivity of the expression peaks method, but also increases the noise. We therefore combined this method with GO analysis and found that this association reduced the number of false positives generated by the exclusive use of the expression peaks method. Combining both analyses reduced the number of potential candidate interactors to a few dozen while the output lists obtained by using either one of these approaches alone contained up to a thousand genes (Additional data file 6). A similar improvement was also recently observed by Corà et al. 44, who successfully combined GO and gene expression analyses in HeLa cell-cycle datasets to extract putative co-regulated genes for the identification of candidate transcription factor binding sites.

It is worthwhile to note that in several cases not all genes of the same complex were strictly co-regulated (Table 3, and Additional data file 3). This represents an intrinsic limitation of any approach based on gene expression data to identify protein interaction. Of course, subcellular localization, and post-transcriptional and post-translational modifications also play a key role in the assembly of both permanent and transient complexes 45464748. Thus, the addition of further information, such as post-translational modifications, could greatly improve the quality of the results, an approach we plan to use in the future.

To enable researchers to test our computational approach, we implemented our pipeline as a web-based, publicly available tool named PEGO 26, which we have queried to identify new protein interactions that we validated experimentally. It is interesting to observe that the PEGO outputs contained additional interacting partners whose genes were not included in our query list. For instance, in the case of the dynactin complex (Additional data file 6c) five new candidates emerged, and two of these, that is, non-erythrocytic spectrins, turned out to be previously characterized interactors of the dynactin complex 4950 (data not shown). While this result confirms the capability of our method to detect functional units, PEGO may actually be applied to a broader class of data types, in particular, to groups of genes without any known and obvious relationship. For example, one could analyze a list of genes that, if silenced, produce the same phenotype and use PEGO to detect any interactions among those candidates. Thus, unlike starting with a list of genes that have similar GO annotation, this approach excludes any prior bias for detection of protein-protein interactions. However, after a list of potential interactors has been generated, further GO analysis will increase the likelihood of detecting new complexes.

The potential of PEGO has been confirmed by 'wet' biological experiments testing the in silico results obtained by submitting PAK1, as a single gene. We selected PAK1 due to our interest in cytoskeleton dynamics in vascular cells. PAK1 relates best to the GO biological process 'Organelle organization and biogenesis', because this category includes both cytoskeleton- and vesicular-related functions that fit well with the subcellular localization of PAK1 in living cells (data not shown). Among the three expression peaks of PAK1 in the Thy-Noc dataset, we selected the 14 h-12 h one because a high percentage of 'Organelle organization and biogenesis' annotated genes peaked there, thus suggesting a novel intriguing role of PAK1 in this process.

PEGO indicated a list of genes to be considered as potential new interactors of PAK1. In light of the PAK1-related literature, we evaluated α-tubulin and EEA1 as potential interactors to be experimentally confirmed. Previous work showed a co-localization of microtubules and PAK1 51 and identified tubulin cofactor B (a cofactor associating with α- and β-tubulin) as an interacting substrate of PAK1 34. These data hint at an interaction between PAK1 and α-tubulin, although no experimental evidence has been obtained for this. We therefore used immunoprecipitation experiments in HeLa cells to demonstrate the physical interaction between PAK1 and α-tubulin supporting the immunofluorescence co-localization data previously reported 34.

The PAK1- EEA1 interaction, however, has not been reported before, and it represents a novel finding highlighting the potential of PEGO to predict unknown protein-protein interactions. Interestingly, we observed that PAK1 and EEA1 co-localize at sites resembling large vesicular structures. This hypothesis is supported by Dharmawardane et al. 36, who described the formation of large macropinocytic vesicles lined by PAK1 in PDGF-stimulated cells. Although the co-localization at these sites suggested a functional relationship between PAK1 and EEA1, the small amounts of overlayed proteins were not sufficient to test their physical interaction by immunoprecipitation. To overcome this technical problem, we screened a phage library to find specific peptides able to bind the CRIB domain of PAK1. Besides binding the small GTPases Cdc42 and Rac1, which trigger the catalytic activity of PAK1, the CRIB domain is also known to bind other transducers 22. The selection of a peptide encompassing an amino acid region of EEA1 (Figure 5d) clearly showed that the observed co-localization in immunofluorescence studies between EEA1 and PAK1 indeed reflects a true interaction.

Our statistical approach to identify protein complexes could be improved by taking into account a greater number of microarray gene expression data obtained by in vitro experiments performed on specific models of cell activation. The same approach applied to in vivo animal models should also allow the discrimination of changes in a putative complex caused by the tissue microenvironment or during development. More interestingly, the comparison of microarray data obtained before and after silencing a specific gene by small RNA interference could allow the identification of new protein complexes and not just simply the identification of new interacting partners.

Finally, a further and relevant progression of our expression peaks method would be the inclusion of other information besides GO to reduce the number of false positives. This could include sequence analysis, evolutionary data or the use of the same experimental design to generate expression data from different animal species.

We studied HeLa cell time series 1925 from the Stanford Microarray Database 24. The analysis was performed for every dataset separately to evaluate gene expression strictly related to specific cellular conditions. Datasets were normalized such that the Euclidean norm of each expression profile was 1 and the average was 0. To evaluate changes in expression level at each step during the cell cycle, for each clone, we built a new expression vector containing differences of expression values computed between two consecutive time points. Since the HeLa cell-cycle datasets were published in 2002, we associated with all IMAGE clones new Entrez Gene IDs according to UniGene database version 183 52, and we excluded from our analysis: clones with an ID different from the IMAGE ID or all numerical IDs; clones with more than one associated Entrez Gene ID; clones whose expression, measured as the log ratio between Cy5 (synchronous cells) and Cy3 (reference sample) channels, varied between -0.2 and +0.2.

The analyzed data set was composed of 32 permanent and transient protein complexes, in which interactions among different components are or are not maintained, respectively 14. Complexes were selected from the following sources: the KEGG database 53 for SRS, large ribosomal subunit (LRS), proteasome, anaphase promoting complex, von hippel-lindau complex, SCF, signal recognition particle, RNA polymerase II, RNA polymerase III, and TAFIID; NCBI 54 for trafficking protein particle complex, nucleopore, mitochondrial small ribosomal subunit, MLRS, adaptor-related protein complex 2, origin recognition complex, pyruvate dehydrogenase, ATP synthase, H+ transporting, mitochondrial F0 complex, ATP synthase, H+ transporting, mitochondrial F1 complex, and SNARE complex; literature for nucleosome 19, focal adhesion 55, centrosome 56, dynactin 57, Arp2/3 58, exosome 59, exocyst 60, axin-related complex (ARC) 61, SWI/SNF 62, cytochrome c oxidase 63, RFC 64, and golgi transport complex 65. To each protein the corresponding Entrez Gene ID was assigned according to UniGene database, version 183 52 (Additional data files 10 and 11).

We evaluated the gene expression correlation between clone ID pairs of genes using the Pearson Coefficient, r:

r = ∑ i = 1 n ( x