From the FS composed of 38,832 interactions involving 9,813 proteins, we extracted a Confidence Subset (CS) composed of interactions that likely occur in macrophages, using functional and statistical parameters. For this, we used principal component analysis (PCA) that allows assembling parameters showing similar behaviours (see Material & Methods for details). According to the correlations obtained, the number of reports and evidences are correlated as well as the number of common Gene Ontology terms describing the cellular components and biological processes in which protein pairs are involved (Figure 2). These statistical observations are used to discriminate the CS interactions.

Considering that gene co-expression is routinely used as a parameter in contextualization attempts 25 26 , we included only interactions between the products of genes co-expressed in macrophages in the CS. Ultimately, considering that proteins composed of known interacting domains have higher confidence in the PPI network, we selected only interactions between partners sharing interacting domains according to PFAM annotations.

Overall, each interaction belonging to the CS obeys the following criteria (Figure 3, Material & Methods for details):

1) the genes encoding the interacting proteins must be co-expressed in normal, uninfected macrophages;

2) the protein partners must share interacting domains according to PFAM;

3) protein partners must share functional Gene Ontology annotations;

4) the interaction must have been identified several times by independent experiments.

In this way, a CS composed of 530 interactions involving 594 proteins was obtained. The analysis of the Gene Ontology terms annotating those proteins showed that the CS is enriched in terms related to immune system development (Fold Change, FC = 3, p-value = 1.06 × 10⁻⁵), to regulation of apoptosis processes (FC = 3.29, p-value = 2.2 × 10⁻²⁶), to regulation of cell death (FC = 3.6, p-value = 8.1 × 10⁻²⁹) and to regulation of I-kappaB kinase/NF-kappaB cascade (FC = 5.1, p-value = 5.7 × 10⁻¹¹) (Additional file 1: Table S1).

Aiming to further gain confidence in the CS, we compared our empirical filtering process to clusters obtained upon applying an unsupervised clustering method to the FS. Interestingly, the Self Organizing Map (SOM) 34 analysis showed that 64% of the interactions contained in the CS are grouped in a single cluster, the remaining interactions being located in 5 out of 16 clusters (Figure 4). This shows that the interactions grouped into the CS according to the criterion empirically chosen (described above) are in agreement with clusters obtained mathematically using an unsupervised algorithm.

In conclusion, the functional enrichment of relevant groups of genes and the satisfactory comparison to unsupervised clustering reinforce the hypothesis that the interactions composing the CS likely occurring in the macrophage.

To identify the most likely macrophage PPI network, the interactions most resembling those of the CS were selected by computing a similarity distance. To this end, the CS interactions barycentre was first identified and compared to the descriptor values of the FS interactions. In this case, the barycentre is computed as the centre of mass of all the CS interactions. In other words, considering that CS interactions represents a cloud of points in a multidimensional space with an axis for each of the variables (descriptors), the barycentre of these interactions is defined by the mean of each variable. The barycentre is identified for CS elements as a centroid point whose coordinates represent a vector as follows: 1 n ∑ i = 1 n CS d 1 , 1 n ∑ i = 1 n CS d 2 , … , 1 n ∑ i = 1 n CD d 8 , where CSd_i represents the confidence subset descriptor index.

Second, we computed and compared the distributions of the Euclidean distance values between the barycentre and the CS interactions on one hand, and the FS interactions on the other hand (Figure 5). We then considered as possible in a macrophage, all the FS interactions showing a distance value to the barycentre less than 4.2, i.e. the value corresponding to 95% of the surface of the CS distribution. In other words, this cut-off was used to select a CS-like behaviour among the FS interactions.

The resulting Contextualized Interactome (= CI) is composed of 30,182 interactions involving 8,633 proteins, corresponding to 75% of the initial FS. This ratio can be taken to mean that nearly 75% of the interactions composing an interactome are possible in a given tissue 35 .

In order to increase our confidence in the contextualisation process, we verified the functional enrichment of the CI compared to the FS. We found that interactions involved in regulation of apoptosis (FC = 1.05, p-value = 1.05 × 10⁻⁷) and cellular death mechanisms (FC = 1.044 p-value = 2.2 × 10⁻⁴) (Additional file 2: Table S2) are enriched, underlining the over-representation of pathways involved in the immune response to pathogenic exposures in the CI.

Although the CI corresponds to three quarters of the FS, its functional terms are more significantly enriched compared to those of the FS (See Figure 6 and Additional file 3: Table S5). Similarly, the CI observed functional annotations terms are more significantly enriched compared to those obtained from randomized interactomes (Additional file 4: Figure S2).

To further assess, statistically, the resulting CI, we compared the host regulatory network following an exposure to MTB 31 with the CI and with randomly obtained interaction sets. Interestingly, the CI is significantly enriched (p-value < 2.2 × 10⁻¹⁶; t-test) in interactions reported in the MTB regulatory network compared to random interaction sets. Likewise, the CI is statistically enriched in interactions experimentally identified in macrophages according to InnateDB 32 , a database devoted to innate immunity (t-test p-value < 2.2 × 10-16, see Materiel & Methods for details).

To complement our analysis, we computed the overlap of CI with interactomes contextualised using other sources: the macrophage proteome from Protein Atlas 29 and HPRD 30 (see the Materiel & Methods for details). The CI overlaps satisfactorily with the HPRD macrophage interactome (p-value = 0.022) and more significantly with the Protein Atlas macrophage interactome (p-value = 3.56 × 10⁻²²) (see Figure 7).

Altogether, these comparisons summarised in Figure 7 emphasize the “macrophagic” specificity of the contextualized interactome (CI).

In order to evaluate the pertinence of the contextualized macrophage interactome, we used it in the following study case.

The expression signatures of macrophages infected with MTB have been characterized in three independent studies 31 36 37 . By combining these data, we obtained two lists of down-regulated and up-regulated genes upon MTB infection. Based on the SAM algorithm (Additional file 5: Figure S1), we have ultimately cumulated 3,724 under-expressed genes and 1,651 over-expressed genes from the three transcriptomic experiments. We focused on these genes, knowing that MTB infection regulates the activity of particular host genes and cellular processes to its own benefits. To evaluate the insights brought by a PPI level analysis versus a classical differential gene expression approach, we extended the list of genes revealed by SAM to their first interactors in the CI, thus defining two sub-networks of 2,966 and 1,435 interactions anchored respectively on the 3,724 under-expressed and the 1,651 genes over-expressed upon infection (note that not all the modulated genes have interactions in the CI). We then compared the functional enrichments of the modulated gene lists and their resulting sub-networks. As shown in Additional file 6: Table S3, whereas the GO terms ‘response to oxygen levels’, ‘cell substrate adhesion’, ‘cell matrix adhesion’, ‘positive T cell selection’ are the most enriched terms when only under-expressed genes are considered, ‘regulation of programmed cell death’, ‘negative and positive regulation of apoptosis’ or ‘response to wounding’ are found to be over-represented when the interactors are taken into account. Similarly, considering the up-regulated genes and their associated sub-network led to the same finding (see Additional file 7: Table S4).

Therefore, focusing on the interactions involving the products of the regulated genes rather than only on the expression of the genes favours the emergence of functional aspects caused by MTB infection. Among these aspects, the regulation of the apoptosis is known to be highly targeted and controlled by the pathogen during the different phase of infection and persistence in the macrophage, as is nicely discussed by Lee and colleagues 38 . Notably, although these regulatory aspects are crucial for the outcome of infection, they are more significantly and extensively revealed at the systemic scale by focusing on the PPI. These findings highlight the need to consider infection of the host by a pathogen at the level of the functional module, defined as a group of interacting proteins involved in the same pathway or biological process, instead of focusing solely on genes or their products.

Moreover, considering the interactome revealed that the products of the down-regulated genes after infection, are closer to each other in the network than the rest of the CI proteins. This supports the hypothesis that MTB targets proteins participating to the same pathways. Indeed, the shortest path values between the down-regulated genes are significantly lower than the shortest path between the CI proteins (Mean paths for CI and down-regulated genes within the CI are respectively 3.3 and 4.5 (p-values 0.002557; t-test)).

Overall, these results suggest that the bacillus acts upon key proteins, which are closely connected within the network to regulate the host response.

From APID we extracted an interactome dataset composed of 38832 interactions involving 9831 proteins. Features were added to compose a dataset of interactions described by functional and quantitative descriptors:

1. # methods: This information is extracted from APID and corresponds to the number of experimental validations describing the interaction according to the molecular interaction controlled vocabulary PSI-MI 43 . Only leaves of the PSI-MI experimental validation tree were selected.

2. # publications: extracted from APID. Corresponds to the number of articles indexed in PubMed and reporting the interaction.

3. iPFAM value: extracted from APID. Identifies whether the interactors pair contains domains known as interacting according to the Pfam database 44 .

4. GO-proxy: this program is part of the GOToolBox suite 45 . It computes a similarity index between the interactors on the basis of the GO annotation terms they share. The similarity index corresponds to Czekanowski-Dice formula 13 46 .

5. # of common GO biological process terms: represents the number of common GO biological processes shared by the interacting proteins. For sake of precision, we only consider terms found at level 3 in the ontology tree.

6. # of common GO cellular component terms: corresponds to the number of common GO cellular components shared by interactors.

7. # of common KEGG pathways: corresponds to the number of KEGG pathways shared by the interactors.

8. Co-expression value: macrophage expression data from Chaussabel and colleagues 36 , downloaded from the Gene Expression Omnibus database (GEO) 47 . Each probe set corresponds to a mRNA and was categorized either by Present, Absent or Marginal. The Presence/Absence call of the mRNA was calculated according to the MAS5.0 algorithm 48 .

To evaluate the occurrence of the interaction considering the Presence/Absence status of the mRNA, we assumed the following hypotheses:

i. the presence of the mRNA implies the presence of the corresponding protein: the mRNA is detected as present according to the SAM algorithm 49 ;

ii. for a couple of proteins interacting in vitro, if both proteins are considered as present within a targeted cell according to the hypothesis (i), we assume that the interaction is bio-physically possible in that condition.

The CS relevance was assessed by using two distinct clustering algorithms.