Rewiring of the inferred protein interactome during blood development studied with the tool PPICompare

Background Differential analysis of cellular conditions is a key approach towards understanding the consequences and driving causes behind biological processes such as developmental transitions or diseases. The progress of whole-genome expression profiling enabled to conveniently capture the state of a cell’s transcriptome and to detect the characteristic features that distinguish cells in specific conditions. In contrast, mapping the physical protein interactome for many samples is experimentally infeasible at the moment. For the understanding of the whole system, however, it is equally important how the interactions of proteins are rewired between cellular states. To overcome this deficiency, we recently showed how condition-specific protein interaction networks that even consider alternative splicing can be inferred from transcript expression data. Here, we present the differential network analysis tool PPICompare that was specifically designed for isoform-sensitive protein interaction networks. Results Besides detecting significant rewiring events between the interactomes of grouped samples, PPICompare infers which alterations to the transcriptome caused each rewiring event and what is the minimal set of alterations necessary to explain all between-group changes. When applied to the development of blood cells, we verified that a reasonable amount of rewiring events were reported by the tool and found that differential gene expression was the major determinant of cellular adjustments to the interactome. Alternative splicing events were consistently necessary in each developmental step to explain all significant alterations and were especially important for rewiring in the context of transcriptional control. Conclusions Applying PPICompare enabled us to investigate the dynamics of the human protein interactome during developmental transitions. A platform-independent implementation of the tool PPICompare is available at https://sourceforge.net/projects/ppicompare/. Electronic supplementary material The online version of this article (doi:10.1186/s12918-017-0400-x) contains supplementary material, which is available to authorized users.

Establishing a fixed expression value that separates silenced from expressed transcripts is a somehow arbitrary choice [4,8]. Previously, authors used TPM thresholds ranging from slightly above 0.0 TPM [2,5] to 1.0 TPM [1] in various downstream analyses. To implement a threshold that reflects the actual protein concentrations in the cell, we used mass spectrometry-based proteome abundance data from HPM [6] for guidance. This independent experimental ressource shares information on CD4 and M cells with our constructed blood cell interactomes. Since the samples were derived from the same tissue of origin in both data sources and the annotational coverage was basically the same for these two cell types in HPM and in the reference PPIN (17,294 annotated HGNC protein-coding genes for HPM data and 17,292 proteins in the reference PPIN), we considered the reported proteome abundances and the number of proteins in our PPINs to be sufficiently comparable.
According to the HPM data, 8,140 proteins were detected in CD4 cells and 7,781 proteins in M cells. We compared these values to the number of proteins found in the constructed CD4 and M networks, i.e. the proteins for which the mRNA levels exceeded different expression thresholds (see Suppl. Table  S1 for details). The optimal TPM threshold for our purpose was determined as the one with the least absolute deviation between the number of proteins in the networks and the respective references (see Figure S4). This metric is considered robust concerning outliers [3]. With an average deviation of 553 proteins per sample, a threshold of 0.31 TPM gave the best agreement. In this case, the mean number of proteins in the 8 CD4 networks differed only by 117 proteins from the proteomic reference (10% of the standard deviation across samples) and the mean of the 5 monocyte networks differed only by 10 proteins (7.4% of the standard deviation across samples). All further analyses presented below were performed on the protein interaction networks constructed with the HPM-derived threshold of 0.31 TPM.

S1.2 Sample-size dependency of PPICompare results
Subsampling of a reference dataset is a common way to test the robustness of a differential analysis and was frequently conducted in the study of differential expression, see [7,9,10], for example. In the same manner, we tested PPICompare on subsets of samples for the transition CLP→CD4 (see Figure  2 in the main text) for which the BLUEPRINT data used contained the largest number of samples (5 samples for CLPs and 8 for CD4 cells). For both cell types we enumerated all subsets with at least 2 samples. We then combined the 26 subsets for CLP and the 247 subsets for CD4 cells to obtain 6,422 pairs of subsamples of the initial data with varying sample size. For all subsets we determined differential interactions among the network samples constructed with HPM-optimal threshold using PPICompare with FDR 0.05. Based on the results of the complete data for CLP→CD4 we calculated the precision (fraction of reference rewirings among all rewirings found by application to the subset) and the recall (fraction of reference rewirings found). From those we derived the balanced F-score (F 1 -score) as the harmonic mean of precision and recall. Figure S5 shows these measures as a function of the number of samples used in the corresponding calculations and the size of the smaller group, respectively. The latter is meant to grasp the effect of unbalanced group sizes.
It was not surprising to find that more data and a better balance generally led to a better overall agreement with the rewiring events detected on the basis of the full set of 13 samples (see F-score in the upper left panel of Figure S5).

Precision increases with the size of the smaller group
For subsets that had a smaller group of equal size, the median precision did often not increase considerably or even decreased with each added sample (see lower left panel of Figure S5). Instead, the precision depended more strongly on the size of the smaller group. This means that a better balance of the sample sizes between the groups was beneficial for the precision when the overall number of samples was fixed. The most remarkable increase of precision was consistently found when subsets with a group that only had 2 samples were compared to subsets that had 3 samples in the smaller group. Then, on average at least 19% (see 9 samples) and up to 34% (see 6 samples) more rewiring events within the results of each subset were found in the results of the complete data. For comparison, the increase from 3 to 4 samples in the smaller group was at most 18% (see 9 samples) and that from 4 to 5 only 15% (see 11 samples). The large number of rewiring events detected (see upper right panel) and a recall level that is similar to that of the more balanced subsets (see lower right panel) suggest that a high group imbalance causes an emphasis on the detection of the individual differences between the subsamples, rather than the more general differences over the group as a whole. As a consequence, more rewiring events were detected but the fraction of those that were significant in the whole dataset was smaller.

Recall increases with the overall number of samples
For the recall, the size of the smaller group had generally a less pronounced direct effect (see lower right panel of Figure S5). For subsets with a small group of only 2 or 3 samples, the increase seemed to happen stepwise. This was best visible for subsets of the data, where the smaller group contained 3 representatives: whereas the main part of the distribution of recall values was in a similar range for 6 to 8 total samples, the recall immediately increased from 9 samples onward to again reach a kind of steady state (see also the number of detected rewiring events in the upper right panel). This observation can be explained by the general implication of pairwise comparison and the strict significance filtering. When comparing pairwise with a small group of size s that includes a single outlier with skewed relevant feature f (say, a change in gene abundance), feature f cannot be recognized in 1 s of all cases and may thus not be found of significance anymore. However, with increasing sample size both the resolution of the sampling and the statistical power of the applied test increase, and may then enable passing a significance-threshold barrier.
Although having more data is of course always beneficial, our practical evaluation shows that even with less than half of the data satisfactory results with relatively few false positives could be obtained (average precision 0.78 with only 3 samples per group). We recommend to use PPICompare with at least 3 samples per group, because this yields a robust precision. Consequently, we only examined cell types with at least that much data available (see Methods section). EB_C006NS  EB_C006MU  EB_C006PO  MK_S00198  MK_C006NS  MK_S0018A  HSC_S001QB  HSC_S0025C  EB_C006X8  CLP_S001QB  HSC_S001FX  MPP_S001FX  MEP_C07002  GMP_C12001  MEP_C000WY  CMP_C12001  MEP_C12001  GMP_C07002  CMP_C07015  CMP_C07002  MEP_C07015  CLP_C07015  MPP_C07015  HSC_C002UU  CLP_C002UU  GMP_C002UU  MPP_C002UU  CD4_C001FR  EB_C0070X  CLP_S000SB  CLP_S001FX  HSC_C07002  HSC_C12001  CD4_S009W4  CD4_S004M7  CD4_S002EV  CD4_S0041C  CD4_S007DD  CD4_S007G7  CD4_S008H1  EB_S002R5  MK_S00K3I  N_PB130513  N_PB100713  N_PB270313  N_S00GRN  N_C000S5  N_C000R7  N_C12012  N_C0010K  N_C0011I  N_C001UY  M_C000S5  M_C0010K  M_C004SQ  M_C0011I  M_C001UY  EB_S002S3  MK_S004BT   MK_S004BT  EB_S002S3  M_C001UY  M_C0011I  M_C004SQ  M_C0010K  M_C000S5  N_C001UY  N_C0011I  N_C0010K  N_C12012  N_C000R7  N_C000S5  N_S00GRN  N_PB270313  N_PB100713  N_PB130513  MK_S00K3I  EB_S002R5  CD4_S008H1  CD4_S007G7  CD4_S007DD  CD4_S0041C  CD4_S002EV  CD4_S004M7  CD4_S009W4  HSC_C12001  HSC_C07002  CLP_S001FX  CLP_S000SB  EB_C0070X  CD4_C001FR  MPP_C002UU  GMP_C002UU  CLP_C002UU  HSC_C002UU  MPP_C07015  CLP_C07015  MEP_C07015  CMP_C07002  CMP_C07015  GMP_C07002  MEP_C12001  CMP_C12001  MEP_C000WY  GMP_C12001  MEP_C07002  MPP_S001FX  HSC_S001FX  CLP_S001QB  EB_C006X8  HSC_S0025C  HSC_S001QB  MK_S0018A  MK_C006NS  MK_S00198  EB_C006PO Figure S2: Comparison of rewiring events caused by co-deregulatory and non-co-deregulatory transcriptomic alterations. We separated all significant rewiring events into four groups of co-deregulatory and non-co-deregulatory types of transcriptomic alterations and examined their behaviour concerning unrelated indicators of modularity. We distinguished between the disjoint groups of rewiring events of mixed cause ("mixed"), those with one deregulated interaction partner ("one deregulated"), and those controlled by simultaneous deregulation of both partners ("both deregulated"). The last category additionally shows the results for rewiring events in which both interaction partners were proteins affected by the small set of likely changes in the transcriptome determined by the internal optimization approach ("reduced set"). The respective classification is introduced in the appropriate paragraphs of the main text. The plots on the left side show the results for all rewiring events affected by this mode of regulation while the plots on the right side depict the results per transition. The dashed red lines always represent the average of the respective measure in the complete reference network. Note that HSC→MPP, MPP→CMP and MPP→CLP had very few (< 5) co-deregulated rewiring events (see also Suppl. Table S2).   Figure S5: Sample-size dependency of PPICompare results for the transition CLP→CD4. We analyzed the rewired interactions found when only a subset of the available data was used and compared their agreement to the result obtained with all 13 samples that was taken as reference and thus set to 1.0.