As previously mentioned, our investigation aimed to confirm current knowledge on the interactomes of PARP-1, PARP-2, and PARG as well as to discover new protein partners which could improve our understanding of into the functions of these key proteins. It was also hoped that a global view of the PARP-1, PARP-2 and PARG interactomes may help to grasp the ramifications of cancer treatment by PARP inhibitors, either in terms of therapeutic efficiency or side effects. Indeed, while PARP inhibitors have been recently shown to be active anticancer agents in BRCA1- and BRCA2-mutant tumors , fairly basic biological questions - including the biological roles of some of the interactors of PARP-1, PARP-2, and PARG - remain to be answered. Several PARP inhibitors are currently undergoing clinical trials and should be marketed as mid-stage results are promising (reviewed in ).
Intrinsic strengths and limitations of AP-MS
Compared to yeast two-hybrid and related methods historically used to define protein interactomes, AP-MS has three major advantages. Firstly, it can be performed under near physiological conditions, in the relevant organism and cell type. Secondly, it does not typically affect post-translational modifications, which are often crucial for the organization and activity of complexes . And thirdly, mass spectrometers can detect every abundant protein present in the immunoprecipitate, whether or not its presence is expected. Western blots, for instance, though very sensitive and specific, are incapable of that: they only target a few selected proteins expected to be in the sample from prior knowledge, and are therefore not conductive to the discovery of new, unanticipated protein interactors. Every analytical methodology - and AP-MS is no exception - possesses unique strengths and limitations, notably in terms of the number of different analytes it can detect simultaneously, and of its potential for false negatives and false positives.
False negatives, or "where did the interactors go?"
As pointed out by Gingras et al
, high-throughput AP-MS datasets typically lack many previously documented protein-protein interactions. This was notably observed for two comprehensive AP-MS studies in yeast [23, 24] and it is clearly the case of the current study's dataset as we failed to detect a number of known protein interaction of PARP-1, PARP-2 and PARG. Four factors likely contribute to this lack of detection : 1) the proteins may not interact under the tested conditions; 2) when a bait is expressed with a tag, the nature and location of the tag may disrupt certain interactions; 3) the conditions of the AP bee too harsh to preserve interactions; and 4) the lack of detection may be due to a problem with the relative abundance of proteins in the AP sample. The latter typically occurs if the bait is in large molar excess over its interacting partners (a problem that is exacerbated when the bait is overexpressed, as in our case) or if the prey is in much lower abundance than other components of the sample (low stoichiometry). AP-MS thus tends to capture more stable and fairly abundant interactors, in contrast to other targeted approaches such yeast two-hybrid assays . Finally interactions detected with western blots may not be detectable by AP-MS since western blot is generally more sensitive than MS for protein detection. We recognize that our investigation most probably missed certain low-abundant and/or low-affinity interactors of PARP-1, PARP-2 and PARG, and that more sensitive techniques will likely expand their interactomes in the future.
False positives, or "are the reported protein-proteininteractions real?"
Lack of specificity, i.e. reporting protein interactions that do not exist in the affinity-purified sample, may stem from either unspecific binding or incorrect protein identification. In the former case, immunoprecipitation material and reagents such as protein G, antibodies, or beads bind to proteins which are then falsely reported as binding the bait. In this study, two measures were taken to avoid reporting such non-specific interactions. Firstly, as detailed in Materials and Methods, BSA was added to the antibody-bead complex; BSA thus acted as a strong competitor (blocking agent) for all non-specific binding sites. Secondly, control immunoprecipitations using mouse immunoglobulin G as a bait were systematically used to discriminate between non-specific interactions and bona fide bait interactors. Proteins present in the controls were not reported. Reporting incorrectly identified proteins interactors was minimized through the use of stringent acceptance criteria for peptide and protein MS identifications, as described in Materials and Methods. Indeed, all reported peptide identifications have a probability of correct identification greater than 95%, and all reported protein identifications have a probability of correct identification greater than 99% and at least two accepted peptides. With these acceptance criteria, the overall MS false positive rate is estimated to be lower than 1%; in other words, the number of reported false positives due to incorrect protein identification should not exceed two. Mass spectrometry is widely recognized as a reliable method of identification of proteins in biological samples when the identifications are statistically validated and the false positive rate is kept low (reviewed in [25–27]). The use of Scaffold  to achieve such a statistical validation is also widely accepted with more than 40 recent peer-reviewed publications containing Scaffold-validated protein identifications. The fact that replicate AP-MS experiments were performed also increases confidence in the validity of the reported interactions as they were all observed twice or more.
Interpretation of the interaction datasets
A few points should be kept in mind while interpreting the PARP-1, PARP-2 and PARG datasets reported in Additional file 1. The first point is that AP-MS cannot distinguish between direct and indirect interactions; that is, reported interactions may correspond to two proteins which interact directly or, probably more frequently, correspond to two proteins which interact indirectly, via one or more bridging molecules (proteins, nucleic acids, or other molecules). The second point relates to overexpression of the bait. Transient over-expression of FLAG-PARP-2 and FLAG-PARG was required to circumvent the low specificity of available anti-PARP-2 and anti-PARG antibodies whereas FLAG-PARP-1 was transiently overexpressed for some of the PARP-1 experiments, in the hope of pulling some weak and/or scarce interactors of PARP-1 above detection limit. Such overexpressions may have resulted in the association of the bait protein with chaperones and could have led to improper intracellular localization . Finally, one must keep in mind that some protein-protein interactions detected in a cell lysate may not actually occur in vivo if the two protein partners never co-localize within the cell.
PARP-1 and PARP-2 may bind some proteins through their pADPr moiety
The PARP-1 and PARP-2 used as baits may have presented some degree of automodification, in which case their pADPr moiety - rather than their protein interaction domains - may have bound some protein interactors. Gagné et al. recently investigated the interactome of pADPr by AP-MS and reported a large number of pADPr-binding partners . Interestingly, the interactors of PARP-1, PARP-2 and PARG listed in Additional file 2 include 53 of these pADPr-binding partners, and thus support the above hypothesis. These 53 common interactors are presented in Additional file 3 and consist mostly of DNA/RNA-binding proteins such as ribonucleoproteins, translation initiation factors, helicases, and ribosomal proteins.
Functional analysis using Gene Ontology
BiNGO analysis suggests that PARP-1 may be involved in more biological processes than PARP-2 or PARG. Indeed, whereas all three enzymes are involved in RNA processing and both PARP-1 and PARP-2 participate in DNA repair, glycolysis and apoptosis, only PARP-1 partakes in cell cycle regulation and ubiquitin conjugation. The involvement of PARP-1 in DNA repair and apoptosis is well known (reviewed in ) and was expected. However, its apparent interaction with glycolysis and RNA metabolism proteins is surprising and may point to new PARP-1 functions. Devalaraja-Narashimha and Padanilam  recently studied the impact on glycolysis of the inhibition of glyceraldehyde phosphate-3-dehydrogenase by PARP-1, while others have focused on how NAD consumption (energy depletion) by PARP-1 influenced glycolysis [31–33]. PARP-1 is responsible for the majority of pADPr formation following DNA damage and excessive pADPr formation promotes shuttling of AIF to the nucleus and caspase-independent cell death . The detailed mechanism is yet to be understood, but it has been suggested that pADPr may be shuttled from the nucleus to the cytoplasm via proteins, resulting in AIF translocation, as well as the sequestration of anti-apoptotic proteins or the activation of pro-apoptotic proteins [35, 36]. The "apoptosis" branch of Figure 4 suggests that PARP-2 probably also plays a role in apoptosis, presumably because it can also add pADPr to proteins. PARP-2 apparently plays a less important role in apoptosis than PARP-1 given the lower complexity of PARP-2's apoptosis branch. The biological processes shown for PARG in Figure 5 are mainly linked to RNA metabolism and confirm the findings of previous studies. Indeed, with the majority of PARG activity located in the cytoplasm, maintaining tight spatial regulation of this enzyme depends on nucleocytoplasmic shuttling proteins such as the ribonucleoprotein (RNP) complexes . Moreover, poly(ADP-ribosyl)ation of heterogenous nuclear ribonucleoproteins (hnRNPs) was determined to be of importance for protein splicing in Drosophila, suggesting that PARP-1 and PARG modulate splicing pathways through regulation of interactions between hnRNPs and RNA .
Involvement of PARP-1, PARP-2 and PARG in RNA metabolism
Figures 3, 4 and 5 indicate that PARP-1, PARP-2 and PARG are all involved in RNA metabolism and, more precisely, in RNA splicing. For instance, hnRNP M and hnRNP U, which are both part of the spliceosome C complex , are among the 11 pulled-down proteins common to PARP-1, PARP-2 and PARG. Moreover, PARP-1, PARP-2 and PARG also interact with several other members of this complex (Additional file 1). PARP-1 interacts with splicing factor 3A1, splicing factor 3B2, splicing factor 3B1, hnRNP C, snRNP EFTUD2 and U5 snRNP 200 kDa; PARP-2 with DEAD box polypeptide 48, ATP-dependent RNA helicase DDX5, hnRNP A1, hnRNP G and Polyadenylate-binding protein 1; and PARG interacts with ATP-dependent RNA helicase DDX5, hnRNP A1, hnRNP Q, hnRNP A2B1 and Polyadenylate-binding protein 1. Many of these proteins share interactions with more than one pADPr-metabolizing enzyme. The fact that each bait pulled down different components of the spliceosome C complex might be explained by the dynamic nature of this complex but is more probably indicative of a lack of affinity for the bait, or of a lack of analytical sensitivity.
Our dataset includes other interesting PARP-1 interactors related to transcription and splicing, namely the THO/TREX and FMRP complexes. THO complex subunits 1, 2, 5 and 6 were identified by mass spectrometry in PARP-1 immunoprecipitates. THO/TREX is required for the nuclear export of mRNA and is coupled to transcriptional elongation [39, 40]. The THO complex is composed of multiple subunits linked to Aly and Bat1, which form the TREX complex responsible for RNA export. The THO/TREX complex is also recruited to splice mRNA, and is indirectly associated to transcription via the splicing machinery . Many components of the FMRP complex, which suppresses translation of selected mRNA within large mRNP complexes , are present in our dataset: Fragile-X mental retardation syndrome-related protein 1 (FXR1) (common to PARP-1, PARP-2 and PARG datasets); FXR2 and FMR1 (PARP-2 and PARG); and nuclear Fragile-X mental retardation-interacting protein 2 (PARG only). While PARP-1  and PARG  have already been identified in mRNP complexes, their role therein remains unclear although some evidence pointed toward pADPr regulation of transcriptional activity. Our finding that PARP-1 and PARP-2 are associated with FMRP particles strengthens this hypothesis. Both PARP-2 and PARG also interact with nuclease-sensitive element-binding protein 1 (YBX-1), which is involved in mRNA binding, regulation of mRNA stability, translation efficiency and, possibly, binding splicing enhancer elements [44, 45].
The involvement of PARP-1, PARP-2 and PARG in the regulation of RNA splicing is corroborated by many studies. Recently, Ji and Tulin  demonstrated in Drosophila that endogenous PARP-1 and PARG regulate pADPr binding to hnRNPs, subsequently altering the RNA-binding ability of hnRNPs and modulating splicing. Another study by Malanga et al.
 confirmed that pADPr can bind the splicing factor ASF/SF2 within domains crucial for splicing activity, thereby regulating splicing. pADPr is also involved in transcription-splicing through its role in Cajal bodies . Small spliceosomal components first accumulate in Cajal bodies and then undergo a spliceosome phase assembly; they are then transported to the cytoplasm for splicing events and are finally reintegrated to Cajal bodies to be recycled [48, 49]. Automodified PARP-1 is responsible for protein shuttling into Cajal bodies and PARP-1 is itself crucial for the integrity of Cajal bodies . Furthermore, PARG deletion leads to an imbalance of proteins in the Cajal bodies, demonstrating that pADPr is an important regulator of shuttling into Cajal bodies. As illustrated above, PARP-1, PARP-2 and PARG share several interactors related to splicing and transcription, and poly(ADP-ribosyl)ation, or pADPr binding, presumably underlies these interactions. Indeed, and as mentioned previously, many interactors of PARP-1, PARP-2 and PARG identified in this study are also known to be associated with pADPr . One can hypothesize that PARP-1, PARP-2 and PARG regulate some transcriptional and splicing events through the addition and subsequent removal of pADPr chains on specific substrates.
Dual DNA-dependent roles of PARP-1 and PARP-2: DNA damage signaling and cell death
It is well known that genomic stress induces pADPr formation via DNA-dependent PARP activation. As the severity of DNA strand breaks increases, more pADPr is synthesized and the extent of poly(ADP-ribosyl)ation determines the cellular response: mild genotoxic stress induces PARP-1 and PARP-2 activation and signaling to promote DNA repair while severe stress generates extreme pADPr accumulation that triggers cell death.
DNA damage signaling
BiNGO analysis revealed similar significance for the "DNA repair" branches of PARP-1 and PARP-2 (Figure 3 and 4). Both PARP-1 and PARP-2 immunoprecipitates contained KU70 and KU80, two proteins involved in DNA double-strand break repair by NHEJ, thus confirming the redundant function of PARP-1 and PARP-2 in the surveillance of genome integrity. Other interactors linked to the DNA repair function of PARP-1 and PARP-2 are RNA-binding protein 14  (observed for both PARPs) and Replication factor C (RFC) subunits 1, 2, 3, 4 and 5 (observed for PARP-1 only). RFC1, RFC2, RFC3, RFC4 and RFC5 form a heteropentamer complex that interacts with PCNA and enables the binding of its N-terminal DNA-binding domain to duplex DNA. This mechanism is essential in the recognition of non-primer template DNA structures during replication and/or repair . PARP-2 interactors also include Replication Factor A protein 1 and the catalytic subunit of DNA-dependent protein kinase (DNA-PK). The former protein participates in the very early stages of initiation in both DNA recombination and DNA replication, through association with DNA-PK and, possibly, through the recruitment of NHEJ proteins . Globally, our interactor datasets are consistent with the known roles of PARP-1 and PARP-2 in DNA repair initiation via the BER [4, 53, 54] and NHEJ [5, 55] pathways. Our results suggest that PARP-1 and PARP-2 may signal DNA damage by means of multiple complexes in order to speed up DNA repair.
As previously mentioned, activation of the DNA-dependent PARP-1 and PARP-2 under severe genomic stress leads to substantial accumulation of pADPr, which triggers AIF translocation, and eventually cell death . Our study reveals some unknown details of this apoptotic pathway. For instance, GRP78, HSC70, HSP90AB1, and HSP70, four heat shock proteins, are all interactors of PARP-1, PARP-2, and PARG. Although heat shock proteins are frequent immunoprecipitation contaminants , the above four proteins were absent from our control datasets and we consequently believe them to be bona fide interactors. These heat shock proteins act as chaperones and are likely to attach to multiple unrelated proteins, given that they bind mutated and unfolded proteins to prevent their secretion. However, some heat shock proteins exhibit interactions or functions similar to PARP-1 and PARP-2, notably in apoptosis. For instance, GRP78 is involved in the negative regulation of apoptosis by suppressing the activation of caspase-7 and caspase-12  while HSC70, an anti-apoptotic co-chaperone, inhibits HSP90 and other apoptotic proteins . HSP90AB1 interacts with TP53 within a multi-chaperone complex in which HSP70 plays an important role in the process of apoptosis . The heat shock protein 60kDa mitochondrial (HSP60) is another interactor that is shared by PARP-1 and PARP-2. One of the known functions of HSP60 is its contribution to the regulation of apoptosis through its association with caspase-3  and BAX . Also of interest among PARP-1 interactors are Baculoviral IAP repeat-containing protein 6 (BIRC6) and BTF. The latter binds to DNA and represses transcription of survival genes , while BIRC6 is an inhibitor of apoptosis through its inhibition of caspases, particularly of caspase-3 .
An interesting hypothesis yet to be confirmed is that poly(ADP-ribosyl)ation affects multiple proteins in order to shift the balance between pro-apoptotic and anti-apoptotic molecules. Indeed, interfering with NAD and pADPr metabolism increases mRNA and protein GRP78 levels [64, 65]. Parp-1-/- fibroblast cells exhibit increased expression of HSP70 , a protein that delays nuclear translocation of AIF . Moreover, increased levels of HSP70 reversibly inactivate PARG, and thus cause pADPr accumulation . Our observation of likely protein interactions between PARP-1/2 and some heat shock proteins raises the possibility that these PARPs could affect the function of heat shock proteins during apoptosis through transcriptional regulation.
A possible new function for PARP-1
The role of pADPr in the cell cycle is generally associated with mitotic spindle functions. This structure consists of an array of microtubules and various molecules that self-organize to align and then segregate chromosomes during mitosis. pADPr is found at the spindle and is required for its function . Our data highlights that another possible role of PARP-1 in the cell cycle, as the regulation of the APC/C complex. Indeed, we report here for the first time that PARP-1 interacts with eight of the twelve proteins belonging to the APC/C complex, one of the two poly-ubiquitylating E3 ligase complexes influencing cell cycle progression. APC/C is largely associated with cell cycle progression and sister chromatid separation (reviewed in ). The four APC/C proteins not observed in PARP-1 immunoprecipitates have low molecular weights (21 kDa and lower) and are therefore more challenging to identify by mass spectrometry since they generate fewer tryptic peptides upon digestion. This is probably why they were not detected. Chang et al.
 demonstrated that PARP-1 is not critical for mitotic progression and that it does not play an essential role in the regulation of anaphase entry in the cell cycle. However, PARP-1 has been shown to bind and poly(ADP-ribosyl)ate BUB3 , which is essential for the recruitment of this protein to centromeric heterochromatin . BUB3 is suggested to act as a regulator of the APC/C complex. It is unclear how PARP-1 could be involved in mitotic progression, but our findings motivate further investigation of this potential function.