Main

The latest craze in science is systems biology. Scientists in different scientific fields, however, define and view “systems biology” in different ways. Systems biology, for example, can mean mathematical modeling of a system for accurate predictions of the behavior of the system. It can also mean the systematic application of different high-throughput approaches to gather information on the system and to combine different information spaces for a more comprehensible view of the biology at hand. To date, the best modeling and predictions have been on systems for which good biological information is available 1. Hence, gathering information on the system is required to develop predictive models. Genomics has been the primary source of large-scale information on systems through genome sequencing, large-scale SNPs analyses, and gene expression studies.

The success of genomics has snow-balled into other areas, such as the large-scale data gathering at the protein level (proteomics). One of the achievements of proteomics is the mapping of protein interactions (interactome; the ensemble of the protein interactions related to a proteome or a genome). We have seen a growing number of large-scale studies of the interactome (or part of) for different species including yeast 2, 3, 4, 5, 6, 7, Drosophilia 8, Helicobacter pylori 9, Caenorhabditis elegans 10, 11, cyanobacterium Synechocystis 12, and human 13, 14, 15, 16, 17. Furthermore, the mapping of individual protein interactions and studies focused on subgroups of proteins, for example, bacteria Flagella 18, the Smad signaling system 19, and C. elegans 26S proteasome 20, are ongoing in many laboratories around the world where mass spectrometry 21 and techniques, such as yeast two hybrid (Y2H) 22, are readily available. Already in yeast we have a semi complete map of the interactome. Unfortunately, in human, we only have a partial map of the interactome, which is complicated by the different cell types and cellular localizations, thus limiting the broader application of systems modeling.

Should the human protein interactome be mapped?

The first question can be easily answered and is reminiscent of early discussions about sequencing the human genome. Many scientists were doubtful that sequencing the human genome would be worth the effort. It is now clear that today's research in human diseases and biological processes benefits from the human genome project. We already have a glimpse of the potential outcomes of mapping the human interactome by looking at the yeast interactome. The importance of mapping the yeast interactome can be measured in different ways; however, the number of references to the original papers by Fields 22 (over 3 000 citations) for interaction mapping by Y2H; and Ho 4 (over 1 300 citations) and Gavin 6 (over 1 400 citations) for protein interaction mapping by mass spectrometry clearly illustrate the changes in understanding yeast biology provided by the yeast interactome. We could reasonably expect a similar or greater impact in human biology by the human interactome.

Are techinques with sufficient throughput available?

From a technical point of view, the broadly accepted techniques of Y2H 22 and affinity purification combined to mass spectrometry 21 can potentially be scaled up. Assuming that 24 000 genes are present in human, testing all the possible interactions using Y2H would require 3×108 tests, not including the repeats and all the splice variants and mutants. This number might well be within reach by the Y2H technique as recent studies have performed over 107 Y2H assays 14, 23. For affinity purification/mass spectrometry, it would require 24 000 experiments, not including the repeats and all the splice variants and mutants. Recently, we published the interaction map for 400 human genes obtained in 293T cells by affinity purification/mass spectrometry 16. Hence, it is conceivable that the human protein interactome could be mapped by Y2H and affinity purification/mass spectrometry.

Will the data quality and coverage be sufficient?

Neither Y2H nor affinity purification/mass spectrometry provide the complete list of interactions 24. Futschik et al. 25 compared eight large-scale maps with a total of over 10 000 unique proteins and 57 000 interactions based either on literature search, orthology, or by Y2H assays. Their comparison reveals a small, but statistically significant overlap. More importantly, their analysis gives clear indications that all interaction maps imply considerable selection and detection biases. Our studies of the overlap between Y2H and affinity purification/mass spectrometry in human indicate at best a 11% overlap 16. A closer assessment of the results indicates that the techniques are complementary because they each provide a partial view of known protein complexes. Furthermore, our current methods to represent and compare Y2H and affinity purification/mass spectrometry do not take into account the differences in the data, i.e. binary interaction data (Y2H) and affinity purification/mass spectrometry data (direct and indirect interactions). We have shown a greater overlap (2.6-fold increase) when taking these differences in consideration 16. Finally, a significant contribution to the low overlap between the two techniques is false positives. False positives can be greatly reduced by taking into account the known localization of proteins and their functions. Reguly et al. 26 demonstrated that of the interaction data obtained in yeast by high-throughput techniques 20% are between proteins that are involved in the same biological process. Furthermore, 27% of the protein interactions are between proteins that have the same cellular localization.

Choosing only one approach to map the human interactome, although appealing from a cost point of view, would likely provide a poor interaction map. We are proposing as a discussion point a process for studying the human interactome that will provide an increasing level of refinement through a multiple pass process with each pass leading to a more confident interaction map (Figure 1). Hence, it is likely that the human interactome will need to be mapped using the combination of different technical approaches.

Figure 1
figure 1

Combination of different protein interaction techniques to increase the level of confidence in the human interactome while minimizing the number of experiments to perform.

Full size image

Are high-throughput techniques standardized?

The mapping of the human protein interactome will require a coordinated international effort. It is not possible to only use one scientific approach; therefore, the coordination of various approaches, such as Y2H and affinity purification/mass spectrometry, needs to be considered. Standardization of methods will be very important for studying the human interactome.

Reagents

Y2H and affinity purification/mass spectrometry both require access to clones. Different collections of clones are available. The latest version of the human ORFeome contains over 12 000 full length clones of human genes 27. Invitrogen has over 35 000 full length clones in their collection while the FLJ-DB database (www.nedo.go.jp/bioiryo/bio-e/index.html) reports over 20 000 clones (although not necessarily full length). Although the availability of full length clones has improved drastically, the annotation and quality of clones remains an important issue. It is likely that clones representing most human genes will be readily available over the next few years.

High-throughput Y2H

The Y2H system is currently one of the most standardized protein interaction mapping techniques. Over the past years, several large-scale protein-protein interaction data-sets have been published 2, 3, 9, 11, 28, 29.

In a Y2H assay, the two proteins to be tested for interactions are expressed with amino-terminal fusion moieties in the yeast Saccharomyces cerevisiae. One protein is fused to a DNA-binding domain (BD) and the other one is fused to a transcription activation domain (AD). An interaction between the two proteins results in the activation of reporter genes that have upstream binding sites for the BD. Large arrays of AD and BD strains representing most of the proteins encoded by a genome have been constructed and used to systematically detect binary interactions. Most large-scale screens have used arrays in a library-screening approach in which the BD strains are individually mated with a library containing all of the AD strains pooled together. Detailed protocols for large-scale two-hybrid analyses have been described in references 30 and are also available on a website (http://vidal.dfci.harvard.edu).

Li et al. 31 recently reported a study on how protein-protein interaction (or “interactome”) networks relate to multicellular functions. They have mapped a large fraction of the C. elegans interactome network. Starting with a subset of metazoan-specific proteins, more than 4 000 interactions were identified from HT-Y2H screens.

Y2H screens are often criticized for generating high rates of false-positives. Vidalain et al. 32 described false-positives as either being biological false-positives (artificially occurring in yeast cell) or technical false-positives (due to limitation in the techniques). They developed various approaches to reduced false-positives due to technical limitations of HT-Y2H. von Mering et al. 24 also demonstrated that the biological false-positives can be reduced by combining results obtained from different studies including results obtained from mass spectrometry based approach 24. As well, Jin et al. 33 recently described a small pool array that increases the screening efficiency by one order of magnitude and reduces the false positives. However, it still remains that the study of human genes using the Y2H approach cannot reproduce the proper cell types, the localizations of the proteins, the basal levels of expression, and it might not provide the necessary post-translational modifications and processing.

Affinity purification/mass spectrometry

Affinity purification has been a basic methodology of modern biology. Its recent coupling with mass spectrometry has increased the speed and ease of identifying protein interactors 34, 35. The combination of affinity purification of protein complexes with mass spectrometry allows the rapid identification of different proteins involved in a complex. This approach is promising and is likely to be very useful for generating part of the human interactome map. The standardization of methodologies for protein interaction mapping by affinity purification/mass spectrometry is lacking and many protocols still need to be developed. “One-size-fit-all” protocols will unlikely emerge. It is more likely, however, that a suite of standardized protocols that are applicable to different protein classes, protein localizations, and potentially cell types will be developed.

At this point, we will discuss some of the technical hurdles that affinity purification faces. The major drawback of this approach is that it cannot be done without disturbing the system. This disturbance of protein interactions can occur at different levels such as: at the level of tagging of the bait protein; at the expression level;when lysis of the cells occurs; and also when the protein complexes are purified using different protocols (Figure 2).

Figure 2
figure 2

Steps involved in protein interaction mapping by affinity purification coupled to mass spectrometry. Possible issues are highlighted.

Full size image

To tag or not to tag

The best situation occurs when monoclonal antibodies against the wild-type proteins are available to perform immunopurification 36. Monoclonal antibodies allow the study of the protein interactions with very limited disturbance of the cells other than the lysis. Unfortunately, good quality antibodies that are sufficient for immunopurification of wild-type proteins are often not available and alternative strategies need to be employed. However, this insufficiency of antibiodies could change with projects such as the Human Protein Atlas 37 that generate a suite of high quality monospecific polyclonal antibodies.

The next strategy requires the tagging of specific bait proteins using molecular biology approaches. To date, this has been the primary approach for high-throughput mapping of protein-protein interactions by affinity purification coupled to mass spectrometry 34. The systematic addition of the tag to proteins provides a universal handle that can be used to affinity purify different complexes. Most current methods for purification and identification of protein complexes use an endogenous expression of an affinity-tagged bait 38. Once the complex has been allowed to form in vivo, the cells are lysed, and the complex is purified using the tag present on the bait protein after which the complexes are denatured and separated by 1D or 2D electrophoresis gel. The protein lanes/spots are excised, digested with trypsin, and analyzed by mass spectrometry (MS). Although the schema is generally the same, different tagging approaches have been reported for the purification of protein complexes.

The tandem affinity purification (TAP) method developed by Seraphin is based on two successive affinity chromatography steps 39, 40. Originally it was applied in yeast by in vivo recombination of the TAP-tag. The tag fused to a target protein is composed of protein A having a very high affinity for IgG, a TEV protease cleavage site, and a calmodulin binding peptide which has a high affinity for calmodulin. The TAP-tagged target protein can be purified using protein A affinity resin followed by incubation with TEV protease which releases the target protein. The protein complex is further purified through a second affinity step based on calmodulin binding in the presence of calcium. The complex is then released using EGTA which depletes the calcium ions that are essential for the bait-calmodulin binding. These two different affinity purification steps enhance the specificity of the purification procedure. The initial high-throughput example of this approach was the mapping of protein complexes in yeast by combining TAP purification and MS as reported by Gavin 6. Additional tagging approaches are also used to purify protein complexes by affinity chromatography. Among those, the FLAG tag has been extensively used 41. This small acidic peptide tag (often coded as a triple FLAG) is selectively recognized by a monoclonal antibody. This tag was used by Ho et al. 4 in combination with recombinant-based cloning to tag 725 yeast genes, which were then transfected in yeast. Other tags such as Glutathione S-transferase (GST), His6, biotinylation substrates, and others were used to purify recombinant proteins over-expressed in Escherichia coli 42. However, these tags were not extensively used, probably due to their limited affinity and/or high background levels 39.

Attempts have been done to compare the results from different tagging approaches. For example, Ito et al. 43 compared two large-scale biochemical protein interaction studies performed by the TAP purification method 6, and by the single-step affinity purification with FLAG tag 4 mentioned above. Unfortunately, the tag was not the only difference found in the experiments. The TAP-tagged proteins, for example, were expressed at their natural levels from the endogenous promoters, and the FLAG tag analysis used proteins over-expressed from plasmids which generated very high levels of expression. In both cases, purification was followed by SDS-PAGE and mass spectrometry. The two studies shared 115 targets and showed only about 10% overlap in the proteins recovered.

Tissue, animal models and cell lines

In human, protein interaction mapping needs to take into account i) the diversity of cells, ii) that proteins are not expressed in all cells, and iii) that protein interactions can change between different cell types. Mapping protein interactions in all of cell types in which proteins are expressed is currently too expensive. Although doing immunopurification from cells isolated from human tissue would provide the most relevant results, it remains an expensive proposition. Instead, approaches that rely on expressions in animal models have been proposed. This could be useful for small sets of proteins, but again, it is likely to be too expensive for all proteins. The more realistic options for the first pass mapping of the human interactome are to create stable cell lines or to perform high-throughput transient transfections. Many reports have already used high-throughput transient transfections 44, 45. The major drawback with transient transfection, however, is that the level of protein expression ends up being one or two orders of magnitude above the endogenous level of expression (based on our experience). This can impact the localization of the protein(s) and cause interactions that are not physiologically relevant. As well, the position and the type of tag employed can drastically affect the localization. For example, GFP as a tag has been shown to lead to miss-localization depending on its N or C termini position 46. Unfortunately, the localization, the basal level of expression, and the effects of tagging are unknown for the vast majority of proteins in human.

On the other hand, although stable cell lines are more expensive to create and reportedly slow to generate, the cells that express different levels of the bait protein can be selected. We have not found any reports of the large scale generation of stable cell lines other than papers reporting the promise of lentivirus in rapidly generating stable cell lines 47 and cell microarrays 48. New systems for regulating expression levels based on the modulation of RNA self-cleavage 49 and modulation of translational termination 50 have been proposed to better control the expression levels during transient and stable transfections.

Immunopurification protocols

The majority of the protocols for immunopurification of proteins were developed for soluble proteins. Unfortunately, these proteins only represent a fraction of the human protein interactome. Membrane proteins often do not work with protocols for soluble proteins, and therefore, remain problematic. Protocols based on cross-linking and different stringencies have been developed, but they need to be assessed on a case by case basis. As well, soluble proteins attached to polymeric macromolecules (DNA, microtubules, and microfilaments) have more complex interactions than what is observed by only looking at their interactions while they are free floating. For example, a few novel approaches have been proposed for the study of proteins associated with DNA 51, 52, 53. It is likely that protein complexes anchored on major polymeric macromolecules are poorly represented by current protocols. Hence, the development of novel affinity purification protocols should be a priority in any global effort to map the human interactome.

How will the protein interaction be accessed?

Public repositories of protein interactions are available such as: The MIPS Mammalian Protein-Protein Interaction Database 54 (mips.gsf.de/proj/ppi/), the Human Protein Reference Database 55, 56 (www.hprd.org), the Biomolecular Interaction Network Database (BIND) 57, The Database of Interacting Proteins (DIP) 58, IntAct 59, 60, Molecular INTeraction database (MINT) 61, and BioGRID 62, 63. Furthermore, HUPO, through the HUPO Proteomics Standard initiative, has recently proposed standards required for the submission of interaction datasets 64. These repositories provide access to the interaction datasets and some levels of annotation.

The protein interaction data submission standard and the minimal requirements for mass spectrometric information from peptide identification that have been defined by the HUPO Proteomics Standard Initiative 65, 66, 67 will also facilitate retrospective studies of protein interaction datasets and the development of new tools. For example, greater attention needs to be focused on the development of confidence scoring approaches for protein-protein interaction. We have developed such a scoring algorithm for our recently released large set of human protein interactions 16. As well, Krogan developed, through machine learning algorithm, a scoring scheme for yeast protein interactions 7. The development of new methods to score protein-protein interactions requires that sufficient information be available in public databases to judge the quality of the information and properly compare different results.

It is clear that the bioinformatic analysis of protein interactions is still evolving. New tools for rapidly analyzing protein-protein interactions need to be developed, particularly for most users who are not interested in seeing the whole interactome, but instead are interested in performing an interaction walk starting from one protein of interest. Software, such as Cytoscape 68, Osprey 69, 70, Biomolecular Interaction Network Database (BIND) 57, 71 and others are good graphical interfaces for interaction datasets. The next step needed is the development of tools that incorporate graphical representations of the interactions with the wealth of annotation available from the protein interaction databases, NCBI, and others. Many plug-ins to Cytoscape have been reported to solve these issues 72, 73, 74, 75, 76, 77, 78. We can also foresee that the “balls and edges” approach for displaying interactions is insufficient because it does not represent the dynamic and the structural aspect of the protein interactions. Already, experiments that study the dynamics of protein interactions are underway (http://dynactome.mshri.on.ca/). As well, high-throughput projects are underway for the elucidation of protein structures; but, the incorporation of structural information in models of protein interactions remains a computational challenge even though Morris et al. have reported a plug-in to Cytoscape to link interaction to structure 72.

Novel approaches

LUMIER

Barrios-Rodiles et al. 13 developed the high-throughput LUMIER (for luminescence-based mammalian interactome mapping) to systematically map protein-protein interactions in mammalian cells. This strategy uses Renilla luciferase enzyme (RL) fused to proteins of interest which are then coexpressed with individual FLAG-tagged partners in mammalian cells. This group determined protein interactions by performing an RL enzymatic assay on immunoprecipitates using an antibody against FLAG. This approach has not yet been as widely used as the Y2H approach.

co-localization

Protein interactions require co-localization; therefore, the study of co-localizations can be used to reinforce the first pass interactions map. Computational approaches have been proposed for the prediction of protein localizations, and there are other tools that can be used to reinforce the confidence in the results 79, 80, 81, 82. However, laboratory validation of co-localizations is also important because many proteins can have multiple localizations that are hardly predictable through current models.

Recently, a high-throughput approach to determine the localization of GFP labeled proteins has been described and applied in yeast 83. In this approach, yeast strains were created for 4 156 proteins and microscopic imaging was used to determine the localization of the GFP labeled proteins. 1 839 of these proteins indicated localizations other than in the nucleus or the cytoplasm. Matsuyama et al. 84 studied the localization of 4 431 proteins in the yeast Schizosaccharomyces pombe by cloning its ORFeome and by tagging each ORF with the yellow fluorescent protein. An automated image analysis software that has over 80% accuracy in evaluating the localization of tagged proteins in yeast has been developed 85. Although the large-scale mapping of protein localizations in human cells using these types of approaches have not been reported, it is a plausible strategy to enhance the confidence of protein-protein interactions.

Conclusions

The mapping of the human protein interactome is one of the key challenges in the post genome era. Mapping the human protein interactome will require a coordinated international effort. Current technologies such as Y2H and affinity purification/mass spectrometry can be scaled up to map the human protein interactome. It is likely that multiple standard methods will be developed for the affinity purification/mass spectrometry studies of proteins to take into account different protein functions and localizations. The development of novel technologies, methodologies, algorithms, and software should also be part of such an international effort. Here again, one can look back at the history of sequencing the human genome which succeeded rapidly because of a systematic effort to continuously improve technologies. Even today, these improvements have led to the recent reports of individual genome sequencing 86. Hence, a combination of current technologies and efforts to improve and develop new technologies should also enable us to reach the goal of mapping the human interactome. The mapping of the human interactome would create an invaluable source of information to better understand human biology.