Protein–protein interactions in the mammalian brain

Abstract

Recent genome-wide high-throughput (HTS) analyses of protein–protein interactions (PPIs) provide molecular-based information to uncover functions of cells and tissues, such as those of the mammalian brain. However, the HTS PPI data contain much false-negatives and false-positives, which should be primarily addressed in experiments. Integrating PPI data sets with other genome-wide data, such as expression profiles and phenotype data sets, provides novel biological insights. Such integration analysis is valuable for addressing the complexity of the mammalian brain. Discovery of novel interactions followed by a detailed analysis is a successful approach to uncover the function of proteins. For example, extensive PPI screens for parkin, a hereditary Parkinson's disease gene, elucidated the function of parkin as an E3 ubiquitin ligase, with localization and activity regulated by contact with its interaction partners, uncovering at least a part of the molecular pathogenesis of Parkinson's disease.

Many proteins are expressed in different mammalian brain cells in a temporally and spatially different manner, adding to the complexity of the brain. Because most proteins work as complexes to regulate biological processes in cells and tissues, discovery of protein–protein interactions (PPIs) and exploration of differences between the PPI networks in many types of brain cells may provide useful information toward understanding the complexity and the pathogenesis of the brain. Here I will first briefly review the current genome-wide PPI research and the false-negatives and false-positives occurring in the high-throughput (HTS) PPI data. Then, I will mention the usefulness of integration analysis of PPI data sets with other data, which may contribute to the understanding of the brain's cellular complexity. Finally, I will raise a working hypothesis derived from such integration analysis for the molecular pathogenesis of Parkinson's disease (PD), a neurodegenerative disease.

Large-scale PPI research and the false interactions

Recent advances in genomic science through large-scale genomic and cDNA sequencing projects have provided both information on and complete physical clone sets of genes in many species. These resources enable systematic screening of PPIs in representative organisms such as yeast, fruit fly and worm, in a high-throughput (HTS) manner (Uetz et al. 2000; Ito et al. 2001; Giot et al. 2003; Li et al. 2004). In these screens, the yeast two-hybrid system was used to obtain binary PPI data. Recently, with the shift of research focus to humans, two human PPI studies have been reported (Rual et al. 2005; Stelzl et al. 2005), although both studies analysed only a part of the entire human PPIs, due to the lack of complete human cDNA sets at this moment. Another large-scale PPI analysis was performed using affinity purification of affinity-tagged proteins followed by mass-spectrometry (TAP-MASS) (Gavin et al. 2002; Ho et al. 2002; Krogan et al. 2006), which enables us to explore the protein complex components rather than binary PPIs. Although the large-scale TAP-MASS approach has only been reported in yeast, the system will soon be applied to representative metazoan cells. Thus, PPI information is rapidly increasing for many species.

The common problems with PPI studies can firstly be seen in a comparison of large-scale yeast PPI studies from two groups, which revealed small overlapping PPIs, indicating several false-negatives and false-positives in the HTS PPI data (von Mering et al. 2002). The small overlap of corresponding PPIs among species (interlogs) has also recently been reported (Gandhi et al. 2006). These false interactions complicate the interpretation of the biological importance of PPIs. The false-negatives, the undetected true interactions, are considered to derive from three origins, improper sample construction, insufficient assay depth and different detection preference (see Table 1 and online Supplemental material, Supplementary text 1). The existence of many false-positives may be a more serious problem with the HTS data sets, where the origins are divided into two class, experimental errors and biological irrelevance (see Table 1 and Supplementary text 1). The way to reduce false interactions seems to be simple to address, but is not easy to realize in actual experiments. There are also several computational approaches to assess the reliability of HTS PPIs (Supplementary text 2).

Table 1.

Origins and treatments of the false interactions

Origins	Treatments
False-negatives
Improper sample construction	Use of more than one assay constructs
N- or C-terminal fusions
Construction errors
Insufficient depth	One by one assay
Pooled samples
Method specificity	Integration of several methods
False-positives
Experimental errors
Non-reproducible interactions	Repetition of assay
Assay specific false-positives	Confirmation using other methods
Biological irrelevance	Endogenous interactions using co-i.p.

Integration of PPIs with other data sets provides novel biological insights

Although current PPI data sets are far from complete, they still form huge PPI networks where several thousands of proteins are connected in a single network. Integration of such PPI data with other data sets provides not only confidence of PPIs (see Supplementary text 2) but also many biological insights (Walhout et al. 2002; Ge et al. 2003). Integrating PPIs with expression profiles revealed for example that genes with similar expression patterns are more likely to associate with each other in yeast and C. elegans (Ge et al. 2001; Teichmann & Babu, 2002; Walhout et al. 2002). Jansen et al. (2002) classified the known yeast protein complexes as permanent or transient based on their expression profiles, where genes in the same permanent complexes have a strong expression correlation. Kemmeren et al. (2002) provided functional annotation of uncharacterized genes in yeast by integration of PPIs and expression profiles. Han et al. (2004) uncovered two types of hub proteins in the PPI network: ‘party’ hubs associate with their partners simultaneously and ‘date’ hubs interact with different a partner in a temporal and spatially different manner.

The integration of phenotype data with PPIs is another approach where the cellular lethality of each protein in yeast is used to evaluate networks; lethal proteins are likely to have more interacting partners than viable ones (Jeong et al. 2001), and lethal proteins tend to interact with each other to form functional units within each cellular function (Saito et al. 2003). Boulton et al. (2002) identified DNA damage response genes in C. elegans by integration analysis of PPIs and phenotype data sets derived from RNAi experiments. In order to predict genetic interactions (Wong et al. 2004), cocomplexed protein pairs (Zhang et al. 2004) and molecular machines involved in C. elegans early embryonesis (Gunsalus et al. 2005), more than three data sets were combined with PPI data.

Recent advances in genomic science have provided systematically annotated information on the genes by using gene ontology (GO), another valuable integration partner with PPI data. The cellular component annotation was, for example, integrated with PPIs in Drosophila to predict functions of uncharacterized genes (Giot et al. 2003). Integrating information on human disease genes revealed attractive suggestions for understanding pathogenic mechanisms and possible therapy targets (Giot et al. 2003; Rual et al. 2005; Stelzl et al. 2005). This analysis uncovered further that many ataxia-causing proteins share interacting partners, some of which have been known to modify neurodegeneration in animal models (Lim et al. 2006).

The integration approach may therefore provide valuable insights into the complexity of the mammalian brain. We are focusing our research on transcription factors because transcription factors play a central role in the complex regulation of gene expression, where many of them work as heterodimers together with their modulating factors. We have constructed expression profiles for 756 transcription factor genes in 11 brain regions of adult mouse using the medium-scale real-time RT-PCR system (Suzuki et al. 2004). We further expanded the expression profiles with more than 1400 transcription factors (Supplementary Table 1). We integrated the expression profiles with the known PPIs of RXRs, RXRB and RXRG, the nuclear receptor members which interact with several nuclear receptors (Mangelsdorf & Evans, 1995; Lane & Bailey, 2005). As shown in Fig. 1, the PPI network generally observed for RXRB and RXRG revealed distinct patterns depending on the brain regions, suggesting a complexity of brain regions due to differently regulated gene sets.

Figure 1. Integration of RXR transcription factor interaction network with the expression profile in mouse brain regions.

Because RXRA is not expressed well in adult brain, the interaction networks for RXRB and RXRG, generally and regionally observed, are shown.

PPI analysis of parkin (PARK2), a hereditary Parkinson's disease gene

Parkinson's disease (PD) is a progressive neurodegenerative disorder characterized by the selective loss of dopaminergic neurons in the substantia nigra (Moore et al. 2005). Although PD is a sporadic disorder in most cases, there are a small number of early onset familial PD patients. Several mutated genes responsible for the familial PD, such as α-synuclein (PCNA), parkin (PARK2), UCHL1, DJ-1, PTEN-induced putative kinase 1 (PINK1) and LRRK2, have been identified by association studies followed by positional cloning (Moore et al. 2005). parkin is the most common of the hereditary PD genes, responsible for autosomal recessive juvenile-onset PD (AR-JP). Parkin is a 465 amino-acid protein that contains a ubiquitin-like (UBL) domain at the amino-terminal region, a linker region (LINKER), and two RING domains and an in-between-RING (IBR) domain at the carboxyl-terminal region (Fig. 2). Because of its domain composition, parkin was verified as an E3 ubiquitin protein ligase involved in the ubiquitin–proteasome protein degradation system, where parkin transfers E2-conjugating ubiquitin to the substrates. Along with this feature, many interaction partners that are involved in the function and targets of parkin have been identified by PPI analysis (see Fig. 2 and Supplementary text 3).

Figure 2. Schematic presentation of the parkin protein and its interaction partners.

Domain composition of parkin is shown in the middle and the interaction partners are shown around parkin. The binding sites of the interaction partners to parkin are roughly shown by lines, except for the target proteins for ubiquitination. The detailed explanation of the interaction partners are described in Supplementary text 3.

Although the discovery of interaction partners followed by detailed analysis has greatly contributed to the understanding of the molecular pathogenesis, the precise mechanism of selective loss of dopaminergic neurons in the substantia nigra still remains unknown. Thus, we systematically analysed expression profiles of the interaction partners, especially in the substantia nigra, using the GNF mouse microarray data (http://symatlas.gnf.org/SymAtlas/). Of all the interaction partners, we found that synaptotagmin XI (Syt11) and b4-tublin (Tubb4), genes of target proteins of parkin, are expressed significantly higher in the substantia nigra than in other tissues (s.d. > 2) (Supplementary Table 2). The high expression of both genes in the substantia nigra is observed even when we compare the expression level among neuronal tissues (Supplementary Table 3). Further, CASK, a protein playing a role of parkin localization at synapses (Fallon et al. 2002), is expressed at significantly lower levels in the substantia nigra than in other neuronal tissues (s.d. < −2.0). The results suggest an attractive working hypothesis where dopaminergic neurons in the substantia nigra are selectively damaged in PD. Low expression of CASK in the dopaminergic neurons may cause a lower concentration of parkin at the synapses than that in other neurons. On the other hand, the concentration of the target proteins, Syt11 and Tubb4, at the synapses of the dopaminergic neurons may be higher than that of the other neurons. Therefore, attenuation of the ubiquitin–proteasome protein degradation system, caused by parkin mutation or for other reasons, may easily result in accumulation of Syt11 and Tubb4, resulting in predominant and selective damage of dopaminergic neurons in the substantia nigra.

Perspectives

At this moment, each researcher is still obligated to experimentally screen interaction partners of their target proteins. However, as I showed above, systematically explored mammalian HTS PPI information is rapidly increasing and its validation, using both experimental and computational approaches, has been greatly improved. Integration of several HTS PPI data sets will soon be sufficient for primary screening of target protein PPIs. Further, PPI networks composed of HTS PPI data sets are integrating all cell and tissue networks. When precisely analysing the cellular complexity of the mammalian brain it is necessary to interpret specific PPI networks for each cell type, where systematic gene expression profiles, such as in situ hybridization data (Sunkin, 2006), should be integrated with PPIs.