Domain-oriented functional analysis based on expression profiling

Recent progress in genomic sequencing has led to the rapid enrichment of protein sequence databases. Computational biology strives to extract the maximum possible information from these sequences by classifying them according to their homologous relationships. Classical protein families are distinguished by members which exhibit sequence similarity, a feature which can often be used to infer that the sequences are related evolutionarily as well as functionally. One of the first goals of any genome-sequencing project is to broadly classify as many genes and their products as possible into putative functional families.

Proteins are translated from their corresponding mRNAs. Cellular mRNA levels are immensely informative about cell state and the activity of genes, and in most cases, changes in mRNA abundance are positively correlated with changes in protein abundance. Co-regulation of proteins often reflects that these proteins are involved in similar biological processes and have related functions [1]. Therefore changes in the expression patterns of protein families under different experimental conditions can provide clues about regulatory mechanisms, the relationships between broader cellular functions and biochemical pathways, as well as interactions between different protein motifs [2–7].

The advent of microarray technology has made simultaneous analysis of the gene expression profiles of tens of thousands of genes a practical reality [8]. Availability of human genome sequences and massive high throughput data on their expression provides an opportunity to study regulatory relationships between gene families [9, 10]. Incyte's LifeExpress RNA Database is a gene expression database that contains raw and normalized data from hundreds of Incyte microarray experiments. Using these large-scale mRNA expression data, we can infer the functional relationships between protein families by comparing the aggregated expression profiles of the members of each protein family.

Pfam is a functionally annotated database of protein domain families [11]. In this study, each member of Pfam families was mapped to Incyte clones which were presented in the Incyte microarray chips based on the sequence identity. We then generated mRNA expression profiles for these Pfam family members using Incyte's LifeExpress RNA (LE) database (Version 3.0, April 2001 release, Incyte Genomics, Inc.). We analyzed 135 Pfam families, whose family members are well represented on the Incyte microarrays. The expression data for various members of a single Pfam family in one experiment was first converted into binary digits and then summarized as the Family Regulation Ratio. The set of Family Regulation Ratios for a particular Pfam family across multiple experiments is called a Family Regulation Profile. By using Pearson Correlation, we analyzed the similarities between these Family Regulation Profiles to impute functional relationships between the different families. This method was validated by comparing Family Regulation Profiles generated based on two different randomly selected LE data subsets.

This study explores an approach to relate protein families based on quantitative comparison of the family mRNA expression profiles. Analysis of the profiles may be useful to address what protein domain families are co-regulated and possibly how they may interact physically or genetically with one another.

Expression monitoring on a genome-wide scale was first successfully demonstrated in budding yeast [1, 13, 14]. Whole-genome expression profiles provide a rich source of information on protein function, protein-protein interactions, and gene pathways, and have been compared with data sets describing transcription factor binding sites, protein families, protein-protein interactions, and protein abundance [15–22] mainly in budding yeast. In this study we used large scale human genome-wide expression data to study the relationship between the expression patterns of different protein families. We summarized the change in gene expression level for a set of protein family members as a binary code of "regulated" and "unregulated". A two-fold change in expression level was chosen as a cutoff for categorizing a change in expression level as significant. This cutoff is arbitrary, but from our experience with Taqman confirmation (data not shown), a two-fold change in expression level reliably distinguishes real signals from background noise for most expression experiments. The gene regulation was then converted to a percentage for gene families to profile the regulation for each microarray experiment.

Since the mapping from Pfam families to Incyte microarray clones is through the Swissprot database, Pfam family members that are not represented in the Swissprot database cannot be mapped to Incyte clones, and are therefore not included in the Family Regulation Profile analysis. That may be one source of false negatives. Other factors such as sensitivity of the Incyte microarray technology, the arbitrary twofold cutoff, etc. may also contribute to false negative results. In order to reduce these types of errors, we included in our analysis only those Pfam families that map to at least 10 Incyte clones. Due to the fact that the domain hierarchy of Pfam classification is not perfectly reflected in the database, multi-domain proteins may fall into several Pfam members leading to a significant number of false positives. To address this potential source of error, we excluded from pair-wise correlation analysis those Incyte clones shared by both Pfam families in the pair.

To validate relationships that are revealed by microarray expression profiling, one direct way is to split all the experiments randomly into two non-overlapping pools and derive the Pfam relationships independently from each of these pools. The 555 LE experiment were divided into two data sets of S1 and S2 and the corresponding Pfam relationships were denoted as PCC1 and PCC2. The correlation coefficient for PCC1 and PCC2 of 0.76 indicates a high degree of correlation between two data sets. In order to ensure the independence of the data, we also performed another analysis where data were split into two sets based on the tissue type. The Pfam relationships (not shown) derived from these latter data sets are consistent with each other and with those derived from the complete data pool, further corroborating the observed relationships.

It is important to note that many of the protein family relationships (based on co-expression) identified here are supported by functional links. Taking the G protein-coupled receptors (GPCRs) as an example, after agonist action, GPCRs activate G proteins and become phosphorylated by G protein-coupled receptor kinases [23]. This event promotes activation of effector enzymes and ion channels by the activated Gα GTP. This GPCR-mediated regulation of ion channels also depends on the coordination and parallel regulation of protein tyrosine kinase and protein tyrosine phosphatase activities [24]. GPCRs can also interact with the growing family of PDZ domain-containing proteins [25]. All of these relationships are reflected in our study by the strong correlation between 7 transmembrane receptors (rhodopsin family) (PF00001), protein kinases (PF00069), protein tyrosine phosphatases (PF00102), ion transport proteins (PF00520), and PDZ domain proteins (PF00595).

Our study revealed many intriguing links between protein families. For example, the link between homeobox proteins (PF00046) and other protein families. Homeobox proteins are a very important family of transcription factors. According to our analysis, homeobox proteins are involved in the regulation of or are regulated by many protein families such as GPCRs, ion channels, tyrosine phosphatases, nuclear hormone receptors, and protein kinases. For example the cone rod homeobox (Crx) binds and transactivates the rhodopsin promoter [26]. However, it has not been determined whether other rhodopsin GPCRs are also regulated by homeobox proteins. In C. elegans, it was reported that the homeobox gene, lim-6, is required for distinct chemosensory representations (ion sensing), which is related to ion channel/transportation mechanisms. Again, it has not been studied systematically whether homeobox proteins are important regulators of ion channel/transporter. Another example of an intriguing relationship revealed in the present study is between the PH domain, the EGF domain and the immunoglobulin superfamily. Pleckstrin-homology (PH) domains are protein modules of approximately 120 amino acids found in a wide variety of signaling proteins in organisms ranging from yeasts to humans. The EGF domain, which often occurs as multiple tandem repeats, is widely distributed among extracellular proteins involved in adhesion, receptor-ligand interactions, extracellular matrix structure, determination of cell fate, and blood coagulation [27]. Cell adhesion usually activates PI 3-kinase activity. Many PH domain proteins bind with high affinity to specific phosphoinositides such as PI-4,5-P2, PI-3,4-P2 or PI-3,4,5-P3 which are generated by PI-3 kinase [28]. This relationship was reflected by the high correlation coefficient (0.79) between the PH domain and the EGF domain seen in our analysis. Furthermore, another major cell adhesion molecular family is the immunoglobulin superfamily (PF00047). Two immunoglobulins which are particularly important in the cell adhesion cascade are Intercellular Adhesion Molecule-1 (ICAM-1) or CD54 and Vascular Cell Adhesion Molecule-1 (VCAM-1). The co-expression that we observed in our analysis for the EGF-like domains and immunoglobin superfamily suggested that they may cooperate with each other in the cell adhesion process. Links between Pfam families like these suggest novel hypotheses about family interactions that may be further explored.

Integrating genome-wide expression information with protein functional classifications and genetic networks is a huge challenge. Our model is based on the binary idealization of gene expression change to generate a Family Regulation Profile. Although this model is simplified, the abstraction may be useful for conceptualizing the nature of functional relationships between families of protein domains.