Functional Genomics: Gene–Protein–Function Association via Mutants

Probably the most important and enduring contribution of the model yeasts (S. cerevisiae and Schizosaccharomyces pombe) and the scientific communities that study the biology of these organisms has been the connection of genes and proteins with the functions that they provide to cells. As we indicated in 1988, the methods for introducing mutations, at will, into and out of the yeast genome, have made it particularly easy to study not only the biochemical function of gene products, but also the biological consequences of failure of the genes to function. Mutations were produced and introduced into yeast strains by each researcher as needed; this soon came to be seen as rate limiting.

Not long after the publication of the yeast genome sequence, the Saccharomyces community organized a cooperative effort that produced a nearly complete set of deletions of every open reading frame (cf. Winzeler et al. 1999 and Giaever et al. 2002). Each gene was replaced by a drug-resistance gene and marked with synthetic “barcode” sequences. These features made each deletion selectable, facilitating transfer by DNA transformation, and made each deletion distinguishable from all the others so that individual mutants can be followed in screens of the entire library of deletion mutants. Mutations in each yeast gene and many ensembles of mutations have been subjected to diverse biological assays, often leading to increased understanding of the biological roles of many of the genes. The library of deletion mutants and its derivatives over the years has been exploited with great effect in genome-scale experiments; many of the successful methods that we summarize are based on deletion libraries (see Scherens and Goffeau 2004 for a review). A similar deletion library for fission yeast (S. pombe) has recently become available (Kim et al. 2010).

Other comprehensive mutant libraries have been constructed and used to characterize yeast gene functions. The introduction of green fluorescent protein (GFP) fusion technology to visualize protein localization and interactions was used to provide localization information for most proteins of yeast (Ghaemmaghami et al. 2003; Huh et al. 2003). Libraries of fusions to other sequence tags have been constructed to facilitate immunoprecipitation and detection of protein interactions. Characteristically, the yeast community made these strains generally available, facilitating the work of all.

Today the Saccharomyces Genome Database (SGD; http://www.yeastgenome.org/) provides information about every yeast gene based not only on the literature, but also on the systematic study of every Saccharomyces gene about which anything has been learned.

Since 1996 the fraction of the nearly 5800 protein-coding Saccharomyces genes for which a rudimentary understanding of their biological role is known has risen from ∼30% to ∼85%. This fraction is much higher for this yeast than for any other eukaryote. Nearly 1000 yeast genes (i.e., ∼17%) are members of orthologous gene families associated with human disease (Heinicke et al. 2007). For the majority of these genes their mammalian homolog is functional in yeast and complements the yeast deletion mutant. In 1988, although we could point to only a handful of cases of interspecies functional complementation, they have long since become routine (see Dolinski and Botstein 2007 for a review).

Databases and Gene Ontology

Comparison of the yeast genomic sequences with those of other model systems, including the human, led quickly to the realization that both protein amino acid sequences and protein functions have been conserved well enough that annotations of function should frequently, if not always, be transferable from one eukaryotic species to another. Since functional information must ultimately be obtained by experiment, this realization emphasized the advantages of yeast as a model experimental system. Virtually all cellular-level technologies for assessing protein function have remained easier with yeast than with most other organisms. This is especially so for the high-throughput technologies that rapidly developed (see below).

Transfer between species of functional information about homologous proteins via sequence homology required better ways to organize and systematize not only the genome sequences, but also the functional information about genes linked to their amino acid sequences. The Gene Ontology Consortium (GO) (Ashburner et al. 2000; http://www.geneontology.org/) was formed to meet this need. The GO collaborators developed, and continue to maintain, three structured, controlled vocabularies (ontologies) that describe gene products in terms of their associated biological processes, cellular components, and molecular functions in a species-independent manner. The goal was to be able to describe what proteins do in ways that are biologically realistic yet compatible with computational methods. This goal has largely been met. GO annotations are a major feature of all the genomic databases and publications in genomics and are also used in many varieties of computational methods that group genes together according to their biological roles, as we describe below.

One of the early insights provided by GO was the realization of the need to improve on the traditional, but very loose rubric “function,” when thinking about what a gene does for an organism. The GO collaborators decided on three roughly orthogonal categories: “biological process” (e.g., DNA replication), “molecular function” (e.g., DNA-dependent DNA polymerase), and “cellular component” (e.g., nucleus). To illustrate, knowing that a gene is a protein kinase is relatively uninformative for biology until one sees that a deletion of this gene results in a DNA replication defect in yeast and that the protein itself is found in the nucleus. A researcher who comes upon the worm or human ortholog of this kinase will, after having consulted GO, have obtained a very promising hypothesis about what this protein might do for the worm or the human.

It should be no surprise that most of the GO annotations, especially those involving biochemistry and cell biology, derive from experiments with yeast. Increasingly, annotations derive from experiments done directly in other model organisms, which is necessary for biological phenomena not found in yeast. It is no longer uncommon to find SGD annotations for genes and proteins that derive from experiments done on another species that are transferred to yeast. Nevertheless, yeast as a model organism remains a disproportionate contributor to the shared knowledge of eukaryotic biology embodied and organized in the gene ontology (for more on this, see Dolinski and Botstein 2005, 2007).

Gene Expression and Regulatory Networks

The advent of the genomic sequences spurred considerable technology development aimed at comprehensive study of all the genes in the genome simultaneously. The earliest and simplest of these were DNA microarrays containing sequences from every open reading frame, which were used to detect labeled copies of mRNAs (see Brown and Botstein 1999 for an early review). Although this technology could be applied to any organism for which gene sequences were known, it was with yeast that this and other comprehensive functional genomic technologies were pioneered and validated. Even more important than the experimental tractability of yeast was the extensive and reliable knowledge of yeast gene functions already available by 1996 from the literature. When DeRisi et al. (1997) published their landmark study of gene expression during exponential growth and the diauxic shift, they could easily validate their new methods because the behavior of many genes under this growth regime had already been well studied.

Shortly thereafter, computational methods capable of analyzing and visualizing the ∼100,000 individual measurements that compose a data set of this kind began to be introduced (Eisen et al. 1998; Alter et al. 2000). Here again, the well-documented prior knowledge of the behavior of a relatively few genes validated the more general results that could be extracted by the new analytic methods from, for example, studies of the cell division cycle (Spellman et al. 1998), sporulation (Chu et al. 1998), and responses to various stresses (Gasch et al. 2000). Some of these data sets have continued to serve the burgeoning computational biology, systems biology, and bioinformatics communities, providing test beds for an array of increasingly sophisticated analytical methods (cf. Botstein 2010).

For yeast biologists, the main results of the early gene expression studies was the discovery of a vast array of interlocking transcriptional regulatory networks and the transcription factors that control them. Genome-wide expression experiments on yeast validated the wide application of the technology. Gene expression technology became the basis for characterizing the functional genomics of everything from fly development to the definition of tumor subtypes. These data form the basis of a general worldview about system-level integration of transcriptional controls in eukaryotic cells that has awakened the interest of many physical and computational scientists.

The success of gene expression technologies naturally led to the development of a great variety of other genome-scale technologies. Virtually every one of these was developed and validated first with yeast—once again because of the well-documented prior knowledge for the functions of at least a few key genes—and then diffused into more general use. Notable among these methods were those for mapping the binding sites of transcription factors in vivo by chromatin precipitation followed by DNA microarray (ChIP-chip) (Iyer et al. 2001; Lieb et al. 2001) or by direct sequencing of the bound DNA (Chip-seq) (Robertson et al. 2007). Like the simpler gene expression methods, these methods could be used to follow natural processes dynamically and observe, for example, the periodic binding of factors involved in cell cycle progression. Also notable are methods that allow the genome-wide assessment of translation rates (Ingolia et al. 2009) or mRNA stability (Wang et al. 2002), each of which has the property of allowing dynamic assessment of regulatory changes.

Protein Interaction Networks

One of the great promises of functional analysis on the genomic scale was always the possibility of advancing beyond analysis of gene and protein functions one by one. Soon after the publication of the yeast genome sequence, a number of such technologies emerged. Each of these technologies identifies interactions among proteins or genes, which typically are visualized as a network. It is in the arena of understanding these networks of functional relationship that the main opportunities and challenges for understanding cellular biology at the system level lie. Yeast biology has led the way into this arena, and this appears to us to be the path forward for the future.

The first of these studies to appear was based on the two-hybrid method for detecting protein interactions (Fields and Song 1989). A number of large-scale efforts using variants of this approach have produced a large body of data. Although lower throughput versions of the two-hybrid method have been successfully employed to identify a great variety of protein–protein interactions, on a large scale the method has been plagued by large numbers of false-positive and false-negative signals, resulting in disappointingly little overlap between the most prominent of the published genome-scale networks (Uetz et al. 2000; Ito et al. 2001; reviewed by Fields 2009).

A much more reliable approach turned out to be a conceptually straightforward biochemical method that can nevertheless be implemented at high throughput. An epitope tag attached to every open reading frame enables precipitation of protein complexes. The identities of the coprecipitating proteins are determined by mass spectrometry. The method is reliable, and the proportion of false-positive signals is low, although it is clear that interactions below a threshold of affinity are unlikely to be detected (Krogan et al. 2006). Of course, care must be taken to ensure that the tagged protein is fully functional biologically, which can usually be ascertained by testing it for the ability to complement the cognate deletion mutation. Several large protein interaction data sets obtained by this method are in general agreement with each other and with other information. Like the other methods, affinity precipitation can, with some effort, be used to follow protein interactions dynamically.

Gene Interaction Networks

The study of double mutants became a mainstay of genetic analysis of biological function >50 years ago. Mutations that suppress or enhance the phenotypes of other mutations were used to discover much of what is known about molecular biology. Analysis of the phenotypes of double mutants was the basis for the analysis of metabolic, morphogenetic, and signal transduction pathways in all the organisms for which genetic manipulation is convenient. Indeed, the extraordinary facility with which mutations can be combined in yeast was one of the reasons that we were so optimistic about yeast as a model organism in 1988.

“Synthetic lethality” occurs when mutations in two different genes, each not lethal by itself, display a lethal phenotype when combined. Early studies in Drosophila and yeast interpreted synthetic lethality as an indication that the two genes have an essential function in common (Dobzhansky 1946; Sturtevant 1956; Novick et al. 1989). This notion led to the idea that one might screen for genes of similar function on this basis (Bender and Pringle 1991; see Guarente 1993 for an early review). Probably the most successful of the post-genome-sequence technologies in yeast has been the extension of this idea to a genomic scale (Tong et al. 2001; Costanzo et al. 2010).

The method, synthetic gene array (SGA) analysis, begins with the library of marked deletion mutants. Using robots to manipulate the thousands of strains of the library simultaneously, Tong et al. (2001) produced double mutants and determined their ability to grow on petri plates. They devised a clever scheme for crossing a strain carrying a mutation in a query gene to all the viable deletion mutants in the library and recovering haploid double-mutant progeny. This scheme has proved to be extremely robust.

Costanzo et al. (2010) studied 5.4 million double mutants using this SGA approach and produced quantitative genetic interaction profiles for ∼75% of genes of S. cerevisiae. The resulting genetic interaction network (Figure 1) is remarkably coherent when examined either in toto or in sharper focus. Like the clustering and visualization schemes developed for DNA microarrays, analysis and visualization of SGA data provide an informative and intuitive picture of gene interactions.

• Download figure
• Open in new tab
• Download powerpoint

Figure 1

The synthetic gene array network as described by Costanzo et al. (2010) and analyzed and visualized by Baryshnikova et al. (2010). Each point represents a functional gene and each edge a functional connection, and the colored regions indicate subnetworks with similar GO process annotations.

Integrating Co-expression and Protein and Gene Interaction Networks

Much of the information about eukaryotic gene functions and interactions is derived from experiments with yeast, not only because of the advantages that we recognized in 1988, but also because experiments at the genome scale that are feasible with yeast are much more difficult to carry out with other eukaryotes. Nevertheless, each genome-scale method has issues of false-positive and false-negative signals. Although the results (seen as networks of co-expression, protein interaction, and gene interaction) are each reasonably self-consistent and (with the possible exception of two-hybrid results) reproducible, it is the integration of these data that provides the best view of a eukaryotic cell at the system level.

Bioinformatic methods for integrating genome-scale functional data have been in development for about a decade. As was the case with genome-scale experimental methods, yeast has been the obvious test bed for such methods because of the unparalleled fund of knowledge of the organism and because a large number of genome-scale data sets became available for yeast. To illustrate with data readily available on the internet, SGD currently provides a program (SPELL; Hibbs et al. 2007) to answer queries about gene expression on the basis of data from literally hundreds of data sets representing thousands of experimental conditions; Biopixie (http://avis.princeton.edu/pixie/index.php) uses a Bayesian network algorithm (Myers et al. 2005) to integrate the gene expression data with other kinds of genome-scale functional data, notably the results of SGA data, two-hybrid protein interaction data, and direct protein interaction data (see Wang et al. 2002 and Huttenhower et al. 2009 for recent reviews).

In the context of yeast as a model organism, it is worth mentioning that it appears to be straightforward to transfer to other organisms the bioinformatic methods devised for yeast data. Indeed, the major types of analysis—from the most basic analysis and display of gene expression and interaction data to the more complex algorithms used to infer gene and protein interactions—have been successfully applied to data on humans (e.g., Huttenhower et al. 2009) as well as other model organisms, including plants (e.g., Arabidopsis: Lee et al. 2011; Pop et al. 2010).

Leveraging Diversity to Understand Complex Inheritance

We have emphasized the role of yeast as a model for experiments that allow inferences of individual gene functions, gene and protein interactions, and network structures, many of which can be transferred to other eukaryotes, including humans, because of the high degree of evolutionary conservation of genome sequences. However, many inherited traits (in humans as well as other organisms) cannot be attributed to one or a few genes. These cases of complex inheritance have become a major roadblock to progress in understanding many common human diseases. The path forward in human genetics clearly involves the exploitation of the diversity in human genome sequences. Current efforts to use such tools have resulted in successful identification of common polymorphisms that correlated with disease, but unfortunately these polymorphisms collectively seem inadequate to account for the observed heritability (see Lander 2011 for a fuller discussion). It is clear that important elements of the rules for complex inheritance remain to be discovered.

Here again yeast has emerged as a model system. Brem et al. (2002) introduced the idea of using natural variation in yeast genomic sequences to model the inheritance of a simple class of quantitative trait: namely gene expression. They found (see also Yvert et al. 2003 and Ehrenreich et al. 2010) that many of these traits (called eQTL) exhibit complex inheritance. The logic of this approach is simple: if complex inheritance involves complex gene interactions, and if many of the genes of two organisms are highly conserved, then we could expect to learn much about the rules of complex inheritance by studying yeast, where any mechanistic ideas are easily followed up because of the experimental tractability and comprehensive knowledge of gene function that made yeast a leading model organism in the first place. Human geneticists (reviewed in Cheung and Spielman 2009) have already obtained substantial evidence that the methods pioneered in yeast can lead the way to understanding complex patterns of inheritance in humans and other eukaryotes.

Evolution

There are two senses in which yeast is a useful model system for the study of evolution. One of these is our ability to infer, mainly from sequence, the course of evolution of the species and then perform experiments that illuminate these inferences. The other is in the use of yeast as an organism amenable to experimental studies of evolutionary adaptation to selective pressures because of its short doubling time.

Human Disease

In our 1988 Science essay, we anticipated that expression of heterologous proteins in yeast cells would facilitate the connection between structure and function in other organisms. “We think that conservation most strongly validates the use of yeasts as models for the primary deduction of functional and mechanistic aspects of proteins and protein systems shared by eukaryotes” (Botstein and Fink 1988, p. 442). What we did not anticipate was the enormous utility of such heterologous expression for assessment of function of human proteins associated with disease states.

Hereditary nonpolyposis colorectal cancer (HNPCC) is associated with defects in DNA mismatch repair. The history of this discovery about a major common human disease is instructive: an observation about instability of simple DNA repeats in tumors (Aaltonen et al. 1993) led to the deliberate isolation of mutations with this phenotype in yeast by Strand et al. (1993), who explicitly predicted that mutations causing HNPCC in humans might lie in the human orthologs of the three yeast genes (PML1, MLH1, and MSH2) that they identified. By the end of 1993, Fishel et al. (1993) had implicated the human ortholog of MSH2 (now called hMSH2) as a causative gene for HNPCC. Mutations in either hMSH2 or hMLH1 underlie the great majority of HNPCC cases.

The functional characterization of defective human proteins in yeast can reveal aberrant enzyme functions that may not be apparent from assays in humans or from inspection of the protein sequence. An excellent example for this is hMSH2. Analysis in yeast of human MSH2 mutant proteins from colon tumors revealed that some of them had defects expected to be caused by mutations known to affect catalytic function (Gammie et al. 2007). In addition, the analysis in yeast revealed that many of the human variants have lost crucial protein–protein interactions, others have reduced steady-state levels, and some influence the ATPase activity of the mismatch recognition complex. These defects, which appear to contribute to tumor formation, were unanticipated by direct study of humans.

Another benefit of studying human disease gene function in yeast is the potential for remediation of the deficiency responsible for the disease. Many missense mutations lead to decreased affinity for substrate or cofactor [altered Michaelis constant (Km)] that result in a metabolic defect because of insufficient product. It has been suggested that such Km defects could be compensated for by dietary intake of high concentrations of the substrate that could ameliorate the defect (Ames et al. 2002). For several human disorders, such as gyrate atrophy and CBS deficiency, which are caused by recessive mutations in B6-dependent enzymes (ornithine aminotransferase and cystathione synthase respectively), vitamin B-6-responsive patients have been identified (Clayton 2006). Recent work with yeast suggests that analysis of potential human CBS B6-responsive alleles in yeast strains carrying the cognate yeast mutations is extremely informative and can, like the Msh2 analysis, reveal defects unanticipated by other methods.

Prospects for the Future: Much Remains To Be Learned

It is clear that yeast has functioned as a model system that allows inference of individual gene functions, of gene and protein interactions, and of network structures through various kinds of experiments ranging from individual assays to high-throughput genome-scale experiments. We have emphasized how yeast has contributed to our understanding of basic biology in other eukaryotes and of human disease. We have every reason to expect that these contributions will continue and even grow in importance because the technology, especially for DNA sequencing, is continuing to improve rapidly.

It is difficult to predict the future, but we can sketch areas in which progress is likely to occur in the short to medium term. There is every expectation that the information in the yeast genome will continue to be mined. Although most of the open reading frames in the genome have now been annotated, there remains a dearth of information on the functions of noncoding RNAs. Moreover, the function of the many proteins shorter than 100 amino acids remains a challenge.

A fundamental unanswered question is the subcellular distribution of proteins and metabolites in the cell cycle. Currently, we have only a rudimentary cartography—a two- dimensional image of the pathways of metabolism and a catalog of corresponding protein functions. We are slowly obtaining a catalog of the protein–protein interactions, but the locations within a cell of most proteins and small molecules, so critical to their function, remain unknown. But imaging technology and analytical methods for molecular detection are improving rapidly, so we look forward to a three-dimensional understanding of the yeast cell that shows where each of the biological processes occurs.

Saccharomyces has already begun to play a central role in both the pharmaceutical and the industrial arenas. Yeast has a number of advantages for processes that require production on a large scale: the low cost of culture media and a history of efficient fermentation technology. Moreover, the yeast itself, a by-product of the production process, is a valuable commodity for animal feed and thus does not incur additional costs required for disposal.

Despite these advantages, yeast has been eschewed for the production of therapeutic proteins because the complex polysaccharides that decorate proteins secreted from yeast are immunogenic. A vaccine prepared from the hepatitis B antigen could be prepared from yeast because the protein is not secreted and forms particles easily separated from the cytosol. However, strains of Picchia with the entire yeast glycosylation pathway replaced with the one from humans (including the genes encoding the completely foreign proteins of the sialic acid pathway) have been constructed. This advance (Hamilton et al. 2003, 2006) has opened the door for the production of secreted proteins, such as monoclonal antibodies and cytokines.

The advantages of yeast alluded to earlier are even more critical in the production of biofuels, where the product is produced on an immense scale (Lam et al. 2010). Yeast is already used in Brazil and China to produce ethanol for fuel. A goal for the future is the conversion of plant polysaccharides (cellulose and xylans) to ethanol or other alcohols. Strains of yeast have been selected that produce ethanol from glucose at close to the theoretical maximum yield. However, the facts that yeast does not use pentoses as a carbon source, is inhibited by many of the breakdown products of lignin (e.g., furfural), and does not have the pathways to produce branched chain alcohols (and is inhibited by them) present challenges for the future. Given the ease of the genetic manipulation of yeasts, it is likely that these problems will be solved.

Yeast is still the most facile organism for studying the relationship of genotype to phenotype in eukaryotic cells. Much is known about the transmission of the traditional carriers of information, DNA and RNA, during mitosis and meiosis, but little is known about the inheritance of organelles, macromolecules such as polysaccharides and lipids, and the myriad small molecules that populate the cells of living organisms. The recognition that the yeast [PSI]-factor is analogous to mammalian prions (Wickner 1994), is transmitted during mitosis and meiosis, and, like DNA, can transform cells has already added to our understanding of non-Mendelian inheritance.