2001 346 Schattauer GmbH
What is Bioinformatics? A Proposed Definition and Overview of the Field
N. M. Luscombe, D. Greenbaum, M. Gerstein Department of Molecular Biophysics and Biochemistry Yale University, New Haven, USA
Summary
Background: The recent flood of data from genome sequences and functional genomics has given rise to new field, bioinformatics, which combines elements of biology and computer science. Objectives: Here we propose a definition for this new field and review some of the research that is being pursued, particularly in relation to transcriptional regulatory systems. Methods: Our definition is as follows: Bioinformatics is conceptualizing biology in terms of macromolecules (in the sense of physical-chemistry) and then applying informatics techniques (derived from disciplines such as applied maths, computer science, and statistics) to understand and organize the information associated with these molecules, on a large-scale. Results and Conclusions: Analyses in bioinformatics predominantly focus on three types of large datasets available in molecular biology: macromolecular structures, genome sequences, and the results of functional genomics experiments (eg expression data). Additional information includes the text of scientific papers and relationship data from metabolic pathways, taxonomy trees, and protein-protein interaction networks. Bioinformatics employs a wide range of computational techniques including sequence and structural alignment, database design and data mining, macromolecular geometry, phylogenetic tree construction, prediction of protein structure and function, gene finding, and expression data clustering. The emphasis is on approaches integrating a variety of computational methods and heterogeneous data sources. Finally, bioinformatics is a practical discipline. We survey some representative applications, such as finding homologues, designing drugs, and performing large-scale censuses. Additional information pertinent to the review is available over the web at http://bioinfo.mbb.yale.edu/what-is-it.
1. Introduction
Biological data are being produced at a phenomenal rate [1]. For example as of April 2001, the GenBank repository of nucleic acid sequences contained 11,546,000 entries [2] and the SWISSPROT database of protein sequences contained 95,320 [3]. On average, these databases are doubling in size every 15 months [2]. In addition, since the publication of the H. influenzae genome [4], complete sequences for nearly 300 organisms have been released, ranging from 450 genes to over 100,000. Add to this the data from the myriad of related projects that study gene expression, determine the protein structures encoded by the genes, and detail how these products interact with one another, and we can begin to imagine the enormous quantity and variety of information that is being produced. As a result of this surge in data, computers have become indispensable to biological research. Such an approach is ideal because of the ease with which computers can handle large quantities of data and probe the complex dynamics observed in nature. Bioinformatics, the subject of the current review, is often defined as the application of computational techniques to understand and organise the information associated with biological macromolecules. Fig. 1 shows that the number of papers
related to this new field has been surging, and now comprise almost 2% of the annual total of papers in PubMed. This unexpected union between the two subjects is attributed to the fact that life itself is an information technology; an organisms physiology is largely determined by its genes, which at its most basic can be viewed as digital information.At the same time, there have been major advances in the technologies that supply the initial data; Anthony Kervalage of Celera recently cited that an experimental laboratory can produce over 100 gigabytes of data a day with ease [5]. This incredible processing power has been matched by developments in computer technology; the most important areas of improvements have been in the CPU, disk storage and Internet, allowing faster computations, better data storage and revolutionalised the methods for accessing and exchanging data.
1.1 Aims of Bioinformatics
In general, the aims of bioinformatics are three-fold. First, at its simplest bioinformatics organises data in a way that allows researchers to access existing information and to submit new entries as they are produced, e.g. the Protein Data Bank for 3D macromolecular structures [6, 7]. While data-curation is an essential task, the information stored in these databases is essentially useless until analysed. Thus the purpose of bioinformatics extends much further. The second aim is to develop tools and resources that aid in the analysis of data. For example, having sequenced a particular protein, it is of interest to compare it with previously characterised sequences.
Keywords
Bioinformatics, Genomics, Introduction, Transcription Regulation Method Inform Med 2001; 40: 34658 Method Inform Med 4/2001
Updated version of an invited review paper that appeared in Haux, R., Kulikowski, C. (eds.) (2001). IMIA Yearbook of Medical Informatics 2001: Digital Libraries and Medicine, pp. 8399. Stuttgart: Schattauer.
347 What is Bioinformatics?
Fig. 1 Plot showing the growth of scientific publications in bioinformatics between 1973 and 2000. The histogram bars (left vertical axis) counts the total number of scientific articles relating to bioinformatics, and the black line (right vertical axis) gives the percentage of the annual total of articles relating to bioinformatics. The data are taken from PubMed.
This needs more than just a simple textbased search, and programs such as FASTA [8] and PSI-BLAST [9] must consider what constitutes a biologically significant match. Development of such resources dictates expertise in computational theory, as well as a thorough understanding of biology. The third aim is to use these tools to analyse the data and interpret the results in a biologically meaningful manner. Traditionally, biological studies examined individual systems in detail, and frequently compared them with a few that are related. In bioinformatics, we can now conduct global analyses of all the available data with the aim of uncovering common principles that apply across many systems and highlight novel features. In this review, we provide a systematic definition of bioinformatics as shown in Box 1. We focus on the first and third aims just described, with particular reference to the keywords: information, informatics, organisation, understanding, large-scale and practical applications. Specifically, we discuss the range of data that are currently being examined, the databases into which they are organised, the types of analyses that are being conducted using transcription regulatory systems as an example, and
finally some of the major practical applications of bioinformatics.
2. the INFORMATION associated with these Molecules
Table 1 lists the types of data that are analysed in bioinformatics and the range of topics that we consider to fall within the field. Here we take a broad view and include subjects that may not normally be listed. We also give approximate values describing the sizes of data being discussed. We start with an overview of the sources of information. Most bioinformatics analyses focus on three primary sources of data: DNA or protein sequences, macromolecular structures and the results of functional genomics experiments. Raw DNA sequences are strings of the four base-letters comprising genes, each typically 1,000 bases long. The GenBank [2] repository of nucleic acid sequences currently holds a total of 12.5 billion bases in 11.5 million entries (all database figures as of April
2001). At the next level are protein sequences comprising strings of 20 amino acidletters. At present there are about 400,000 known protein sequences [3], with a typical bacterial protein containing approximately 300 amino acids. Macromolecular structural data represents a more complex form of information. There are currently 15,000 entries in the Protein Data Bank, PDB [6, 7], containing atomic structures of proteins, DNA and RNA solved by x-ray crystallography and NMR. A typical PDB file for a medium-sized protein contains the xyz-coordinates of approximately 2,000 atoms. Scientific euphoria has recently centred on whole genome sequencing. As with the raw DNA sequences, genomes consist of strings of base-letters, ranging from 1.6 million bases in Haemophilus influenzae [10] to 3 billion in humans [11, 12]. The Entrez database [13] currently has complete sequences for nearly 300 archaeal, bacterial and eukaryotic organisms. In addition to producing the raw nucleotide sequence, a lot of work is involved in processing this data. An important aspect of complete genomes is the distinction between coding regions and non-coding regions -junk repetitive sequences making up the bulk of base sequences especially in eukaryotes. Within the coding regions, genes are annotated with their translated protein sequence, and often with their cellular function.
Bioinformatics a Definition1
(Molecular) bio informatics: bioinformatics is conceptualising biology in terms of molecules (in the sense of Physical chemistry) and applying informatics techniques (derived from disciplines such as applied maths, computer science and statistics) to understand and organise the information associated with these molecules, on a large scale. In short, bioinformatics is a management information system for molecular biology and has many practical applications.
1As
submitted to the Oxford English Dictionary.
Method Inform Med 4/2001
348 Luscombe, Greenbaum, Gerstein
Table 1 Sources of data used in bioinformatics, the quantity of each type of data that is currently (April 2001) available, and bioinformatics subject areas that utilize this data.
quences. While more biological information can be derived from a single structure than a protein sequence, the lack of depth in the latter is compensated by analysing larger quantities of data.
3. ORGANISE the Information on a LARGE SCALE
3.1 Redundancy and Multiplicity of Data
A concept that underpins most research methods in bioinformatics is that much of the data can be grouped together based on biologically meaningful similarities. For example, sequence segments are often repeated at different positions of genomic DNA [27]. Genes can be clustered into those with particular functions (eg enzymatic actions) or according to the metabolic pathway to which they belong [28], although here, single genes may actually possess several functions [29]. Going further, distinct proteins frequently have comparable sequences organisms often have multiple copies of a particular gene through duplication and different species have equivalent or similar proteins that were inherited when they diverged from each other in evolution. At a structural level, we predict there to be a finite number of different tertiary structures estimates range between 1,000 and 10,000 folds [30, 31] and proteins adopt equivalent structures even when they differ greatly in sequence [32]. As a result, although the number of structures in the PDB has increased exponentially, the rate of discovery of novel folds has actually decreased. There are common terms to describe the relationship between pairs of proteins or the genes from which they are derived: analogous proteins have related folds, but unrelated sequences, while homologous proteins are both sequentially and structurally similar. The two categories can sometimes be difficult to distinguish especially if the relationship between the two proteins is remote [33, 34]. Among homologues, it is useful to distinguish between orthologues, proteins in different species that have evolv-
More recent sources of data have been from functional genomics experiments, of which the most common are gene expression studies. We can now determine expression levels of almost every gene in a given cell on a whole-genome level, however there is currently no central repository for this data and public availability is limited. These experiments measure the amount of mRNA that is produced by the cell [14-18] under different environmental conditions, different stages of the cell cycle and different cell types in multi-cellular organisms. Much of the effort has so far focused on the yeast [19-24] and human genomes [25, 26]. One of the largest dataset for yeast has made approximately 20 time-point measurements for 6,000 genes [19]. However, there is potential for much greater quanMethod Inform Med 4/2001
tities of data when experiments are conducted for larger organisms and at more time-points. Further genomic-scale data include biochemical information on metabolic pathways, regulatory networks, proteinprotein interaction data from two-hybrid experiments, and systematic knockouts of individual genes to test the viability of an organism. What is apparent from this list is the diversity in the size and complexity of different datasets. There are invariably more sequence-based data than others because of the relative ease with which they can be produced.This is partly related to the greater complexity and information-content of individual structures or gene expression experiments compared to individual se-
349 What is Bioinformatics?
ed from a common ancestral gene, and paralogues, proteins that are related by gene duplication within a genome [35]. Normally, orthologues retain the same function while paralogues evolve distinct, but related functions [36]. An important concept that arises from these observations is that of a finite parts list for different organisms [37-39]: an inventory of proteins contained within an organism, arranged according to different properties such as gene sequence, protein fold or function. Taking protein folds as an example, we mentioned that with a few exceptions, the tertiary structures of proteins adopt one of a limited repertoire of folds. As the number of different fold families is considerably smaller than the number of genes, categorising the proteins by fold provides a substantial simplification of the contents of a genome. Similar simplifications can be provided by other attributes such as protein function. As such, we expect this notion of a finite parts list to become increasingly common in future genomic analyses. Clearly, an essential aspect of managing this large volume of data lies in developing methods for assessing similarities between different biomolecules and identifying those that are related. There are well-documented classifications for all of the main types of data we described earlier. Although detailed descriptions of these classification systems are beyond the scope of the current review, they are of great importance as they ease comparisons between genomes and their products. Links to the major databases are available from our supplementary website.
straightforward to access and crossreference these sources of information because of differences in nomenclature and file formats. At a basic level, this problem is frequently addressed by providing external links to other databases. For example in PDBsum, web-pages for individual structures direct the user towards corresponding entries in the PDB, NDB, CATH, SCOP and SWISS-PROT databases. At a more advanced level, there have been efforts to integrate access across several data sources. One is the Sequence Retrieval System, SRS [41], which allows flat-file databases to be indexed to each other; this allows the user to retrieve, link and access entries from nucleic acid, protein sequence, protein motif, protein structure and bibliographic databases. Another is the Entrez facility [42], which provides similar gateways to DNA and protein sequences, genome mapping data, 3D macromolecular structures and the PubMed bibliographic database [43].A search for a particular gene in either database will allow smooth transitions to the genome it comes from, the protein sequence it encodes, its structure, bibliographic reference and equivalent entries for all related genes. In our own group, we have developed the SPINE [44] and PartsList [39] web resources; these databases integrate many types of experimental data and organise them using the concept of the finite parts list we described above.
4. UNDERSTAND and Organise the Information
Having examined the data, we can discuss the types of analyses that are conducted.As shown in Table 1, the broad subject areas in bioinformatics can be separated according to the type of information that is used. For raw DNA sequences, investigations involve separating coding and non-coding regions, and identification of introns, exons and promoter regions for annotating genomic DNA [45, 46]. For protein sequences, analyses include developing algorithms for sequence comparisons [47], methods for
3.2 Data Integration
The most profitable research in bioinformatics often results from integrating multiple sources of data [40]. For instance, the 3D coordinates of a protein are more useful if combined with data about the proteins function, occurrence in different genomes, and interactions with other molecules. In this way, individual pieces of information are put in context with respect to other data. Unfortunately, it is not always
producing multiple sequence alignments [48], and searching for functional domains from conserved sequence motifs in such alignments. Investigations of structural data include prediction of secondary and tertiary protein structures, producing methods for 3D structural alignments [49, 50], examining protein geometries using distance and angular measurements, calculations of surface and volume shapes and analysis of protein interactions with other subunits, DNA, RNA and smaller molecules. These studies have lead to molecular simulation topics in which structural data are used to calculate the energetics involved in stabilising macromolecular structures, simulating movements within macromolecules, and computing the energies involved in molecular docking. The increasing availability of annotated genomic sequences has resulted in the introduction of computational genomics and proteomics large-scale analyses of complete genomes and the proteins that they encode. Research includes characterisation of protein content and metabolic pathways between different genomes, identification of interacting proteins, assignment and prediction of gene products, and large-scale analyses of gene expression levels. Some of these research topics will be demonstrated in our example analysis of transcription regulatory systems. Other subject areas we have included in Table 1 are: development of digital libraries for automated bibliographical searches, knowledge bases of biological information from the literature, DNA analysis methods in forensics, prediction of nucleic acid structures, metabolic pathway simulations, and linkage analysis linking specific genes to different disease traits. In addition to finding relationships between different proteins, much of bioinformatics involves the analysis of one type of data to infer and understand the observations for another type of data. An example is the use of sequence and structural data to predict the secondary and tertiary structures of new protein sequences [51]. These methods, especially the former, are often based on statistical rules derived from structures, such as the propensity for certain amino acid sequences to produce
Method Inform Med 4/2001
350 Luscombe, Greenbaum, Gerstein
Paradigm shifts during the past couple of decades have taken much of biology away from the laboratory bench and have allowed the integration of other scientific disciplines, specifically computing. The result is an expansion of biological research in breadth and depth. The vertical axis demonstrates how bioinformatics can aid rational drug design with minimal work in the wet lab. Starting with a single gene sequence, we can determine with strong certainty, the protein sequence. From there, we can determine the structure using structure prediction techniques. With geometry calculations, we can further resolve the proteins surface and through molecular simulation determine the force fields surrounding the molecule. Finally docking algorithms can provide predictions of the ligands that will bind on the protein surface, thus paving the way for the design of a drug specific to that molecule. The horizontal axis shows how the influx of biological data and advances in computer technology have broadened the scope of biology. Initially with a pair of proteins, we can make comparisons between the between sequences and structures of evolutionary related proteins. With more data, algorithms for multiple alignments of several proteins become necessary. Using multiple sequences, we can also create phylogenetic trees to trace the evolutionary development of the proteins in question. Finally, with the deluge of data we currently face, we need to construct large databases to store, view and deconstruct the information. Alignments now become more complex, requiring sophisticated scoring schemes and there is enough data to compile a genome census a genomic equivalent of a population census providing comprehensive statistical accounting of protein features in genomes. Fig. 2 Organizing and understanding biological data
different secondary structural elements. Another example is the use of structural data to understand a proteins function; here studies have investigated the relationship different protein folds and their functions [52, 53] and analysed similarities between different binding sites in the absence of homology [54]. Combined with similarity measurements, these studies provide us with an understanding of how much biological information can be accurately
Method Inform Med 4/2001
transferred between homologous proteins [55].
4.1 The Bioinformatics Spectrum
Fig. 2 summarises the main points we raised in our discussions of organising and understanding biological data the development of bioinformatics techniques has allowed an expansion of biological
analysis in two dimension, depth and breadth. The first is represented by the vertical axis in the figure and outlines a possible approach to the rational drug design process. The aim is to take a single gene and follow through an analysis that maximises our understanding of the protein it encodes. Starting with a gene sequence, we can determine the protein sequence with strong certainty. From there, prediction algorithms can be used to calcu-
351 What is Bioinformatics?
late the structure adopted by the protein. Geometry calculations can define the shape of the proteins surface and molecular simulations can determine the force fields surrounding the molecule. Finally, using docking algorithms, one could identify or design ligands that may bind the protein, paving the way for designing a drug that specifically alters the proteins function. In practise, the intermediate steps are still difficult to achieve accurately, and they are best combined with experimental methods to obtain some of the data, for example characterising the structure of the protein of interest. The aim of the second dimension, the breadth in biological analysis, is to compare a gene or gene product with others. Initially, simple algorithms can be used to compare the sequences and structures of a pair of related proteins. With a larger number of proteins, improved algorithms can be used to produce multiple alignments, and extract sequence patterns or structural templates that define a family of proteins. Using this data, it is also possible to construct phylogenetic trees to trace the evolutionary path of proteins. Finally, with even more data, the information must be stored in large-scale databases. Comparisons become more complex, requiring multiple scoring schemes, and we are able to conduct genomic scale censuses that provide comprehensive statistical accounts of protein features, such as the abundance of particular structures or functions in different genomes. It also allows us to build phylogenetic trees that trace the evolution of whole organisms.
son methods such as text search and onedimensional alignment algorithms. Motif and pattern identification for multiple sequences depend on machine learning, clustering and data-mining techniques. 3D structural analysis techniques include Euclidean geometry calculations combined with basic application of physical chemistry, graphical representations of surfaces and volumes, and structural comparison and 3D matching methods. For molecular simulations, Newtonian mechanics, quantum mechanics, molecular mechanics and electrostatic calculations are applied. In many of these areas, the computational methods must be combined with good statistical analyses in order to provide an objective measure for the significance of the results.
6.1 Structural Studies
As of April 2001, there were 379 structures of protein-DNA complexes in the PDB. Analyses of these structures have provided valuable insight into the stereochemical principles of binding, including how particular base sequences are recognized and how the DNA structure is quite often modified on binding. A structural taxonomy of DNA-binding proteins, similar to that presented in SCOP and CATH, was first proposed by Harrison [56] and periodically updated to accommodate new structures as they are solved [57]. The classification consists of a two-tier system: the first level collects proteins into eight groups that share gross structural features for DNA-binding, and the second comprises 54 families of proteins that are structurally homologous to each other. Assembly of such a system simplifies the comparison of different binding methods; it highlights the diversity of protein-DNA complex geometries found in nature, but also underlines the importance of interactions between -helices and the DNA major groove, the main mode of binding in over half the protein families. While the number of structures represented in the PDB does not necessarily reflect the relative importance of the different proteins in the cell, it is clear that helix-turn-helix, zinc-coordinating and leucine zipper motifs are used repeatedly.These provide compact frameworks to present the -helix on the surfaces of structurally diverse proteins. At a gross level, it is possible to highlight the differences between transcription factor domains that just bind DNA and those involved in catalysis [58]. Although there are exceptions, the former typically approach the DNA from a single face and slot into the grooves to interact with base edges. The latter commonly envelope the substrate, using complex networks of secondary structures and loops. Focusing on proteins with -helices, the structures show many variations, both in amino acid sequences and detailed geometry. They have clearly evolved independently in accordance with the requirements of the context in which they are found. While achieving a close fit between the
Method Inform Med 4/2001
6. Transcription Regulation a Case Study in Bioinformatics
DNA-binding proteins have a central role in all aspects of genetic activity within an organism, participating in processes such as transcription, packaging, rearrangement, replication and repair. In this section, we focus on the studies that have contributed to our understanding of transcription regulation in different organisms. Through this example, we demonstrate how bioinformatics has been used to increase our knowledge of biological systems and also illustrate the practical applications of the different subject areas that were briefly outlined earlier. We start by considering structural analyses of how DNA-binding proteins recognise particular base sequences. Later, we review several genomic studies that have characterised the nature of transcription factors in different organisms, and the methods that have been used to identify regulatory binding sites in the upstream regions. Finally, we provide an overview of gene expression analyses that have been recently conducted and suggest future uses of transcription regulatory analyses to rationalise the observations made in gene expression experiments. All the results that we describe have been found through computational studies.
5. applying INFORMATICS TECHNIQUES
The distinct subject areas we mention require different types of informatics techniques. Briefly, for data organisation, the first biological databases were simple flat files. However with the increasing amount of information, relational database methods with Web-page interfaces have become increasingly popular. In sequence analysis, techniques include string compari-
352 Luscombe, Greenbaum, Gerstein
-helix and major groove, there is enough flexibility to allow both the protein and DNA to adopt distinct conformations. However, several studies that analysed the binding geometries of -helices demonstrated that most adopt fairly uniform conformations regardless of protein family. They are commonly inserted in the major groove sideways, with their lengthwise axis roughly parallel to the slope outlined by the DNA backbone. Most start with the N-terminus in the groove and extend out, completing two to three turns within contacting distance of the nucleic acid [59, 60]. Given the similar binding orientations, it is surprising to find that the interactions between each amino acid position along the -helices and nucleotides on the DNA vary considerably between different protein families. However, by classifying the amino acids according to the sizes of their side chains, we are able to rationalise the different interactions patterns. The rules of interactions are based on the simple premise that for a given residue position on -helices in similar conformations, small amino acids interact with nucleotides that are close in distance and large amino acids with those that are further [60, 61]. Equi-valent studies for binding by other structural motifs, like -hairpins, have also been conducted [62]. When considering these interactions, it is important to remember that different regions of the protein surface also provide interfaces with the DNA. This brings us to look at the atomic level interactions between individual amino acid-base pairs. Such analyses are based on the premise that a significant proportion of specific DNA-binding could be rationalised by a universal code of recognition between amino acids and bases, ie whether certain protein residues preferably interact with particular nucleotides regardless of the type of protein-DNA complex [63]. Studies have considered hydrogen bonds, van der Waals contacts and water-mediated bonds [64-66]. Results showed that about 2/3 of all interactions are with the DNA backbone and that their main role is one of sequence-independent stabilisation. In contrast, interactions with bases display some strong preferences, including the
Method Inform Med 4/2001
interactions of arginine or lysine with guanine, asparagine or glutamine with adenine and threonine with thymine. Such preferences were explained through examination of the stereochemistry of the amino acid side chains and base edges. Also highlighted were more complex types of interactions where single amino acids contact more than one base-step simultaneously, thus recognising a short DNA sequence. These results suggested that universal specificity, one that is observed across all protein-DNA complexes, indeed exists. However, many interactions that are normally considered to be non-specific, such as those with the DNA backbone, can also provide specificity depending on the context in which they are made. Armed with an understanding of protein structure, DNA-binding motifs and side chain stereochemistry, a major application has been the prediction of binding either by proteins known to contain a particular motif, or those with structures solved in the uncomplexed form. Most common are predictions for -helix-major groove interactions given the amino acid sequence, what DNA sequence would it recognise [61, 67]. In a different approach, molecular simulation techniques have been used to dock whole proteins and DNAs on the basis of force-field calculations around the two molecules [68, 69]. The reason that both methods have been met with limited success is because even for apparently simple cases like helix-binding, there are many other factors that must be considered. Comparisons between bound and unbound nucleic acid structures show that DNA-bending is a common feature of complexes formed with transcription factors [58, 70].This and other factors such as electrostatic and cationmediated interactions assist indirect recognition of the nucleotide sequence, although they are not well understood yet. Therefore, it is now clear that detailed rules for specific DNA-binding will be family specific, but with underlying trends such as the arginine-guanine interactions.
6.2 Genomic Studies
Due to the wealth of biochemical data that are available, genomic studies in bioinformatics have concentrated on model organisms, and the analysis of regulatory systems has been no exception. Identification of transcription factors in genomes invariably depends on similarity search strategies, which assume a functional and evolutionary relationship between homologous proteins. In E. coli, studies have so far estimated a total of 300 to 500 transcription regulators [71] and PEDANT [72], a database of automatically assigned gene functions, shows that typically 2-3% of prokaryotic and 6-7% of eukaryotic genomes comprise DNA-binding proteins.As assignments were only complete for 40-60% of genomes as of August 2000, these figures most likely underestimate the actual number. Nonetheless, they already represent a large quantity of proteins and it is clear that there are more transcription regulators in eukaryotes than other species. This is unsurprising, considering the organisms have developed a relatively sophisticated transcription mechanism. From the conclusions of the structural studies, the best strategy for characterising DNA-binding of the putative transcription factors in each genome is to group them by homology and to analyse the individual families. Such classifications are provided in the secondary sequence databases described earlier and also those that specialise in regulatory proteins such as RegulonDB [73] and TRANSFAC [74]. Of even greater use is the provision of structural assignments to the proteins; given a transcription factor, it is helpful to know the structural motif that it uses for binding, therefore providing us with a better understanding of how it recognises the target sequence. Structural genomics through bioinformatics assigns structures to the protein products of genomes by demonstrating similarity to proteins of known structure [75]. These studies have shown that prokaryotic transcription factors most frequently contain helix-turnhelix motifs [71, 76] and eukaryotic factors contain homeodomain type helix-turnhelix, zinc finger or leucine zipper motifs.
353 What is Bioinformatics?
From the protein classifications in each genome, it is clear that different types of regulatory proteins differ in abundance and families significantly differ in size. A study by Huynen and van Nimwegen [77] has shown that members of a single family have similar but functions, as the requirements of this function vary over time, so does the presence of each gene family in the genome. Most recently, using a combination of sequence and structural data, we examined the conservation of amino acid sequences between related DNA-binding proteins, and the effect that mutations have on DNA sequence recognition. The structural families described above were expanded to include proteins that are related by sequence similarity, but whose structures remain unsolved. Again, members of the same family are homologous, and probably derive from a common ancestor. Amino acid conservations were calculated for the multiple sequence alignments of each family [78]. Generally, alignment positions that interact with the DNA are better conserved than the rest of the protein surface, although the detailed patterns of conservation are quite complex. Residues that contact the DNA backbone are highly conserved in all protein families, providing a set of stabilising interactions that are common to all homologous proteins. The conservation of alignment positions that contact bases, and recognise the DNA sequence, are more complex and could be rationalised by defining a three-class model for DNA-binding. First, protein families that bind non-specifically usually contain several conserved base-contacting residues; without exception, interactions are made in the minor groove where there is little discrimination between base types. The contacts are commonly used to stabilise deformations in the nucleic acid structure, particularly in widening the DNA minor groove. The second class comprise families whose members all target the same nucleotide sequence; here, base-contacting positions are absolutely or highly conserved allowing related proteins to target the same sequence. The third, and most interesting, class comprises families in which binding is also
specific but different members bind distinct base sequences. Here protein residues undergo frequent mutations, and family members can be divided into subfamilies according to the amino acid sequences at base-contacting positions; those in the same subfamily are predicted to bind the same DNA sequence and those of different subfamilies to bind distinct sequences. On the whole, the subfamilies corresponded well with the proteins functions and members of the same subfamilies were found to regulate similar transcription pathways. The combined analysis of sequence and structural data described by this study provided an insight into how homologous DNA-binding scaffolds achieve different specificities by altering their amino acid sequences. In doing so, proteins evolved distinct functions, therefore allowing structurally related transcription factors to regulate expression of different genes. Therefore, the relative abundance of transcription regulatory families in a genome depends, not only on the importance of a particular protein function, but also in the adaptability of the DNA-binding motifs to recognise distinct nucleotide sequences. This, in turn, appears to be best accommodated by simple binding motifs, such as the zinc fingers. Given the knowledge of the transcription regulators that are contained in each organism, and an understanding of how they recognise DNA sequences, it is of interest to search for their potential binding sites within genome sequences [79]. For prokaryotes, most analyses have involved compiling data on experimentally known binding sites for particular proteins and building a consensus sequence that incorporates any variations in nucleotides. Additional sites are found by conducting word-matching searches over the entire genome and scoring candidate sites by similarity [80-83]. Unsurprisingly, most of the predicted sites are found in non-coding regions of the DNA [80] and the results of the studies are often presented in databases such as RegulonDB [73]. The consensus search approach is often complemented by comparative genomic studies searching upstream regions of orthologous genes in closely related organisms. Through such an
approach, it was found that at least 27% of known E. coli DNA-regulatory motifs are conserved in one or more distantly related bacteria [84]. The detection of regulatory sites in eukaryotes poses a more difficult problem because consensus sequences tend to be much shorter, variable, and dispersed over very large distances. However, initial studies in S. cerevisiae provided an interesting observation for the GATA protein in nitrogen metabolism regulation. While the 5 base-pair GATA consensus sequence is found almost everywhere in the genome, a single isolated binding site is insufficient to exert the regulatory function [85]. Therefore specificity of GATA activity comes from the repetition of the consensus sequence within the upstream regions of controlled genes in multiple copies. An initial study has used this observation to predict new regulatory sites by searching for overrepresented oligonucleotides in non-coding regions of yeast and worm genomes [86, 87]. Having detected the regulatory binding sites, there is the problem of defining the genes that are actually regulated, commonly termed regulons. Generally, binding sites are assumed to be located directly upstream of the regulons; however there are different problems associated with this assumption depending on the organism. For prokaryotes, it is complicated by the presence of operons; it is difficult to locate the regulated gene within an operon since it can lie several genes downstream of the regulatory sequence. It is often difficult to predict the organisation of operons [88], especially to define the gene that is found at the head, and there is often a lack of long-range conservation in gene order between related organisms [89]. The problem in eukaryotes is even more severe; regulatory sites often act in both directions, binding sites are usually distant from regulons because of large intergenic regions, and transcription regulation is usually a result of combined action by multiple transcription factors in a combinatorial manner. Despite these problems, these studies have succeeded in confirming the transcription regulatory pathways of well-characterized systems such as the heat shock response
Method Inform Med 4/2001
354 Luscombe, Greenbaum, Gerstein
system [83]. In addition, it is feasible to experimentally verify any predictions, most notably using gene expression data.
6.3 Gene Expression Studies
Many expression studies have so far focused on devising methods to cluster genes by similarities in expression profiles. This is in order to determine the proteins that are expressed together under different cellular conditions. Briefly, the most common methods are hierarchical clustering, self-organising maps, and K-means clustering. Hierarchical methods originally derived from algorithms to construct phylogenetic trees, and group genes in a bottomup fashion; genes with the most similar expression profiles are clustered first, and those with more diverse profiles are included iteratively [90-92]. In contrast, the self-organising map [93, 94] and K-means methods [95, 96] employ a top-down approach in which the user pre-defines the number of clusters for the dataset. The clusters are initially assigned randomly, and the genes are regrouped iteratively until they are optimally clustered. Given these methods, it is of interest to relate the expression data to other attributes such as structure, function and subcellular localisation of each gene product. Mapping these properties provides an insight into the characteristics of proteins that are expressed together, and also suggest some interesting conclusions about the overall biochemistry of the cell. In yeast, shorter proteins tend to be more highly expressed than longer proteins, probably because of the relative ease with which they are produced [97]. Looking at the amino acid content, highly expressed genes are generally enriched in alanine and glycine, and depleted in asparagine; these are thought to reflect the requirements of amino acid usage in the organism, where synthesis of alanine and glycine are energetically less expensive than asparagine.Turning to protein structure, expression levels of the TIM barrel and NTP hydrolase folds are highest, while those for the leucine zipper, zinc finger and transmembrane helix-containing folds are lowest. This
Method Inform Med 4/2001
relates to the functions associated with these folds; the former are commonly involved in metabolic pathways and the latter in signalling or transport processes [98]. This is also reflected in the relationship with subcellular localisations of proteins, where expression of cytoplasmic proteins is high, but nuclear and membrane proteins tend to be low [99, 100]. More complex relationships have also been assessed. Conventional wisdom is that gene products that interact with each other are more likely to have similar expression profiles than if they do not [101, 102]. However, a recent study showed that this relationship is not so simple [103]. While expression profiles are similar for gene products that are permanently associated, for example in the large ribosomal subunit, profiles differ significantly for products that are only associated transiently, including those belonging to the same metabolic pathway. As described below, one of the main driving forces behind expression analysis has been to analyse cancerous cell lines [104]. In general, it has been shown that different cell lines (eg epithelial and ovarian cells) can be distinguished on the basis of their expression profiles, and that these profiles are maintained when cells are transferred from an in vivo to an in vitro environment [105]. The basis for their physiological differences were apparent in the expression of specific genes; for example, expression levels of gene products necessary for progression through the cell cycle, especially ribosomal genes, correlated well with variations in cell proliferation rate. Comparative analysis can be extended to tumour cells, in which the underlying causes of cancer can be uncovered by pinpointing areas of biological variations compared to normal cells. For example in breast cancer, genes related to cell proliferation and the IFN-regulated signal transduction pathway were found to be upregulated [25, 106]. One of the difficulties in cancer treatment has been to target specific therapies to pathogenetically distinct tumour types, in order to maximise efficacy and minimise toxicity. Thus, improvements in cancer classifications have been central to advances in cancer treatment. Although
the distinction between different forms of cancer for example subclasses of acute leukaemia has been well established, it is still not possible to establish a clinical diagnosis on the basis of a single test. In a recent study, acute myeloid leukaemia and acute lymphoblastic leukaemia were successfully distinguished based on the expression profiles of these cells [26]. As the approach does not require prior biological knowledge of the diseases, it may provide a generic strategy for classifying all types of cancer. Clearly, an essential aspect of understanding expression data lies in understanding the basis of transcription regulation. However, analysis in this area is still limited to preliminary analyses of expression levels in yeast mutants lacking key components of the transcription initiation complex [19, 107].
7 many PRACTICAL APPLICATIONS
Here, we describe some of the major uses of bioinformatics.
7.1 Finding Homologues
As described earlier, one of the driving forces behind bioinformatics is the search for similarities between different biomolecules. Apart from enabling systematic organisation of data, identification of protein homologues has some direct practical uses. The most obvious is transferring information between related proteins. For example, given a poorly characterised protein, it is possible to search for homologues that are better understood and with caution, apply some of the knowledge of the latter to the former. Specifically with structural data, theoretical models of proteins are usually based on experimentally solved structures of close homologues [108]. Similar techniques are used in fold recognition in which tertiary structure predictions depend on finding structures of remote homologues and checking whether the prediction is energetically viable [109]. Where biochemical or structural data are lacking, studies
355 What is Bioinformatics?
Fig. 3 Above is a schematic outlining how scientists can use bioinformatics to aid rational drug discovery. MLH1 is a human gene encoding a mismatch repair protein (mmr) situated on the short arm of chromosome 3. Through linkage analysis and its similarity to mmr genes in mice, the gene has been implicated in nonpolyposis colorectal cancer. Given the nucleotide sequence, the probable amino acid sequence of the encoded protein can be determined using translation software. Sequence search techniques can be used to find homologues in model organisms, and based on sequence similarity, it is possible to model the structure of the human protein on experimentally characterised structures. Finally, docking algorithms could design molecules that could bind the model structure, leading the way for biochemical assays to test their biological activity on the actual protein.
could be made in low-level organisms like yeast and the results applied to homologues in higher-level organisms such as humans, where experiments are more demanding. An equivalent approach is also employed in genomics. Homologue-finding is extensively used to confirm coding regions in newly sequenced genomes and functional data is frequently transferred to annotate individual genes. On a larger scale, it also
simplifies the problem of understanding complex genomes by analysing simple organisms first and then applying the same principles to more complicated ones this is one reason why early structural genomics projects focused on Mycoplasma genitalium [75]. Ironically, the same idea can be applied in reverse. Potential drug targets are quickly discovered by checking whether homologues of essential microbial proteins are
missing in humans. On a smaller scale, structural differences between similar proteins may be harnessed to design drug molecules that specifically bind to one structure but not another.
7.2 Rational Drug Design
One of the earliest medical applications of bioinformatics has been in aiding rational
Method Inform Med 4/2001
356 Luscombe, Greenbaum, Gerstein
drug design. Fig. 3 outlines the commonly cited approach, taking the MLH1 gene product as an example drug target. MLH1 is a human gene encoding a mismatch repair protein (mmr) situated on the short arm of chromosome 3 [110]. Through linkage analysis and its similarity to mmr genes in mice, the gene has been implicated in nonpolyposis colorectal cancer [111]. Given the nucleotide sequence, the probable amino acid sequence of the encoded protein can be determined using translation software. Sequence search techniques can then be used to find homologues in model organisms, and based on sequence similarity, it is possible to model the structure of the human protein on experimentally characterized structures. Finally, docking algorithms could design molecules that could bind the model structure, leading the way for biochemical assays to test their biological activity on the actual protein.
protein folds are often related to specific biochemical functions [52, 53], these findings highlight the diversity of metabolic pathways in different organisms [36, 89]. As we discussed earlier, one of the most exciting new sources of genomic information is the expression data. Combining expression information with structural and functional classifications of proteins we can ask whether the high occurrence of a protein fold in a genome is indicative of high expression levels [97]. Further genomic scale data that we can consider in large-scale surveys include the subcellular localisations of proteins and their interactions with each other [114-116]. In conjunction with structural data, we can then begin to compile a map of all protein-protein interactions in an organism.
this way, we are able to examine individual systems in detail and also compare them with those that are related in order to uncover common principles that apply across many systems and highlight unusual features that are unique to some.
Acknowledgments We thank Patrick McGarvey for comments on the manuscript.
References
1. Reichhardt T. Its sink or swim as a tidal wave of data approaches. Nature 1999. 399 (6736): 517-20. 2. Benson DA, et al. GenBank. Nucleic Acids Res 2000; 28 (1): 15-8. 3. Bairoch A, Apweiler R. The SWISS-PROT protein sequence database and its supplement TrEMBL in 2000. Nucleic Acids Res 2000; 28 (1): 45-8. 4. Fleischmann RD, et al. Whole-genome random sequencing and assembly of Haemophilus influenzae Rd. Science 1995; 269 (5223): 496-512. 5. Drowning in data. The Economist 1999 (26 June 1999). 6. Bernstein FC, et al. The Protein Data Bank. A computer-based archival file for macromolecular structures. Eur J Biochem 1977; 80 (2): 319-24. 7. Berman HM, et al. The Protein Data Bank. Nucleic Acids Res 2000; 28 (1): 235-42. 8. Pearson WR, Lipman DJ. Improved tools for biological sequence comparison. Proc Natl Acad Sci USA 1988; 85 (8): 2444-8. 9. Altschul SF, et al. Gapped BLAST and PSIBLAST: a new generation of protein database search programs. Nucleic Acids Res 1997; 25 (17): 3389-402. 10. Fleischmann RD, et al. Whole-genome random sequencing and assembly of Haemophilus influenzae Rd. Science 1995; 269 (5223): 496-512. 11. Lander ES, et al. Initial sequencing and analysis of the human genome. Nature 2001; 409: 860-921. 12. Venter JC, et al. The sequence of the human genome. Science 2001; 291 (5507): 1304-51. 13. Tatusova TA, Karsch-Mizrachi I, Ostell JA. Complete genomes in WWW Entrez: data representation and analysis. Bioinformatics 1999; 15 (7-8): 536-43. 14. Eisen MB, Brown PO. DNA arrays for analysis of gene expression. Methods Enzymol, 1999; 303: 179-205. 15. Cheung VG, et al. Making and reading microarrays. Nat Genet 1999; 21 (1 Suppl): 15-9. 16. Duggan DJ, et al. Expression profiling using cDNA microarrays. Nat Genet 1999. 21 (1 Suppl): 10-4. 17. Lipshutz RJ, et al. High density synthetic oligonucleotide arrays. Nat Genet 1999; 21 (1): 20-4.
8. Conclusions
With the current deluge of data, computational methods have become indispensable to biological investigations. Originally developed for the analysis of biological sequences, bioinformatics now encompasses a wide range of subject areas including structural biology, genomics and gene expression studies. In this review, we provided an introduction and overview of the current state of field. In particular, we discussed the types of biological information and databases that are commonly used, examined some of the studies that are being conducted with reference to transcription regulatory systems and finally looked at several practical applications of the field. Two principal approaches underpin all studies in bioinformatics. First is that of comparing and grouping the data according to biologically meaningful similarities and second, that of analysing one type of data to infer and understand the observations for another type of data. These approaches are reflected in the main aims of the field, which are to understand and organise the information associated with biological molecules on a large scale. As a result, bioinformatics has not only provided greater depth to biological investigations, but added the dimension of breadth as well. In
7.3 Large-scale Censuses
Although databases can efficiently store all the information related to genomes, structures and expression datasets, it is useful to condense all this information into understandable trends and facts that users can readily understand. Broad generalisations help identify interesting subject areas for further detailed analysis, and place new observations in a proper context. This enables one to see whether they are unusual in any way. Through these large-scale censuses, one can address a number of evolutionary, biochemical and biophysical questions. For example, are specific protein folds associated with certain phylogenetic groups? How common are different folds within particular organisms? And to what degree are folds shared between related organisms? Does this extent of sharing parallel measures of relatedness derived from traditional evolutionary trees? Initial studies show that the frequency of folds differs greatly between organisms and that the sharing of folds between organisms does in fact follow traditional phylogenetic classifications [37, 112, 113]. We can also integrate data on protein functions; given that the particular
Method Inform Med 4/2001
357 What is Bioinformatics?
18. Velculescu VE, et al. Serial Analysis of Gene Expression. Detailed Protocol 1999. 19. Holstege FC, et al. Dissecting the regulatory circuitry of a eukaryotic genome. Cell 1998; 95 (5): 717-28. 20. Roth FP, Estep PW, Church GM. Finding DNA regulatory motifs within unaligned noncoding sequences clustered by whole-genome mRNA quant...