After reading this chapter, you should know the answers to these questions:
KeywordsSequence Information Protein Data Bank Severe Acute Respiratory Syndrome Basic Local Alignment Search Tool Human Genome Sequence
Why are sequence, structure, and biological pathway information relevant to medicine?
Where on the Internet should you look for a DNA sequence, a protein sequence, or a protein structure?
What are two problems encountered in analyzing biological sequence, structure, and function?
How has the age of genomics changed the landscape of bioinformatics?
What two changes should we anticipate in the medical record as a result of these new information sources?
What are two computational challenges in bioinformatics for the future?
24.1 The Problem of Handling Biological Information
Bioinformatics is the study of how information is represented and analyzed in biological systems, especially information derived at the molecular level. Whereas clinical informatics deals with the management of information related to the delivery of health care, bioinformatics focuses on the management of information related to the underlying basic biological sciences. As such, the two disciplines are closely related—more so than generally appreciated (see Chap. 1). Bioinformatics and clinical informatics share a concentration on systems that are inherently uncertain, difficult to measure, and the result of complicated interactions among multiple complex components. Both deal with living systems that generally lack straight edges and right angles. Although reductionist approaches to studying these systems can provide valuable lessons, it is often necessary to analyze those systems using integrative models that are not based solely on first principles. Nonetheless, the two disciplines approach the patient from opposite directions. Whereas applications within clinical informatics usually are concerned with the social systems of medicine, the cognitive processes of medicine, and the technologies required to understand human physiology, bioinformatics is concerned with understanding how basic biological systems conspire to create molecules, organelles, living cells, organs, and entire organisms. Remarkably, however, the two disciplines share significant methodological elements, so an understanding of the issues in bioinformatics can be valuable for the student of clinical informatics and vice versa.
The discipline of bioinformatics continues to be in a period of rapid growth, because the needs for information storage, retrieval, and analysis in biology—particularly in molecular biology and genomics—have increased dramatically over the past decade. History has shown that scientific developments within the basic sciences tend to have a delayed effect on clinical care and there is typically a lag of a decade before the influence of basic research on clinical medicine is realized. The types of information being gathered by biologists today will drastically alter the types of information and technologies available to the health care workers of tomorrow.
24.1.1 Many Sources of Biological Data
There are three sources of information that are revolutionizing our understanding of human biology and that are creating significant challenges for computational processing. The most dominant new type of information is the sequence information produced by genetic studies. This was enabled by the Human Genome Project, an international undertaking intended to determine the complete sequence of human DNA as it is encoded in each of the 23 human chromosomes. The first draft of the sequence was published in 2001 (Lander et al. 2001) and a final version was announced in 2003 coincident with the 50th anniversary of the solving of the Watson and Crick structure of the DNA double helix. The sequence continues to be revised and refined and efforts are underway to sequence the genomes of many different individuals (1,000 Genomes Consortium, 2010). Essentially, the entire set of genetically driven events from conception through embryonic development, childhood, adulthood, and aging are encoded by the DNA blueprints within most human cells. Given a complete knowledge of these DNA sequences, we are in a position to understand these processes at a fundamental level and to consider the possible use of DNA sequences for diagnosing and treating disease.
While some scientists are studying the human genome, other researchers are studying the functions of the genomes of numerous other biological organisms, including important model organisms (such as mouse, rat, and yeast) as well as important human pathogens (such as Mycobacterium tuberculosis or Haemophilus influenzae). The genomes of these organisms have been determined, and efforts are underway to characterize them. These allow two important types of analysis: the analysis of mechanisms of pathogenicity and the analysis of animal models for human disease. In both cases, the functions encoded by genomes can be studied, classified, and categorized, allowing us to decipher how genomes affect human health and disease.
These ambitious scientific projects are not only proceeding at a furious pace, but also are accompanied in many cases by a new approach to biology, which produces a third source of biomedical information: proteomics, the study of the protein gene products of the genome – the proteome. Proteomics enables researchers to discover the state (quantity and configuration) of proteins within an organism. These protein states can be correlated with different physiological conditions, including disease states. Some of these protein states can be used as identifying markers of human disease.
Additionally, large-scale experimental methodologies are used to collect data on thousands or millions of molecules simultaneously. Scientists apply these methodologies longitudinally over time and across a wide variety of organisms or (within an organism) organs to observe the development of various physiological phenomena. Technologies give us the abilities to follow the production and degradation of molecules on gene expression microarrays (Lashkari et al. 1997), to study the expression of large numbers of genes with one another (Bai and Elledge 1997), and to create multiple variations on a genetic theme to explore the implications of changes in genome function on human disease. All these technologies, along with the genome-sequencing projects, are conspiring to produce a volume of biological information that at once contains secrets to age-old questions about health and disease and threatens to overwhelm our current capabilities of data analysis. Thus, bioinformatics is becoming critical for medicine in the twenty-first century.
24.1.2 Implications for Clinical Informatics
Genetic sequence information in the medical record. With the first set of human genomes now available and prices for gene sequencing rapidly decreasing, it is now cost-effective to consider sequencing every patient genome or at least genotyping key sections of the genomes. The sequence of a gene involved in disease may provide the critical information that we need to select appropriate treatments. For example, the set of genes that produces essential hypertension may be understood at a level sufficient to allow us to target antihypertensive medications based on the precise configuration of these genes. Clinical trials now use information about genetic sequence to define precisely the population of patients who would benefit from a new therapeutic agent. Finally, clinicians may learn the sequences of infectious agents (such as of the Escherichia coli strain that causes recurrent urinary tract infections) and store them in a patient’s record to record the precise pathogenicity and drug susceptibility observed during an episode of illness. In any case, it is likely that genetic information will need to be included in the medical record and will introduce special problems (see also Chap. 2 and Masys et al. 2012). Raw sequence information, whether from the patient or the pathogen, is meaningless without context and thus is not well suited to a printed medical record. Like images (CAT scans, MRIs), sequence data can come in high information density and must be presented to the clinician in novel ways. As there are for laboratory tests, there may be a set of nondisease (or normal) values to use as comparisons, and there may be difficulties in interpreting abnormal values. Fortunately, most of the human genome is shared and identical among individuals; less than 1 % of the genome seems to be unique to individuals. Nonetheless, the effects of sequence information on clinical databases will be significant. As an example, Kaiser Permanente and UCSF recently completed identifying genome wide genetic variants in 100,000 individuals linked to their electronic health records.1 The EMERGE research network is specifically focused on genetic discovery and replication using electronic medical records and genetics (McCarty et al. 2011).
New diagnostic and prognostic information sources. One of the main contributions of the genome-sequencing projects (and of the associated biological innovations) is that we are likely to have unprecedented access to new diagnostic and prognostic tools. Single nucleotide polymorphisms (SNPs) and other genetic markers are used to identify how a patient’s genome differs from the draft genome. Diagnostically, the genetic markers from a patient with an autoimmune disease, or of an infectious pathogen within a patient, will be highly specific and sensitive indicators of the subtype of disease and of that subtype’s probable responsiveness to different therapeutic agents. For example, the severe acute respiratory syndrome (SARS) virus was determined to be a corona virus using a gene expression array containing the genetic information from several common pathogenic viruses.2 In general, diagnostic tools based on the gene sequences within a patient are likely to increase greatly the number and variety of tests available to the physician. Physicians will not be able to manage these tests without significant computational assistance. Moreover, genetic information will be available to provide more accurate prognostic information to patients. What is the standard course for this disease? How does it respond to these medications? Over time, we will be able to answer these questions with increasing precision, and will develop computational systems to manage this information.
Several genotype-based databases have been developed to identify markers that are associated with specific phenotypes and identify how genotype affects a patient’s response to therapeutics. The Human Gene Mutation Database (HGMD) annotates mutations with disease phenotype.3 This resource has become invaluable for genetic counselors, basic researchers, and clinicians. Additionally, the Pharmacogenomics Knowledge Base (PharmGKB) collects genetic information that is known to affect a patient’s response to a drug (more on PharmGKB is described in Translational Bioinformatics, Chap. 25).4 As these data sets, and others like them, continue to improve, the first clinical benefits from the genome projects will be realized.
Ethical considerations. One of the critical questions facing the genome-sequencing projects is “Can genetic information be misused?” The answer is certainly yes. With knowledge of a complete genome for an individual, it may be possible in the future to predict the types of disease for which that individual is at risk years before the disease actually develops. If this information fell into the hands of unscrupulous employers or insurance companies, the individual might be denied employment or coverage due to the likelihood of future disease, however distant. There is even debate about whether such information should be released to a patient even if it could be kept confidential. Should a patient be informed that he or she is likely to get a disease for which there is no treatment? What about that patient’s relatives, who share genetic information with the patient? This is a matter of intense debate, and such questions have significant implications for what information is collected and for how and to whom that information is disclosed (Durfy 1993; see Chap. 10). Passage of the Genetic Information Nondiscrimination Act in 2008 set initial federal guidelines on use of genetic information.5 Additionally, the Personal Genome Project (PGP) has been working to define open consent models for releasing genetic information.6
24.2 The Rise of Bioinformatics
A brief review of the biological basis of medicine will bring into focus the magnitude of the revolution in molecular biology and the tasks that are created for the discipline of bioinformatics. The genetic material that we inherit from our parents, that we use for the structures and processes of life, and that we pass to our children is contained in a sequence of chemicals known as deoxyribonucleic acid (DNA).7 The total collection of DNA for a single person or organism is referred to as the genome. DNA is a long polymer chemical made of four basic subunits. The sequence in which these subunits occur in the polymer distinguishes one DNA molecule from another, and the sequence of DNA subunits in turn directs a cell’s production of proteins and all other basic cellular processes. Genes are discreet units encoded in DNA and they are transcribed into ribonucleic acid (RNA), which has a composition very similar to DNA. Genes are transcribed into messenger RNA (mRNA) and a majority of mRNA sequences are translated by ribosomes into protein. Not all RNAs are messengers for the translation of proteins. Ribosomal RNA, for example, is used in the construction of the ribosome, the huge molecular engine that translates mRNA sequences into protein sequences. Additionally, mRNAs can be modified through alternative splicing, degradation, and formation of secondary structures that influence transcriptions. Once expressed, proteins are frequently modified (such as phosphorylation), and these modifications can change the function of the protein.
Understanding the basic building blocks of life requires understanding the function of genomic sequences, genes, and proteins. When are genes expressed? Once genes are transcribed and translated into proteins, into what cellular compartment are the proteins directed? How do the proteins function once there? Do the proteins need to be modified in order for them to become active? How are the proteins turned off? Experimentation and bioinformatics have divided the research into several areas, and the largest are: (1) DNA and protein sequence analysis, (2) macromolecular structure–function analysis, (3) gene expression analysis, (4) proteomics, and (5) systems biology.
24.2.1 Roots of Modern Bioinformatics
Practitioners of bioinformatics have come from many backgrounds, including medicine, molecular biology, chemistry, physics, mathematics, engineering, and computer science. It is difficult to define precisely the ways in which this discipline emerged. There are, however, two main developments that have created opportunities for the use of information technologies in biology. The first is the progress in our understanding of how biological molecules are constructed and how they perform their functions. This dates back as far as the 1930s with the invention of electrophoresis, and then in the 1950s with the elucidation of the structure of DNA and the subsequent sequence of discoveries in the relationships among DNA, RNA, and protein structure. The second development has been the parallel increase in the availability of computing power. Starting with mainframe computer applications in the 1950s and moving to modern workstations, there have been hosts of biological problems addressed with computational methods.
24.2.2 The Genomics Explosion
The benefit of the human genome sequence to medicine is both in the short and in the long term. The short-term benefits lie principally in diagnosis; the availability of sequences of normal and variant human genes will allow for the rapid identification of these genes in any patient (e.g., Babior and Matzner 1997). The long-term benefits will include a greater understanding of the proteins produced from the genome: how the proteins interact with drugs; how they malfunction in disease states; and how they participate in the control of development, aging, and responses to disease.
The effects of genomics on biology and medicine cannot be overstated. We now have the ability to measure the activity and function of genes within living cells. Genomics data and experiments have changed the way biologists think about questions fundamental to life. Whereas in the past, reductionist experiments probed the detailed workings of specific genes, we can now assemble those data together to build an accurate understanding of how cells work.
24.3 Biology Is Now Data-Driven
Twenty years ago, the use of computers was proving to be useful to the laboratory researcher. Today, computers are an essential component of modern research. This has led to a change in thinking about the role of computers in biology. Before, they were optional tools that could help provide insight to experienced and dedicated enthusiasts. Today, they are required by most investigators, and experimental approaches rely on them as critical elements. This is because advances in research methods such as microarray chips, drug screening robots, X-ray crystallography, nuclear magnetic resonance spectroscopy, proteomic mass spectrometry and DNA sequencing experiments have resulted in experiments that generate massive amounts of data. These data pose new problems for basic researchers on how the data are properly stored, analyzed, and disseminated.
The volume of data being produced by genomics projects is staggering. There are now more than 135 million sequences in GenBank comprising more than 126 billion digits. Since 2008, sequencing has bested Moore’s law (see Chap. 1).8 But these data do not stop with sequence data: PubMed contains over 23 million literature citations, the Protein Data Bank (PDB) contains three-dimensional structural data for over 86,000 protein sequences, and the Gene Expression Omnibus (GEO) contains over 900,000 arrayed samples. These data are of incredible importance to biology, and in the following sections we introduce and summarize the importance of sequences, structures, gene expression experiments, systems biology, and their computational components to medicine.
24.3.1 Sequences in Biology
Sequence information (including DNA sequences, RNA sequences, and protein sequences) is critical in biology: DNA, RNA, and protein can be represented as a set of sequences of basic building blocks (bases for DNA and RNA, amino acids for proteins). Computer systems within bioinformatics thus must be able to handle biological sequence information effectively and efficiently.
One major difficulty within bioinformatics is that standard database models, such as relational database systems, are not well suited to sequence information. Any given position in a sequence can be important because of its own identity, because it is part of a larger subsequence, or perhaps because it is part of a large set of overlapping subsequences, all of which have different significance. It is necessary to support queries such as, “What sequence motifs are present in this sequence?” It is often difficult to represent these multiple, nested relationships within standard relational database schema. In addition, the neighbors of a sequence element are also critical, and it is important to be able to perform queries such as, “What sequence elements are seen 20 elements to the left of this element?” For these reasons, researchers in bioinformatics are developing object-oriented databases (see Chap. 6) in which a sequence can be queried in different ways, depending on the needs of the user.
24.3.2 Structures in Biology
How can we analyze the structures of molecules to learn their associated function? Approaches range from detailed molecular simulations (Levitt 1983) to statistical analyses of the structural features that may be important for function (Wei and Altman 1998).
How can we extend the limited structural data by using information in the sequence databases about closely related proteins from different organisms (or within the same organism, but performing a slightly different function)? There are significant unanswered questions about how to extract maximal value from a relatively small set of examples.
How should structures be grouped for the purposes of classification? The choices range from purely functional criteria (“these proteins all digest proteins”) to purely structural criteria (“these proteins all have a toroidal shape”), with mixed criteria in between. One interesting resource available today is the Structural Classification of Proteins (SCOP),10 which classifies proteins based on shape and function.
24.3.3 Expression Data in Biology
The development of DNA microarrays has led to a wealth of data and unprecedented insight into the fundamental biological machine. The traditional premise is relatively simple; tens of thousands of gene sequences derived from genomic data are fixed onto a glass slide or filter. The sequences for each spot are derived from a single gene sequence and the sequences are attached at only one end, creating a forest of sequences in each spot that are all identical. An experiment is performed where two groups of cells are grown in different conditions, one group is a control group and the other is the experimental group. The control group is grown normally, while the experimental group is grown under experimental conditions. For example, a researcher may be trying to understand how a cell compensates for a lack of sugar. The experimental cells will be grown with limited amounts of sugar. As the sugar depletes, some of the cells are removed at specific intervals of time. When the cells are removed, all of the mRNA from the cells is separated and converted back to DNA, using reverse transcriptase (a special enzyme that can create a DNA copy from an RNA template). This leaves a pool of cDNA molecules (DNA derived from mRNA is called complementary DNA or cDNA) that represent the genes that were expressed (turned on) in that group of cells. Using a chemical reaction, a red fluorescent molecule is attached to the experimental cDNA molecules (the “red group”) and the cDNA from the control group is attached to green fluorescent molecules (the “green group”). These two samples are pooled together and then washed over the glass slide. The labeled cDNA in the pool sticks or hybridizes to the corresponding gene sequences in the spots on the glass slide. The molecules from the two groups will compete with each other to hybridize to the complementary sequences in the spots with the group that had more of a particular sequence binding more of the molecules stuck to the glass slide. Using a scanning confocal microscope and a laser to fluoresce the cDNA with their dye molecules, the amount of red and green fluorescence in each spot can be measured. The ratio of red to green fluorescence for a spot is a measure of the ratio of the expression of that gene product between the red and green groups. This ratio is a measure of whether a gene is being turned off (downregulated) in the experimental group or whether the gene is being turned on (upregulated). As each glass slide can hold up to 40,000 spots each experiment has the potential of measuring the activity of all the genes in a cell due to some experimental change. Expressed genes can include genes that are translated into proteins (mRNA), ribosomal RNA (rRNA), regulatory microRNAs (miRNA) or other gene products. Although protein coding genes were initially the area most studied for disease, miRNAs have recently become increasingly important for disease associations because, unlike mRNA, miRNA’s serve a regulatory role that determines how genes are expressed.
While is it possible for a researcher to measure the fluorescence of a spot on the gene expression glass slide, it is entirely impractical to measure the fluorescence of each of the 40,000 spots on each slide. Additionally, an experiment may be composed of several gene expression slides. Computers become critical for analyzing these data because it is impossible for a researcher to measure and analyze all of those red and green spots. Currently scientists are using gene expression experiments to study how cells from different organisms compensate for environmental changes, how pathogens fight antibiotics, and how cells grow uncontrollably (as is found in cancer). A challenge for biological computing is to develop methods to analyze these data, tools to store these data, and computer systems to collect the data automatically.
24.3.4 Genome Sequencing Data in Biology
Advances in sequencing technology are pivotal in enabling the practice of genomic medicine. Whereas the first human genome sequence was carried out over approximately 13 years at a cost of $2.7 billion (Davies 2010), whole human genomes can now be sequenced in a matter of days at a cost that is growing ever-closer to the magic, if somewhat arbitrary, $1000 price tag. This amount is commonly seen as the price at which it becomes feasible to sequence a patient in the course of clinical care, justifiable both clinically and financially. In 2004, and again in 2011, the National Human Genome Research Institute (part of the National Institutes of Health) funded a number of efforts specifically aimed at increasing speed and decreasing the cost of genome scale sequencing. In addition, in 2006 the X Prize Foundation announced the genomics X PRIZE for sequencing ten high quality genomes in 10 days for under $10,000.
Traditional sequencing involves a method referred to as Sanger sequencing. This method typically is applied to sequences ranging from 300 to 1,000 nucleotides in a non-high throughput manner.11 In the early to mid 2000s, several technologies were introduced to sequence large amounts of DNA in parallel. These technologies, including 454 Pyrosequencing, SOLiD or Solexa (now Illumina) sequencing typically determine much shorter sequences than Sanger based approaches, but can generate as many as 300 gigabases (300 × 109 bases) of sequence in short 75-base fragments at a cost well under $10,000.12 These next generation sequencing methods are being used for many applications, including identification of genetic variants in clinical studies, characterizing genome function with specific experiments and sequencing novel species genomes. These studies have already discovered the genetic basis of rare genetic disorders by sequencing entire families (Ng et al. 2010), and we have seen a glimpse of the future of genome sequencing for routine health care in the analysis of both a single genome of a healthy man (Ashley et al. 2010) and a family of four (Dewey et al. 2011). Currently, we are seeing the introduction of third generation sequencing technologies that promise higher throughput, single molecule sequencing, higher accuracy, or lower cost. One emergent area of research is metagenomics, the study of microorganism ecosystems using DNA sequencing, including the association of human gut flora populations to disease phenotypes in humans (Qin et al. 2010).
24.3.5 Epigenetics Data
Epigenetics consists of heritable changes that are not encoded in the primary DNA sequence. Several types of epigenetic effects can now be studied in the laboratory, and they have been associated to disease and risks of disease (Goldberg et al. 2007). First, the regional structure of chromosomes affects which regions of the genome can be transcribed, that is which regions can be expressed. Large proteins, called histones, coordinate the structure of chromosomes and their structure and positions are regulated with protein posttranslational modifications to the histones bound to the DNA. These changes have been associated with spontaneous mutations in cancer, complex genetic diseases, and Mendelian inherited genetic diseases. Second, cytosine bases in the DNA can be methylated and this can affect gene expression. DNA methylation patterns can be passed on when DNA is replicated. Like chromosome structure, these modifications have been associated with human disease (Bird 2002).
24.3.6 Systems Biology
Recent advances in high throughput technologies have enabled a new, dynamic approach to studying biology, that of systems biology. In contrast to the historically reductionist approach to biology, studying one molecule at a time, systems biology looks at the entirety of a system including dynamic relationships between the different components. With that said, systems biology is still only in its infancy, and even within the field, researchers have not reached consensus on a single definition (Wishart 2008). As an analogy, consider an airplane. Having a “parts list” for a Boeing 747 does not enable us to understand how those parts work together to make the airplane operate. If the airplane breaks, the parts list alone does not tell us how to remedy the situation. Rather, we need to understand how the parts interact, how one affects another, and how perturbations to one part of the system affect the rest of the system. Similarly, systems biology involves understanding not only the “parts list”, i.e. the list of all genes, proteins, metabolites, etc., but also the dynamic networks of interactions among these parts. Recently an integrated simulation of an entire bacterial cell has shown the feasibility of accurate computational simulations of cell physiology (Karr et al. 2012).
Current research in -omics technologies have both enabled and catalyzed the advancement of systems biology. However, a systems biology approach goes beyond simply performing these high bandwidth methods for the purpose of biological discovery. Rather, systems biology implies a systematic, hypothesis-driven approach based on omic-scale (very large) hypotheses. Once the interactions in a biological network are understood, one can model that network to make predictions regarding the system’s behavior, particularly in light of specific perturbations. Understanding how the system has evolved to work can also help us understand what goes wrong when the system breaks down, and how to intervene in order to restore the system to normal.
24.4 Key Bioinformatics Algorithms
There are a number of common computations that are performed in many contexts within bioinformatics. In general, these computations can be classified as sequence alignment, structure alignment, pattern analysis of sequence/structure, gene expression analysis, and pattern analysis of biochemical function.
24.4.1 Early Work in Sequence and Structure Analysis
As it became clear that the information from DNA and protein sequences would be voluminous and difficult to analyze manually, algorithms began to appear for automating the analysis of sequence information. The first requirement was to have a reliable way to align sequences so that their detailed similarities and distances could be examined directly. Needleman and Wunsch (1970) published an elegant method for using dynamic programming techniques to align sequences in time related to the cube of the number of elements in the sequences. Smith and Waterman (1981) published refinements of these algorithms that allowed for searching both the best global alignment of two sequences (aligning all the elements of the two sequences) and the best local alignment (searching for areas in which there are segments of high similarity surrounded by regions of low similarity). A key input for these algorithms is a matrix that encodes the similarity or substitutability of sequence elements: When there is an inexact match between two elements in an alignment of sequences, it specifies how much “partial credit” we should give to the overall alignment based on the similarity of the elements, even though they may not be identical. Looking at a set of evolutionarily related proteins, Dayhoff et al. (Dayhoff 1974) published one of the first matrices derived from a detailed analysis of which amino acids (elements) tend to substitute for others.
Within structural biology, the vast computational requirements of the experimental methods (such as X-ray crystallography and nuclear magnetic resonance) for determining the structure of biological molecules drove the development of powerful structural analysis tools. In addition to software for analyzing experimental data, graphical display algorithms allowed biologists to visualize these molecules in great detail and facilitated the manual analysis of structural principles (Langridge 1974; Richardson 1981). At the same time, methods were developed for simulating the forces within these molecules as they rotate and vibrate (Gibson and Scheraga 1967; Karplus and Weaver 1976; Levitt 1983).
The most important development to support the emergence of bioinformatics, however, has been the creation of databases with biological information. In the 1970s, structural biologists, using the techniques of X-ray crystallography, set up the Protein Data Bank (PDB) specifying the Cartesian coordinates of the structures that they elucidated (as well as associated experimental details) and made PDB publicly available. The first release, in 1977, contained 77 structures. The growth of the database is chronicled on the Web: the PDB now has over 75,000 detailed atomic structures and is the primary source of information about the relationship between protein sequence and protein structure.13 Similarly, as the ability to obtain the sequence of DNA molecules became widespread, the need for a database of these sequences arose. In the mid-1980s, the GENBANK database was formed as a repository of sequence information. Starting with 606 sequences and 680,000 bases in 1982, the GENBANK has grown by much more than 135 million sequences and 125 billion bases.14 The GENBANK database of DNA sequence information supports the experimental reconstruction of genomes and acts as a focal point for experimental groups. Numerous other databases store the sequences of protein molecules15 and information about human genetic diseases.16
Included among the databases that have accelerated the development of bioinformatics is the Medline database of the biomedical literature and its paper-based companion Index Medicus (see Chap. 21).17 Including articles as far back as 1809 and brought online free on the Web in 1997, Medline provides the glue that relates many high-level biomedical concepts to the low-level molecule, disease, and experimental methods. In fact, this “glue” role was the basis for creating the Entrez and PubMed systems for integrating access to literature references and the associated databases.
24.4.2 Sequence Alignment and Genome Analysis
Unfortunately, the dynamic programming algorithms are computationally expensive to apply, so a number of faster, more heuristic methods have been developed. The most popular algorithm is the Basic Local Alignment Search Tool (BLAST) (Altschul et al. 1990). BLAST is based on the observation that sections of proteins are often conserved without gaps (so the gaps can be ignored—a critical simplification for speed) and that there are statistical analyses of the occurrence of small subsequences within larger sequences that can be used to prune the search for matching sequences in a large database. Another tool that has found wide use in mining genome sequences is BLAT (Kent 2003). BLAT is often used to search long genomic sequences with significant performance increases over BLAST. It achieves its 50-fold increase in speed over other tools by storing and indexing long sequences as short, non-overlapping sequences, allowing efficient storage, searching, and alignment on modest hardware.
24.4.3 Prediction of Structure and Function from Sequence
One of the primary challenges in bioinformatics is taking a newly determined DNA sequence (as well as its translation into a protein sequence) and predicting the structure of the associated molecules, as well as their function. Both problems are difficult, being fraught with all the dangers associated with making predictions without hard experimental data. Nonetheless, the available sequence data are starting to be sufficient to allow good predictions in a few cases. For example, there is a Web site devoted to the assessment of biological macromolecular structure prediction methods.18 Recent results suggest that when two protein molecules have a high degree (more than 40 %) of sequence identity and one of the structures is known, a reliable model of the other can be built by analogy. In the case that sequence similarity is less than 25 %, however, performance of these methods is much less reliable.
When scientists investigate biological structure, they commonly perform a task analogous to sequence alignment, called structural alignment. Given two sets of three-dimensional coordinates for a set of atoms, what is the best way to superimpose them so that the similarities and differences between the two structures are clear? Such computations are useful for determining whether two structures share a common ancestry and for understanding how the structures’ functions have subsequently been refined during evolution. There are numerous published algorithms for finding good structural alignments. We can apply these algorithms in an automated fashion whenever a new structure is determined, thereby classifying the new structure into one of the protein families (such as those that SCOP maintains).
A related problem is that of using the structure of a large biomolecule and the structure of a small organic molecule (such as a drug or cofactor) to try to predict the ways in which the molecules will interact. An understanding of the structural interaction between a drug and its target molecule often provides critical insight into the drug’s mechanism of action. The most reliable way to assess this interaction is to use experimental methods to solve the structure of a drug–target complex. Once again, these experimental approaches are expensive, so computational methods play an important role. Typically, we can assess the physical and chemical features of the drug molecule and can use them to find complementary regions of the target. For example, a highly electronegative drug molecule will be most likely to bind in a pocket of the target that has electropositive features.
Prediction of function often relies on use of sequential or structural similarity metrics and subsequent assignment of function based on similarities to molecules of known function. These methods can guess at general function for roughly 60–80 % of all genes, but leave considerable uncertainty about the precise functional details even for those genes for which there are predictions, and have little to say about the remaining genes.
24.4.4 Clustering of Gene Expression Data
Analysis of gene expression data often begins by clustering the expression data. A typical experiment is represented as a large table, where the rows are the genes on each chip and the columns represent the different experiments, whether they be time points or different experimental conditions. Each row is then a vector of values that represent the results of the experiment with respect to a specific gene. Clustering can then be performed to determine which genes are being expressed similarly. Genes that are associated with similar expression profiles are often functionally associated. For example, when a cell is subjected to starvation (fasting), ribosomal genes are often downregulated in anticipation of lower protein production by the cell. It has similarly been shown that genes associated with neoplastic progression could be identified relatively easily with this method, making gene expression experiments a powerful assay in cancer research (see Yan and Gu 2009, for a review). In order to cluster expression data, a distance metric must be determined to compare a gene’s profile with another gene’s profile. If the vector data are a list of values, Euclidian distance or correlation distances can be used. If the data are more complicated, more sophisticated distance metrics may be employed. These methods fall into two categories: supervised and unsupervised. Supervised learning methods require some preconceived knowledge of the data at hand (discussed below). Usually, the method begins by selecting profiles that represent the different groups of data, e.g., genes that represent certain pathways, and then the clustering method associates each of the genes with the representative profile to which they are most similar. Unsupervised methods are more commonly applied because these methods require no knowledge of the data, and can be performed automatically.
The future clinical usefulness of these algorithms cannot be overstated. In 2002, van’t Veer et al. (2002) found that a gene expression profile could predict the clinical outcome of breast cancer. The global analysis of gene expression showed that some cancers were associated with different prognosis, not detectable using traditional means. Another exciting advancement in this field is the potential use of microarray expression data to profile the molecular effects of known and potential therapeutic agents. This molecular understanding of a disease and its treatment will soon help clinicians make more informed and accurate treatment choices (for more, see Chap. 25 on Translational Bioinformatics).
220.127.116.11 Classification and Prediction
A high level description of some common approaches to classification or supervised learning are described below, but note that entire courses could be, and are, taught on each of these methods. For further details we refer readers to the suggested texts at the end of this chapter.
One of the simplest methods for classification is that of k-nearest-neighbor, or KNN. Essentially, KNN uses the classification of the k closest instances to a given input as a set of votes regarding how that instance should be classified. Unfortunately, KNN tends not to be useful for omics-based classification because it tends to break down in high-dimensional space. For high-dimensional data, KNN has difficulty in finding enough neighbors to make prediction, which will lead to large variation in the classification. This breakdown is one aspect of the “curse of dimensionality,” described in more detail below (Hastie et al. 2009).
A more general statistical approach to supervised learning, and one which encompasses a number of popular methods, is that of function approximation. In this approach, one attempts to find a useful approximation of the function f(x) that underlies the actual relation between the inputs and outputs. In this case, one chooses a metric by which to judge the accuracy of the approximation, for example the residual sum of squares, and uses this metric to optimize the model to fit the training data. Bayesian modeling, logistic regression, and Support Vector Machines all use variations on this approach.
Finally, there is the class of rule-based classifiers. This type of classifier may be thought of as a series of rules, each of which splits the set of instances based on a given characteristic. Details such as what criteria are used to choose the feature on which to base a rule, and whether the algorithm uses enhancements such ensemble learning (i.e., multiple models together) determine the specifics of the classifier type, for example decision trees, random forests, or covering rules.
Which approach to use depends both on the nature of the data and the question being asked. The question might prioritize sensitivity over specificity or vice versa. For example, for a test to detect a life-threatening infection that is easily treatable by readily available antibiotics, one might want to err on the side of sensitivity. In addition, data may be numeric or categorical or have differing degrees of noise, missing values, correlated features or non-linear interactions among features. These different qualities are better handled by different methods. In many cases the best approach is actually to try a number of different methods and to compare the results. Such comparative analysis is facilitated through freely available software packages such as R/Bioconductor20 and Weka.21
24.4.5 The Curse of Dimensionality
In the post-genomic era, there is no shortage of data to analyze. Rather, many researchers have more data than they know what to do with. However this overabundance tends to be a factor of the dimensionality of the data, rather than the number of subjects. This mismatch can lead to challenges for experimental design and statistical analysis. Type 1 error, or the tendency to incorrectly reject the null hypothesis and say that indeed there is statistical significance to a pattern (see Chap. 11), is amplified by looking at high-dimensional data. This is one aspect of what is known as the “curse of dimensionality” (Hastie et al. 2009). Consider analysis of gene expression data for 20,000 genes, trying to detect a pattern that can predict outcome. In a sample of, say, 30 subjects—a reasonable number when testing a single hypothesis—by random chance, some number of genes will correlate with outcome. Essentially one is testing not one but 20,000 hypotheses simultaneously. One must therefore correct for multiple hypothesis testing. The Bonferroni method is a common and straightforward approach to correct for multiple hypothesis testing.22 It entails dividing the threshold p-value one would use, traditionally 0.05, by the number of hypotheses. So, for a test of 20,000 genes, one would require a p-value of 2.5 × 10−6 to call a gene significant. Typically, analyses using high dimensional data such as gene expression are not sufficiently powered to pass this stringent test. One would need thousands of samples to be sufficiently powered. Another approach is to use q-value, or false discovery rate (Storey and Tibshirani 2003), rather than p-value. This approach relies on empirical permutation to determine the expected number of false positives if indeed the null hypothesis is correct, which enables approximation of the proportion of false positives among all reported positives. Consider again the microarray experiment above in which each array includes 20,000 genes. We want to know whether gene X was differentially expressed in cases versus controls. Choosing a threshold p-value, or false positive rate, of 0.05 means that 1 time in 20 we will erroneously reject the null hypothesis and predict a false positive. If a statistical test returns 2,000 positives, i.e. 2,000 genes appear to be significantly differentially expressed, we expect 1 in 20 of the genes being analyzed (20,000 * (1/20) = 1,000) or approximately half of them to be false positives. A false discovery rate of 0.05, on the other hand, would mean that 5 % of those called positive, in this case 100 out of 2,000, are false positives. Q-value is thus less stringent than p-value, but may be of greater utility in a high-dimensional omics context than a traditional p-value or correction for multiple hypotheses.
Another approach to analysis of high dimensional data sets is to use dimensionality reduction methods such as feature selection or feature extraction. Feature selection entails extracting only a subset of the features at hand, in this case genes. This may be done in a number of ways, based on which genes vary the most, or on which genes seem to best predict the categorization at hand. In contrast, feature extraction creates a new smaller set of features that captures the essence of the original variation. As an example, imagine a plane flight from Seattle, WA to Key West, FL. One could use a 3-dimensional vector consisting of latitude and longitude to describe the plane’s position at any given point along the way. In this case, one value would describe how far the plane had gone in the north/south direction, and one would indicate how far the plane had gone in the east/west direction. However, if we change the axis along which we are measuring to instead be the direct route along which the plane is flying, then we only need 1 dimension to describe where the plane is located. The distance flown tells us where the plane is located at any given time. This approach of changing the axes is the basis for principle components analysis (PCA), a common method for feature extraction. Instead of going from two dimensions to one, PCA on gene expression data typically goes from tens of thousands of features to just a few. Both for feature selection and feature extraction, it is important to replicate the findings in an independently generated data set in order to be sure the model is not over fitting the data on which it was trained.
24.5 Current Application Successes from Bioinformatics
Biologists have embraced the Web in a remarkable way and have made Internet access to data a normal and expected mode for doing business. Hundreds of databases curated by individual biologists create a valuable resource for the developers of computational methods who can use these data to test and refine their analysis algorithms. With standard Internet search engines, most biological databases can be found and accessed within moments. The large number of databases has led to the development of meta-databases that combine information from individual databases to shield the user from the complex array that exists. There are various approaches to this task.
The Entrez system from the National Center for Biotechnology Information (NCBI) gives integrated access to the biomedical literature, protein, and nucleic acid sequences, macromolecular and small molecular structures, and genome project links (including both the Human Genome Project and sequencing projects that are attempting to determine the genome sequences for organisms that are either human pathogens or important experimental model organisms) in a manner that takes advantages of either explicit or computed links between these data resources.23 Newer technologies are being developed that will allow multiple heterogeneous databases to be accessed by search engines that can combine information automatically, thereby processing even more intricate queries requiring knowledge from numerous data sources. One example is the Bioconductor project, a toolbox for bioinformatics in the R programming language.24
24.5.1 Data Standards and Metadata
Chapter 7 on standards in biomedical informatics addresses standardized terminologies as well as standards for data exchange, and terminologies for translational research are discussed in Chap. 25 (Sect. 25.4.2). A number of specialized exchange formats are relevant to this space. The Human Proteome Organization (HUPO) Proteomics Standards Initiative (PSI), founded in 2002, has a family of markup languages (ML) for different aspects of proteomics: mzML for mass spectrometry, PSI-MI XML for molecular interactions, and mzIdentML for proteomics informatics. The FGED (Functional Genomics Data Society, formed in 1999 as MGED for Microarray Gene Expression Data Society) developed MAGE-ML (Microarray and Gene Expression Markup Language) to enable sharing of microarray data. However, it was determined that MAGE-ML was too technical and cumbersome for labs without dedicated bioinformatics support, so FGED now recommends use of MAGE-TAB, a tab-delimited, spreadsheet-based format, instead.25 The Metabolomics Standards Initiative was formed in 2005 to coordinate efforts among a number of existing groups and initiatives, though metabolomics standards have lagged somewhat behind the progress of their genomic and proteomic siblings (Sansone et al. 2007; Scalbert et al. 2009).
A third type of standard that has been emphasized in bioinformatics is the minimum information standard. The MIBBI (Minimum Information for Biological and Biomedical Investigations) consortium established the MIBBI project in 2008 to provide a freely accessible, web-based resource for minimum information standards and to enable owners of those standards to harmonize overlapping elements where applicable. The most well-known of these standards is MIAME, or Minimum Information About a Microarray Experiment, but the MIBBI portal lists over 30 minimum information reporting guidelines (Taylor et al. 2008). These guidelines identify only the types of information that must be included. They do not specify format or semantics. For example, MIAME enumerates six required pieces of information: (1) raw data, (2) processed data, (3) sample annotation, (4) experimental design including, e.g., which raw data file relates to which sample, which hybridizations are technical versus biological replicates, (5) array annotation, and (6) lab and data processing protocols. Array annotation could include a file that identifies the sequence of each individual spot on the array, or it could simply name a standard commercial microarray chip for which annotation is readily available from the manufacturer. Either of these methods would be MIAME compliant for the array annotation component in that it would enable someone looking at the data to understand what each spot represents. One could argue that these guidelines do not go far enough—without specifying semantics or format, the information cannot be computed over and is therefore less useful than it could be. But some feel that the standards are already too onerous and represent a barrier to data sharing (Galbraith 2006).
Note that of the six facets included in the MIAME standard, only two apply directly to data generated through the experiment. The other four involve data about the data, or metadata. Without knowing what the data fields mean, and how the data was generated, the data itself are of no use. As discussed below, policies to ensure data sharing are becoming more common, but simply putting data on a server is not sufficient. The data must be annotated in a sufficiently complete and structured manner to enable data comprehension and reuse. To this end, a workshop was held in 2007 to explore the possibilities for a simple, common data format to describe data and metadata derived from biological, biomedical, and environmental studies involving omics data as well as more traditional data types (Sansone et al. 2008). The outcome from this workshop was the development of the ISA-Tab format for representation of experimental metadata, along with the open source ISA software suite of tools (Rocca-Serra et al. 2010).
24.5.2 Data Sharing Policies
In 1996, the First International Strategy Meeting on Human Genome Sequencing was held in Bermuda. In this meeting, a set of principles was agreed upon regarding sharing of human genome sequencing data. These principles came to be known as the Bermuda principles. They stipulated that (1) all sequence assemblies larger than 1 kb should be released as soon as possible, ideally within 24 h; (2) finished annotated sequences should be published immediately to public databases; and (3) that all human sequence data generated in large-scale sequencing centers should be made available in the public domain.26
Increasingly, journals and funders require that researchers deposit all types of research data in publicly available repositories (Fischer and Zigmond 2010). In 2009, President Obama announced an Open Government Directive that included plans to make federally funded research data available to the public.27 This announcement describes the NIH's policy regarding published manuscripts in particular, but also notes that the results of government-funded research can take many forms, including data sets. Currently the NIH requires that proposals for funding of over $500,000 include a data sharing plan.28
18.104.22.168 Sequence and Genome Databases
In addition to GENBANK, there are numerous special-purpose DNA databases for which the curators have taken special care to clean, validate, and annotate the data. The work required of such curators indicates the degree to which raw sequence data must be interpreted cautiously. GENBANK can be searched efficiently with a number of algorithms and is usually the first stop for a scientist with a new sequence who wonders “Has a sequence like this ever been observed before? If one has, what is known about it?” There are increasing numbers of stories about scientists using GENBANK to discover unanticipated relationships between DNA sequ-ences, allowing their research programs to leap ahead while taking advantage of information collected on similar sequences.
24.5.3 Structure Databases
Although sequence information is obtained relatively easily, structural information remains expensive on a per-entry basis. The experimental protocols used to determine precise molecular structural coordinates are expensive in time, materials, and human power. Therefore, we have only a small number of structures for all the molecules characterized in the sequence databases. The two main sources of structural information are the Cambridge Structural Database30 for small molecules (usually less than 100 atoms) and the PDB31 for macromolecules (see Sect. 24.3.2), including proteins and nucleic acids, and combinations of these macromolecules with small molecules (such as drugs, cofactors, and vitamins). The PDB has approximately 75,000 high-resolution structures, but this number is misleading because many of them are small variants on the same structural architecture. There are approximately 100,000 proteins in humans; therefore many structures remain unsolved (e.g., Burley and Bonanno 2002). In the PDB, each structure is reported with its biological source, reference information, manual annotations of interesting features, and the Cartesian coordinates of each atom within the molecule. Given knowledge of the three-dimensional structure of molecules, the function sometimes becomes clear. For example, the ways in which the medication methotrexate interacts with its biological target have been studied in detail for two decades. Methotrexate is used to treat cancer and rheumatologic diseases, and it is an inhibitor of the protein dihydrofolate reductase, an important molecule for cellular reproduction. The three-dimensional structure of dihydrofolate reductase has been known for many years and has thus allowed detailed studies of the ways in which small molecules, such as methotrexate, interact at an atomic level. As the PDB increases in size, it becomes important to have organizing principles for thinking about biological structure. SCOP32 provides a classification based on the overall structural features of proteins. It is a useful method for accessing the entries of the PDB.
24.5.4 Analysis of Biological Pathways and Understanding of Disease Processes
The ECOCYC project is an example of a computational resource that has comprehensive information about biochemical pathways. ECOCYC is a knowledge base of the metabolic capabilities of E. coli; it has a representation of all the enzymes in the E. coli genome and of the chemical compounds those enzymes transform.33 It also links these enzymes to their genes, and genes are mapped to the genome sequence.
EcoCyc also encodes the genetic regulatory network of E. coli, describing all protein and RNA regulators of E. coli genes. The network of pathways within ECOCYC provides an excellent substrate on which useful applications can be built. For example, they provide: (1) the ability to guess the function of a new protein by assessing its similarity to E. coli genes with a similar sequence, (2) the ability to ask what the effect on an organism would be if a critical component of a pathway were removed (would other pathways be used to create the desired function, or would the organism lose a vital function and die?), and (3) the ability to provide a rich user interface to the literature on E. coli metabolism. Similarly, the Kyoto Encyclopedia of Genes and Genomes (KEGG) provides pathway datasets for organism genomes.34
24.5.5 Postgenomic Databases
24.6 Future Challenges as Bioinformatics and Clinical Informatics Converge
Bioinformatics did solve all of its problems with the sequencing of the human genome. Indeed, if we can identify a set of challenges for the next generations of investigators, then we can more comfortably claim disciplinary status for the field. Fortunately, there is a series of challenges for which the completion of the first human genome sequence is only the beginning.
24.6.1 Completion of Multiple Human Genome Sequences
With the first human genome in hand, the possibilities for studying the role of genetics in human disease multiply. A new challenge immediately emerges, however: collecting individual sequence data from patients who have disease. Initial studies are promising for identification of the genetic cause of disease using individual or family based sequencing (Lupski et al. 2010). Currently, many researchers have chosen to sequence either entire genomes, or exomes, the regions of the genome that are transcribed into mRNA, and therefore the most likely to cause disease. Researchers now know that more than 99 % of the DNA sequences within humans are identical, but the remaining sequences are different and account for our variability in susceptibility to and development of disease states. It is not unreasonable to expect that for particular disease syndromes, the detailed genetic information for individual patients will provide valuable information that will allow us to tailor treatment protocols and perhaps let us make more accurate prognoses. There are significant problems associated with obtaining, organizing, analyzing, and using this information.
24.6.2 Linkage of Molecular Information with Symptoms, Signs, and Patients
There is currently a gap in our understanding of disease processes. Although we have a good understanding of the principles by which small groups of molecules interact, we are not able to explain fully how thousands of molecules interact within a cell to create both normal and abnormal physiological states. As the databases continue to accumulate information ranging from patient-specific data to fundamental genetic information, a major challenge is creating the conceptual links among these databases to create an audit trail from molecular-level information to macroscopic phenomena, as manifested in disease. The availability of these links will facilitate the identification of important targets for future research and will provide a scaffold for biomedical knowledge, ensuring that important literature is not lost within the increasing volume of published data.
24.6.3 Computational Representations of the Biomedical Literature
An important opportunity within bioinformatics is the linkage of biological experimental data with the published papers that report them. Electronic publication of the biological literature provides exciting opportunities for making data easily available to scientists. Already, certain types of simple data that are produced in large volumes are expected to be included in manuscripts submitted for publication, including new sequences that are required to be deposited in GENBANK and new structure coordinates that are deposited in the PDB. However, there are many other experimental data sources that are currently difficult to provide in a standardized way, either because the data are more intricate than those stored in GENBANK or PDB or they are not produced in a volume sufficient to fill a database devoted entirely to the relevant area. Knowledge base technology can be used, however, to represent multiple types of highly interrelated data.
Knowledge bases can be defined in many ways (see Chap. 22); for our purposes, we can think of them as databases in which (1) the ratio of the number of tables to the number of entries per table is high compared with usual databases, (2) the individual entries (or records) have unique names, and (3) the values of many fields for one record in the database are the names of other records, thus creating a highly interlinked network of concepts. The structure of knowledge bases often leads to unique strategies for storage and retrieval of their content. To build a knowledge base for storing information from biological experiments, there are some requirements. First, the set of experiments to be modeled must be defined. Second, the key attributes of each experiment that should be recorded in the knowledge base must be specified. Third, the set of legal values for each attribute must be specified, usually by creating a controlled terminology for basic data or by specifying the types of knowledge-based entries that can serve as values within the knowledge base.
The development of such schemes necessitates the creation of terminology standards, just as in clinical informatics. There are now many controlled vocabularies (or ontologies) and metadata standards for annotation of genomic or proteomic data. For example, Gene Ontology (GO) is an ontology used for annotation of gene function, and arguably the most widely used ontology in basic research. Metadata standards help define information which should be collected and annotated upon various types of datasets. For example, the MIAME standard (Minimum Information About A Microarray Experiment) defines metadata one would annotate upon a microarray dataset. Recently, the Biosharing website36 is providing links to terminologies and standards for a wide range of data types. Furthermore, a great many tools have been developed to help researchers access and analyze this data. For example, the previously mentioned Bioconductor project provides bioinformatic tools in the R language for solving common problems. Other commonly used tools include BioPerl, BioPython and MATLAB.37
24.6.4 Computational Challenges with an Increasing Deluge of Biomedical Data
An increasing challenge in biomedicine is storing, interpreting and integrating the massive amount of datasets the biomedical community is generating, largely from modern technologies in high throughput experimentation. The amount of DNA sequence data being generated over time has dwarfed Moore’s Law, for example. This issue is important for all areas of biomedical informatics, and is discussed in more detail in the Chapter on Translational Bioinformatics ( Chap. 25).
Bioinformatics is closely allied to clinical informatics. It differs in its emphasis on a reductionist view of biological systems, starting with sequence information and moving to structural and functional information. The emergence of the genome sequencing projects and the new technologies for measuring metabolic processes within cells is beginning to allow bioinformaticians to construct a more synthetic view of biological processes, which will complement the whole-organism, top-down approach of clinical informatics. More importantly, there are technologies that can be shared between bioinformatics and clinical informatics because they both focus on representing, storing, and analyzing biological or biomedical data. These technologies include the creation and management of standard terminologies and data representations, the integration of heterogeneous databases, the organization and searching of the biomedical literature, the use of machine learning techniques to extract new knowledge, the simulation of biological processes, and the creation of knowledge-based systems to support advanced practitioners in the two fields.
Questions for Discussion
How are DNA and protein sequence information changing the way that medical records are managed? Which types of systems are or will be most affected (laboratory, radiology, admission and discharge, financial, order entry)?
It has been postulated that clinical informatics and bioinformatics are working on the same problems, but in some areas one field has made more progress than the other. Identify three common themes. Describe how the issues are approached by each subdiscipline.
Why should an awareness of bioinformatics be expected of clinical informatics professionals? Should a chapter on bioinformatics appear in a clinical informatics textbook? Explain your answers.
Why should an awareness of clinical informatics be expected of bioinformatics professionals? Should a chapter on clinical informatics appear in a bioinformatics textbook? Explain your answers.
One major problem with introducing computers into clinical medicine is the extreme time and resource pressure placed on physicians and other health care workers. Do you think that the same problems are arising in basic biomedical research?
Why have biologists and bioinformaticians embraced the Web as a vehicle for disseminating data so quickly, whereas clinicians and clinical informaticians have been more hesitant to put their primary data online?
If a patient’s entire genome were present in their medical record how would one go about interpreting it clinically? Similarly, if we had an entire electronic health record database that included human genomes, how would a researcher go about finding new or novel genetic associations?
With the many high throughput experiments that are used in biomedical research, how are some ways to integrate those datasets using systems biology? For example, if you had a microarray dataset that annotated gene expression levels and a proteomics dataset that identified protein interactions, how could you jointly use both datasets to identify markers for a disease?
Altman, R. B., Dunker, A. K., Hunter, L., & Klein, T. E. (2003). Pacific symposium on Biocomputing’03. Singapore: World Scientific Publishing. The proceedings of one of the principal meetings in bioinformatics, this is an excellent source for up-to-date research reports. Other important meetings include those sponsored by the International Society for Computational Biology (ISCB, http://www.iscb.org/), Intelligent Systems for Molecular Biology (ISMB, http://iscb.org/conferences.shtml.35), and the RECOMB meetings on computational biology (http://www.ctw-congress.de/recomb/). ISMB and PSB have their proceedings indexed in PubMed.
Baldi, P., & Brunak, S. (2001). Bioinformatics: The machine learning approach. Cambridge, MA: MIT Press. This introduction to the field of bioinformatics focuses on the use of statistical and artificial intelligence techniques in machine learning.
Baldi, P., & Hatfield, G. W. (2002). DNA microarrays and gene expression. Cambridge: Cambridge University Press. Introduces the different microarray technologies and how they are analyzed.
Berg, J. M., Tymoczko, J. L., & Stryer, L. (2010). Biochemistry. New York: W.H. Freeman. The textbook by Stryer and colleagues is well written, and is illustrated and updated on a regular basis. It provides an excellent introduction to basic molecular biology and biochemistry.
Durbin, R., Eddy, S. R., Krogh, A., & Mitchison, G. (1998). Biological sequence analysis: Probabilistic models of proteins and nucleic acids. Cambridge: Cambridge University Press. This edited volume provides an excellent introduction to the use of probabilistic representations of sequences for the purposes of alignment, multiple alignment, and analysis.
Gusfield, D. (1997). Algorithms on strings, trees and sequences: Computer science and computational biology. Cambridge: Cambridge University Press. Gusfield’s text provides an excellent introduction to the algorithmics of sequence and string analysis, with special attention paid to biological sequence analysis problems.
Malcolm, S., & Goodship, J. (Eds.). (2007). Genotype to phenotype (2nd ed.). Oxford: BIOS Scientific Publishers. This volume illustrates the different efforts to understand how diseases are linked to genes.
Pevsner, P. (2009). Bioinformatics and functional genomics. Hoboken, NJ: Wiley. A widely used excellent introduction to bioinformatics algorithms.
http://www.ucsf.edu/news/2011/07/10305/ucsf-and-kaiser-permanente-complete-massive-genotyping-project (accessed November 30, 2012)
http://www.cdc.gov/sars/index.html (accessed November 30, 2012)
http://www.hgmd.org/ (accessed November 30, 2012)
http://www.pharmgkb.org/ (accessed December 20, 2012)
http://www.genome.gov/10002328 (accessed November 30, 2012)
http://www.personalgenomes.org/ (accessed November 30, 2012)
If you are not familiar with the basic terminology of molecular biology and genetics, reference to an introductory textbook in the area would be helpful before you read the rest of this chapter.
http://www.genome.gov/sequencingcosts/ (accessed November 30, 2012)
For more information see http://www.rcsb.org/pdb/ (accessed November 30, 2012)
http://scop.mrc-lmb.cam.ac.uk/scop/ (accessed November 30, 2012)
http://en.wikipedia.org/wiki/DNA_sequencing (accessed November 30, 2012)
http://www3.appliedbiosystems.com/cms/groups/mcb_marketing/documents/generaldocuments/cms_061241.pdf ; (accessed November 30, 2012)
See http://www.rcsb.org/pdb/ (accessed April 26, 2013)
http://www.ncbi.nlm.nih.gov/genbank/ (accessed November 30, 2012)
http://www.uniprot.org/ (accessed November 30, 2012)
http://www.ncbi.nlm.nih.gov/omim (accessed November 30, 2012)
http://www.ncbi.nlm.nih.gov/pubmed (accessed November 30, 2012)
http://predictioncenter.org/. (accessed November 30, 2012)
http://www.cgl.ucsf.edu/Research/minrms/ (accessed November 30, 2012)
http://bioconductor.org/ (accessed November 30, 2012)
http://www.cs.waikato.ac.nz/ml/weka/ (accessed November 30, 2012)
http://en.wikipedia.org/wiki/Bonferroni_correction (accessed November 30, 2012)
http://www.ncbi.nlm.nih.gov/Entrez/ (accessed December 20, 2012)
http://bioconductor.org/ (accessed November 30, 2012)
http://www.mged.org/mage-tab/ (accessed November 30, 2012)
http://www.ornl.gov/sci/techresources/Human_Genome/research/bermuda.shtml (accessed November 30, 2012)
http://edocket.access.gpo.gov/2009/E9-29322.htm (accessed November 30, 2012)
http://grants.nih.gov/grants/guide/notice-files/NOT-OD-03-032.html (accessed November 30, 2012)
http://genome.ucsc.edu/ (accessed November 30, 2012)
http://www.ccdc.cam.ac.uk/products/csd/ (accessed November 30, 2012)
http://www.rcsb.org/pdb/ (accessed November 30, 2012)
http://scop.mrc-lmb.cam.ac.uk/scop/ (accessed November 30, 2012)
http://ecocyc.org/ (accessed November 30, 2012)
http://www.genome.jp/kegg/pathway.html (accessed November 30, 2012)
http://www.ncbi.nlm.nih.gov/omim/ (accessed November 30, 2012)
http://biosharing.org/ (accessed November 30, 2012)
http://www.open-bio.org/ (accessed November 30, 2012)
- Dayhoff, M. O. (1974). Computer analysis of protein sequences. Federal Proceedings, 33(12), 2314–2316.Google Scholar
- Durfy, S. J. (1993). Ethics and the human genome project. Archives of Pathology & Laboratory Medicine, 117(5), 466–469.Google Scholar
- Hastie, T., Tibshirani, R., et al. (2009). The elements of statistical learning: Data mining, inference, and prediction. New York: Springer.Google Scholar
- Kent, W. J. (2003). BLAT – the BLAST-like alignment tool. Genome Research, 12(4), 656–664.Google Scholar
- Lashkari, D. A., DeRisi, J. L., McCusker, J. H., Namath, A. F., Gentile, C., Hwang, S. Y., Brown, P. O., & Davis, R. W. (1997). Yeast microarrays for genome wide parallel genetic and gene expression analysis. Proceedings of the National Academy of Sciences, 94(24), 13057–13062.CrossRefGoogle Scholar
- Masys, D. R., Jarvik, G. P., Abernethy, N. F., Anderson, N. R., Papanicolaou, G. J., Paltoo, D. N., Hoffman, M. A., Kohane, I. S., & Levy, H. P. (2012). Technical desiderata for the integration of genomic data into electronic health records. Journal of Biomedical Informatics, 45(3), 419–422.PubMedCrossRefGoogle Scholar
- Wei, L., Altman, R.B. (1998). Recognizing protein binding sites using statistical descriptions of their 3D environments. Proceedings of the pacific symposium on Biocomputing ’98 (pp. 497–508), Singapore.Google Scholar