Computational Gold: Data mining and bioinformatics software for the next millennium

Table of Bioinformatics Systems (Enhanced version) The entire process of discovery and invention is a marriage of inquiry, experimentation, and observation, and the product of learning and extending upon the findings of previous investigations. Many of the major scientific and technological discoveries of the 20th century were in part a product of knowledgeable, intuitive researchers who utilized innovative problem-solving approaches and protocols and had the experimental means and comprehensive access to critical information. Having the experimental means and appropriate information can make the difference between success and failure. A case in point is Watson and Crick's access to Rosalind Franklin's DNA X-ray crystallographic data, which provided key clues toward their elucidation of the three-dimensional (3-D) structure of DNA in 1953. The structure was deduced using a novel experimental approach: the construction of a physical 3-D model of DNA using paper and wire. Affymetrix's LIMS Process diagram Experimental means have advanced significantly since the early years of paper, wire, stick, and ball modeling techniques and now include mathematical approaches that use advanced computer systems and data processing techniques. And more dramatically so, information and ready access to information--from unpublished results (for example, personal communications and poster presentations) to data in voluminous data sets (for example, GenBank, SWISS-PROT, PDB)--continue to have a significant impact on the application of experimental approaches and hence scientific discovery. In the early years of modern discovery, research findings would appear in a journal, then get buried in some library archive. In addition, the information existed in various formats--textual, graphic, hard copy, tape, film, and so on--and was not easily archived or retrievable using technologies of the day. Data analyses were generally limited to the slide rule and manual manipulation. Over the past decade, however, technological advances in computational science and scientific instrumentation have facilitated the exponential growth in data and the tools to record and analyze this data. What was the Computer Age as we entered the 1990s has been supplanted by the Information Age. This change was made possible by the advent of the Internet, in particular the World Wide Web. This innovative, truly universal mechanism of information dissemination, in concert with new computation-based analytical tools, has provided practically endless opportunities for scientific discovery. Some of the earliest biology-related accessible information databases were of bibliographic records. These were soon followed by DNA and protein sequence and structure (coordinate) data resources and then by physico- chemical information databases on everything from atoms to macromolecules. Recently joining the information foray have been specialized information resources on protein families; metabolic pathways; information sets on the etiology, biochemistry, and genetic basis of diseases; sequence motif or pattern databases; and so on. The means to analyze this data have also grown, from the basic modeling programs of the mid-1960s to sophisticated multifunction computer-based applications. The volumes of data currently available and the mass of new data generated every day have far exceeded our human capacity to grasp its significance, much less analyze it. This has spawned the field of "bioinformatics" (biological information), which involves the development of computational tools, not only to analyze this information but to provide mechanisms to record, store, and retrieve it. Many of the scientific advances in molecular biology, biochemistry, drug design, and medicine can be attributed to these new computational tools and information resources. Innovations in microchip design, computer architecture, database design and storage systems, and networking have broadened the use of the wealth of information in the databases and the efficiency with which computational programs use it. In addition, new experimental approaches, "algorithms," to problem solving have advanced the specificity and sensitivity of computational analyses. Bioinformatics methods today generally resolve problems of sequence comparison or alignment, database searching, evolutionary relatedness, pattern recognition, and structure prediction. Currently available bioinformatics programs contain routines that provide rigorous implementations of these methods, as well as numerous ancillary functions. These functions include, for example, basic sequence manipulation (editing), calculation of an isoelectric point or molecular mass, determination of amino composition or sequence length, creation of theoretical peptide or restriction maps for specific chemical or enzymic digestions, translations and back-translations, determination of optimal sequence primers for PCR reactions, and assembly of a contiguous sequence from a series of sequencing experiments. The complexity of current biological problems usually requires a suite of bioinformatics routines, each capitalizing on the specific analytical assets of the various methods. A researcher can also run these analyses independently and integrate the results (data) between routines, a tedious but not impossible process. Another way to accomplish these tasks is to use integrated, multifunction applications, such as those profiled in Table 1. Many multifunction bioinformatics applications build upon various tools and methods and include routines that work synergistically to analyze data. In addition, many of the routines within a given package may have been optimized for commonly used analyses, thus requiring minimal input from the researcher at the onset of an analysis. Bioinformatics-based data analysis usually encompasses a broad range of computational processes. Understanding some of the fundamentals of these analytical techniques and methods is important to the appropriate and efficient use of bioinformatics software applications. Thus, a basic description of some of the key problem-solving methods is presented below. Keep in mind that this is not an exhaustive essay. Readers requiring more comprehensive and in-depth explanations of informatics theory should consult M. Gribskov and J. Devereux's Sequence Analysis Primer.1,2 The most common bioinformatics task is the search of a sequence database with a query sequence provided by the researcher. This process is basically a sequence alignment that is reiterated for each known sequence in the database. One of the earliest and most fundamental sequence alignment routines is the dot matrix (plot) comparison--the search for short, defined-length perfect matches between two sequences. The analysis is essentially a one-to-one comparison between residues. This technique was originally developed in the early 1970s but didn't see widespread use until the 1980s when Maizel and Lenk introduced an implementation of the dot matrix method for use on desktop computer systems.3 Shortly thereafter, improved versions of this method incorporated the use of sequence segment ("window") comparisons and output filtering (scoring matrices). These methods compare segments of overlapping sequence and establish matches as a function of exceeding a minimum threshold defined by a scoring matrix. Scoring matrices can be based upon any number of physical, chemical, structural, and/or evolutionary similarities of and between amino acids or nucleotides. The simplest, unitary or identical, matrices are those in which comparisons between identical molecules are scored as positive numbers. They are used primarily for DNA sequence comparisons. Protein sequence comparison matrices are generally based upon the chemical or mutational similarity of the amino acids. The most commonly used protein scoring matrix is the mutation data matrix.4 A relatively new matrix finding more widespread use is the blocks substitution matrix (BLOSSUM) developed by Heinikoff and Heinikoff.5 Paracel's GeneMatcher™ computer hardware For many years dot plot methods were the stalwarts of sequence analysis. Although these methods still find much use today (they are still excellent tools for comparing two sequences), they have been superseded by more sensitive and specific methods. Practically all new sequence alignment or comparison routines are based upon dynamic programming methods, initially introduced by Needleman and Wunch.6 Dynamic programming allows for the introduction of artificial gaps in the aligned sequences to create an optimal alignment. Obviously, there has to be a penalty for introducing a gap into a perfectly complete sequence, hence the term "gap penalty" in many alignment applications. Gap penalties are length independent or dependent. In the former case, the same penalty is applied to all gaps regardless of length; in the latter, the degree of penalty is a function of the gap length. Most dynamic programming methods align all the residues or nucleotides of two sequences from end to end, commonly referred to as "global alignment." One disadvantage of global alignment is that it can mask internal regions of high similarity since the alignment is directed toward a global "view" of the sequences. To find these internal regions of high similarity, Smith and Waterman introduced a local alignment method, which finds the two most similar regions of a sequence and aligns them.7 Global alignment applications are best for comparing sequences that are evolutionarily related, whereas local alignment methods are best used on sequences of unknown relatedness. The methods described above are basically concerned with a one-to-one comparison between two sequences. But what about comparisons between three, four, five, or one or two dozen sequences? One fundamental approach to this problem is the progressive alignment method introduced by Feng and Doolittle.8 This method capitalizes upon the efficiency of pairwise alignment methods by constructing an initial pairwise alignment (cluster) of the two most similar sequences. This cluster is then progressively aligned with other clusters of similarly aligned sequences. In essence, this approach breaks the multiple alignment into a series of pairwise alignments and then compares these alignments to establish relatedness. As alluded to previously, database searching is accomplished through alignment comparison. The large number of sequences in the typical DNA or protein database and the length of the sequences can be major impediments to complete, timely searches (especially on desktop computer systems). These hurdles can be overcome by simplifying the search process using hashing techniques. In this approach, a table ("hash") of the location(s) of short sequences found in the query sequence is created, then sequences in the database are examined against the hash table. The FASTA routine is one of the first implementations of this computational approach and is used widely in most protein database searches.9 A newer implementation of the hash approach, the Basic Local Alignment Search Tool (BLAST) focuses on the significance of ungapped alignments and makes a statistical ranking of matches.10 This approach has enhanced search sensitivity and specificity. Since its inception, the BLAST routine has become one of the most commonly used database search methods. Many of the ancillary sequence analysis routines within software packages are derivatives or extensions of the sequence comparison and database searching methods. One example is pattern-finding routines that search for a particular sequence motif in a query sequence or compare the query sequence against a database of sequence motifs. "Homology" is a commonly used expression to define evolutionary relatedness, and the phylogenetic tree is a graphical representation of this relatedness. Evolutionary relatedness is typically determined by calculating a distance matrix for multiply aligned sequences. This matrix is subsequently used by clustering routines to calculate a sum distance or "relatedness" for the aligned sequences. Distance matrices are essentially base-by-base or residue-by-residue counts of the number of matches and gaps between two aligned sequences. Commonly used "clustering" methods include the unweighted pair group method using arithmetic averages (UPGMA) and the neighbor joining method. The former determines relatedness by progressively working through the pairwise alignments recording distance minima. The latter method is an extension of the UPGMA method in that it seeks to minimize the distance sum of the branch lengths for the phylogenetic tree it creates. Research into the development of accurate methods of predicting secondary protein structure was at the forefront of protein research in the 1960s and 1970s and continues to receive attention today, mostly in relation to how these prediction methods impact tertiary structure prediction methods. One of the earliest and most universally used methods was developed by Chou and Fasman.11 In this method, the spatial orientation of amino acids in a protein sequence (that is, geometries consistent with alpha-helices, beta-sheets, loops, etc.) were inferred from the distribution of amino acids in the secondary structures of protein structures resolved experimentally. There are other approaches to secondary structure prediction, but all are only fair predictors of secondary structure (<70 percent correct for all residues in a given prediction). For those tools that use a database, the researcher should keep in mind that the analytical tools are only as good as the information they work on. In the competitive research environment today, it is imperative that researchers perform computational analyses using not only the most appropriate tool but also the most current database(s). And "the most current database" usually refers to the database that was updated last night! Depending upon one's field of study, using a database that has not been updated in the last week can mean the difference between finding something novel today or taking a back seat to a colleague who makes "your" discovery tomorrow. Many proprietary bioinformatics applications are packaged with a select set of biodatabases.12 In addition, some program suites require an Internet connection to automatically update internal databases or access external databases on an as-needed basis. Other applications may require the user to find, download, and integrate these databases, or supply the appropriate data on a case-by-case basis. Finding these databases and having a mechanism to download them are major hurdles to many information consumers (i.e., bench researchers). To some extent, these problems have been resolved with the advent of the Internet. The Internet has become one of the chief mediums by which information is disseminated to and by the user community. It has become so popular that the tools to manage and index Internet resources (e.g., search engines such as HotBot, AltaVista, Yahoo) have not been able to keep up with the amount of new resources added every day. So although the Internet provides a simple mechanism to retrieve data, finding the data you need can be a bit of a problem. To facilitate your search, compendiums of molecular biology tools, databases, and information resources are now readily available on the Web. Some of the more comprehensive guides include the CMS Molecular Biology Resource;13 (www.sdsc.edu/restools), ABIM (www-biol.univ-mrs.fr/english/biology.html), and www.infobiogen.fr/services/dbcat). For those who are looking to dive right into a database, here are annotated descriptions of various databases and biological information resources, some better known than others. They all provide the research community with invaluable information. Perhaps one of the oldest and foremost of the publicly available protein structure data-bases is the Protein DataBank (PDB). Hosted by Brookhaven National Laboratories for over 17 years, the PDB recently moved to the Research Collaboratory for Structural Bioinformatics (RCSB; www.rcsb.org). The PDB is the primary repository of the 3-D atomic coordinate data files for macromolecules. Other structure databases and information resources, many derived from PDB data, include the protein structure-sequence information library, 3D-ALI (www.embl-heidelberg.de/argos/ali/ali.html); the database of molecular movements, MolMovDB (bioinfo.mbb.yale.edu/MolMovDB); the biological macromolecule crystallization database, BMCD (bioidc2.carb.nist.gov:4040/bmcd/bmcd.html); the Nucleic Acids Database (ndbserver.rutgers.edu); the RNA Secondary Structure [Images] Resource (pundit.icmb.utexas.edu); the Cambridge Crystallographic Data Center, CCDS (www.ccdc.cam.ac.uk); the Prosthetic Groups and Metal Ions in Protein Active Sites Database, PROMISE (bioinf.leeds.ac.uk/promise/MAIN.html); and the Monosaccharide Browser (bmbwww.leeds.ac.uk/designs/monosac/home.htm). Paracel's GeneMatcher™ computer hardware Some of the oldest and best protein sequence databases include SWISS-PROT (www.expasy.ch/sprot/); the Protein Information Resource, PIR (www-nbrf.georgetown.edu/pirwww/pirhome.shtml); and the protein data within the European Molecular Biology Laboratory (EMBL) and GenBank sequence databases. GenBank protein sequence data, as well as data from EMBL, the DNA DataBank of Japan (DDBJ), PIR, SWISS-PROT, and the PDB, are accessible through the Entrez Web interface (www.ncbi.nlm.nih.gov/Entrez/protein.html) provided by the National Center for Biotechnology Information (www.ncbi.nlm.nih.gov). Other more specialized protein sequence resources include the Yeast Protein Databank (www.proteome.com/YPDhome.html); the E. coli Genome and Proteome Database, GenProtEC (DBASE.mbl.edu/ecoli/menu.htf); and the Immuno-globulin Database, IMGT (imgt.cnusc.fr:8104). GenBank, EMBL, and DDBJ (www.ddbj.nig.ac.jp) are the best known and most widely used DNA sequence databases. Examples of specialized DNA databases include the Database of Information is available for many living and questionably living organisms, ranging from viruses to higher eukaryotes, including protists, protozoa, bacteria, fungi, higher plants, insects, and animals. The most accessed genomic databases are those associated with highly active sequencing projects or for which complete genomes are available. These include the Human Genome database, GDB (www.gdb.org); the Mouse Genome Database, MGD (www.informatics.jax.org); the A. thaliana Database, AtDB (genome-www.stanford.edu/Arabidopsis); the Drosophila Genome Database, FlyBase (flybase.bio.indiana.edu); the C. elegans Database, AceDB (www.sanger.ac.uk/Projects/C_elegans/); and the E. coli Database (susi.bio.uni-giessen.de/ecdc.html). Examples of more esoteric databases include the B. subtillus database (pbil.univ-lyon1.fr/nrsub/nrsub.html); the Calicivirus sequence database (www.iah.bbsrc.ac.uk/virus/Caliciviridae/marine.htm); the Maize database, MaizeDB (www.agron.missouri.edu/); and the Puffer Fish database, FUGU project (fugu.hgmp.mrc.ac.uk). Other information resources that are seeing more use are those that serve up unprocessed data from experiments. These data sets typically consist of two-dimensional protein fractionation gel maps but can also contain data from other experimental techniques, for example, peptide fragment mapping or mass spectral analyses. The 2D-PAGE Database (www.expasy.ch/ch2d/ch2d-top.html) is one of many examples of the former, while MassBank (prowl.rockefeller.edu/PROWL/massbank.html), represents one of the latter. Another class of information resources that may not play a direct part in a computational analysis but could be of great benefit to researchers themselves are those that provide biological function information or physical, chemical, and atomic constants data on specific molecules or macromolecules. An excellent example of this group is the Amino Acid Information resource (prowl.rockefeller.edu/aainfo/contents.htm) at PROWL, which provides information on amino acid chemical structure, classification, residue volumes, areas, pKa's, bond geometries, genetic codes, and so on. The IMB-Jena amino acid image collection (gopher://gopher.imb-jena.de/11/ftp/images/PROTEINS/amino_acids) provides 3-D space-filling structure images for all the amino acids. Information on specific biomacromolecules is accessible through the Klotho (www.ibc.wustl.edu/moirai/klotho), LIGAND (www.genome.ad.jp/dbget/ligand.html), and Biocatalysis /Biodegradation (www.labmed.umn.edu/umbbd) databases. The author, Chris Smith, can be contacted at csmith@sdsc.edu. Table of Bioinformatics Systems (Enhanced version) R.F. Doolittle, Methods in Enzymology, New York, Academic Press, 1990. A. Baxevanis, Methods of Biochemical Analysis, New York, John Wiley & Sons, 1998. J.V. Maizel et al., "Enhanced graphic matrix analysis of nucleic acid and protein sequences," Proceedings of the National Academy of Sciences, 78:7665-9, 1981. Dayhoff et al., "A model of evolutionary change in proteins," Atlas of Protein Sequence and Structure, 5:345-8, 1978. S. Heinikoff et al., "Amino acid substitution matrices from protein blocks," Proceedings of the National Academy of Sciences, 89:10915-9, 1992. S.B. Needleman et al., "A general method applicable to the search of similarities in the amino acid sequences of two proteins," Journal of Molecular Biology, 48:443-53, 1970. T.F. Smith et al., "Comparison of biosequences," Advances in Applied Mathematics, 2:482-9, 1981. D. Feng et al., "Progressive multiple sequence alignment," Journal of Molecular Evolution, 25:351-60, 1987. D.J. Lipman et al., "Rapid and sensitive protein similarity searches," Science, 227:1435-41, 1985. S.F. Altschul et al., "Basic Local Alignment Search Tool," Journal of Molecular Biology, 215:403-10, 1990. P.Y. Chou et al., "Prediction of secondary structure of proteins from their amino acid sequence," Advances in Enzymology and Related Areas of Molecular Biology, 47:45-147, 1978. L. DeFrancesco, "Getting down to bases," The Scientist, [20]:18, Oct. 12, 1998. C.M. Smith, "CMS Molecular Biology Resource," Trends in Genetics, 13:416, 1997.

Computational Gold: Data mining and bioinformatics software for the next millennium

Interested in reading more?

Become a Member of