Gene Expression Data Mining

Image: Courtesy of Rosetta Biosoftware The Rosetta Resolver system's Image Viewer application showing an Affymetrix GeneChip probe array. There is a saying: "Be careful what you wish for, you just may get it." Biologists long pined for faster, more efficient ways to gather data; now they generate genomic information faster than they can assimilate it. The result: information overload. The solution: data mining. Though data mining is an ambiguous term, most definitions include the idea of dealing with very large data sets and enabling exploratory data analysis, says Simon Lin, manager, Bioinformatics Core Facility, Duke University. That approach is handy when you're not sure what you're looking for. Traditional analysis, in contrast, tests a hypothesis. "With data mining," Lin says, "you're always getting something unexpected." He cites a clinical collaboration in which a second, heretofore unknown, disease subtype was found, helping to explain why the standard treatment failed for some patients. "We're using data mining to generate more hypotheses, not to confirm therapies," he emphasizes. In another collaboration, data mining provided one principal investigator enough hypotheses to test for the next three years, Lin adds. As yet, no single data-mining system is considered best. Tools designed for business applications don't always deal well with genomics data. The problems are the variability and types of data, and the prevalence of relative but not absolute numbers, Lin explains. As the quantity of genomic data exploded, data-mining systems were designed specifically to handle gene-expression applications. These systems incorporate broader types of data, perform literature searches, and handle high-dimensionality studies involving hundreds of samples and thousands of genes. FRILLS AND THRILLS One of the newest such programs is CAGED (Cluster Analysis of Gene Expression Dynamics),1 developed at the Children's Hospital Informatics Program at Harvard Medical School and the Department of Mathematics and Statistics at the University of Massachusetts, Amherst. The software's developers designed CAGED to mine gene-expression time-series data as well as independent observations. The values of time-series observations depend upon previous values and are known as the memory of the machine, explains Marco Ramoni of Children's Hospital. "In CAGED, you can choose the length of memory of the machine," that is, how many previous time points to consider for each new time point. "Memory can be zero," he adds, allowing the system to mine independent observations as a traditional clustering program would. CAGED combines the visualization capabilities of distance-based clustering with a principled method of distinguishing between two different clusters. Because it is designed for gene-expression data, CAGED can query online databases using accession numbers and, says Ramoni, offers "other frills and thrills for the bioinvestigator." He adds, "The method underpinning the program is general and so can handle other kinds of data." METABOLITE NEIGHBORHOODS Another system, just completed by Pedro Mendes, assistant research professor at the Virginia Bioinformatics Institute in Blacksburg, goes beyond data mining to combine measurements of small molecules (metabolomics) and gene expression (transcriptomics) with genomics. Called the MetExViewer, the software incorporates knowledge collected since the 1930s with new data from mass spectroscopy, microarray, and real-time PCR studies. For each condition being studied-- normal and stressed states of a cell, for example--MetExViewer creates ratios of the data and translates them into colors, Mendes says, so that "it's quick and simple for humans to comprehend." With these colored diagrams, he says, researchers can "better comprehend the relationships of genes without looking for patterns in millions of numbers. The results are as accurate as the technology we used to measure [them] with." MetExViewer introduces the concept of metabolite neighborhoods--groups of metabolites surrounding and interacting with a specific metabolite. That visualization technique shows all reactions with which the specific metabolite is involved, rather than the traditional linear biochemical pathway that displays only one reaction. Mendes' analogy is a map of a city showing all possible routes for incoming and outgoing traffic, versus a traditional set of directions from point A to point B. A public version of this software is in development. A team of scientists at Virginia Tech, using funding from the National Science Foundation's Next Generation Software Program, has developed a data-mining system that addresses the entirety of microarray experimentation. Called Expresso, the software captures both the upstream stages of microarray design and the downstream stages of postprocessing. "Researchers get a screen showing patterns in the context of the biology--not in a vacuum," says Naren Ramakrishnan, one of the project's principle investigators. Currently, among other projects, Expresso is being used to determine what makes loblolly pine trees resistant to drought and stress. "We would like to be able to identify genes that actually help the plant adapt to the stress," Ramakrishnan says. Expresso's role is to highlight interesting patterns, thus helping biologists reformulate their hypotheses and better know which questions to ask. DIRECT SUPERVISION A number of commercial entities develop data-mining software for gene-expression applications. One such company is X-Mine of Brisbane, Calif. Says Sandip Ray, president, founder, and CEO, "We're mining data at a molecular level with a system that simultaneously mines textual information and multivarient data, and therefore distills information into knowledge." Called X-Miner, this program mines microarray and mass spectrometry data, and includes single nucleotide polymorphisms and other items of interest. "It's very flexible," Ray says. It includes a supervised learning suite, unsupervised tools, annotations, and a knowledge- discovery platform. Supervised learning, Ray explains, helps scientists find the proverbial needle in the haystack--hundreds of samples and tens of thousands of variables. "How do you figure out the one or two variables that make several classes of samples different from another without statistical ambiguity?" Ray asks. Most off-the-shelf programs rely upon pattern matching and neural nets or clustering and therefore, he says, lack criteria for classifying data points. In contrast, Ray elaborates that "supervised analysis allows scientists to input data with annotations, such as a time-series experiment, compound dosages, or cell types. By incorporating knowledge from the domain [annotations], you get answers faster with statistical relevance." As an example, Ray cites a customer who used X-Miner for three months and identified biomarkers for organ transplant rejection. By looking at gene expression in white blood cells, researchers gained a relatively painless way to assess tissue with 70% accuracy. The usual method, which involves a "very painful" tissue biopsy and microscopic analysis, is about 60% accurate, Ray notes. Likewise, researchers at Stanford University are identifying biomarkers in heart disease, and researchers at Massachusetts General Hospital Cancer Center are defining markers to identify the effects of environmental estrogen in breast cancer, using X-Miner. "Such questions would be very, very difficult to answer with conventional clinical or pathology tools," Ray adds. X-Miner also helps researchers sift through the published literature. "PubMed gives you all the relevant abstracts for a query," Ray says, "but look at [gene] p53 and you get 24,000 abstracts!" X-Miner's search engine looks for associated memory rather than a direct match. Specifically, Ray says, "Once a gene or term is mentioned, the search path forks. One path looks for all terms associated with the search term, and the other looks for the search term plus its associations." In the case of p53, "you get all the genes that p53 is associated with and why." BEEFING UP STATISTICS "The single most important trend [in data mining] is the advent of statistics into microarray data analysis," says Peter Dansky, senior director of marketing and informatics at Affymetrix, Santa Clara, Calif. It's very important to understand statistical algorithms before the experiment is designed, Lin emphasizes, because algorithms' strengths and weaknesses directly affect interpretations of results and can lead to faulty conclusions.2 "An algorithm only finds what you tell it to look for," says Bill Ladd, senior director of analytic applications at Somerville, Mass.-based Spotfire. Biologists don't always fully appreciate this, and therefore, says Dansky, "tend to come to statisticians too late." In response, data-mining firms known for robust statistics are adding more biological analysis capabilities, and firms concentrating on biological data are adding higher-level statistical analyses to their programs. The goal, for each, is to be more user-friendly. Image: Courtesy of Affymetrix IN THE MINE: Data obtained using Affymetrix's HG-U133A GeneChip array. Spotfire is a case in point. According to Ladd, the company is bringing a variety of mathematics and database tools together to help ensure researchers can access the right analytic methods. To do that, Spotfire provides guides to walk users through the choice of analyses. "Power users can do it without the guides," Ladd notes, and users may add new algorithms as they are developed. Spotfire's visualization engine "lets end-users look at the data they brought together, to look for relationships and trends by subsets and variables," Ladd says, aggregating queries rather than requerying to drill deeper into the data. Specifically, Spotfire takes a larger subset of data into the searchable group at once, so users can tweak the query without starting over. For example, a query such as, "What are the kinases with two-fold levels of activity?" can be adjusted to three-fold levels of activity by updating--not repeating--the query, using an onscreen control. Text-based clustering and pattern matching have been integrated into Spotfire's searches, Ladd says, as well as multidimensional visual environments, which allow things like functional categories or gene-expression patterns to be seen "in a context in which they make sense," Ladd says. "Most analysis programs apply statistical clustering methods, [looking at] up- and down-regulations, and try to determine what's happening," says Dansky. "But, there's a higher order analysis that helps predict networks between genes and how they map out to biologic pathways." One program that takes data mining to this higher level is Clementine. Released this past July by SPSS Science of Chicago, Clementine version 6.5 offers predictive data modeling for gene-expression analysis. It allows the merger and mining of disparate datasets (e.g., microarray, clinical, and toxicological data) using neural networks, clustering, decision trees, and association rules. Logistical regression, machine learning, and prediction are included, but other programs and techniques can be added using the Clementine External Module Interface. "You don't have to be a PhD in statistics to use it," says Colin Shearer, company vice president. The program is based upon a graphical user interface and Cross Industry Standard Process for Data Mining (CRISP-DM)--the public standard. "Templates are the skeleton," of this version, Shearer says. "People build [models] with a collection of icons," transforming raw data to usable forms that reduce variability, aide visualization, and explain specific outcomes using predictive and clustering models. BEYOND TEXT MINING A number of companies have begun offering tools to mine the scientific literature. SPSS, for instance, offers Lexiquest, released this spring. "It works with a spider ... that goes into the text and works out what's in there," and determines what's interesting, Shearer says. "It's more intelligent," than a keyword search, he says, because it performs linguistic analysis and is designed "to watch for new things, frequency of words and concepts, and may notice how research is evolving." Currently, the application can search in English, Dutch, French, German, and Spanish. A Japanese version is under development. In a similar vein, Bernardo Huberman, Hewlett-Packard Fellow, and his team at Hewlett-Packard Laboratories in Palo Alto, Calif., developed a statistical method to swiftly identify and analyze sets of genes associated with specific diseases in the biomedical literature.3 "One big problem," he says, "is that genes are named many ways and abbreviations have multiple meanings. For example, HUGO [Human Genome Organization] found 13,800 redundant gene symbols this year, and 'ER' can stand for emergency room as well as for estrogen receptor." To minimize the babble, the Hewlett-Packard system looks for a correlation between gene names and abbreviations and then scores their relevance based upon the frequency of the gene name in the literature. "We also have a technique called 'disambiguation' that puts words in context by looking at the words that precede [the term's] mention in the text. So far, its success rate exceeds 95%," Huberman says. The system searched the 11 million documents in Medline and created a table of genes and their associated diseases. The process is orders of magnitude faster than creating a table manually. INDUSTRY PIONEER Since Affymetrix pioneered DNA arrays in 1991, it has focused on the data its systems generate, Dansky says. Although Affymetrix software cannot analyze data from other technologies, newer Affymetrix systems combine the company's data with annotations. "Affymetrix has developed new expression-calling algorithms based on statistical methods to ensure that the data going into data-mining applications is robust," Dansky adds. The Affymetrix Data Mining Tool, now in version 3.0, sorts genes by descriptions and by expression results. "It has gene clustering and the common statistical analysis capabilities, but it's not entirely sufficient for most customers," Dansky says. It is, however, open source software so it's easy to export data to other systems. "We work with a long list of third parties," he adds, to incorporate the results from Affymetrix DNA arrays into robust data-mining systems. Affymetrix also offers the Netaffx™ Analysis Center, a Web-based resource "that includes all of our array design information, such as annotations, sequence representations, [related] proteins and pathways, and current knowledge from the public domain and from Affymetrix's own research group," says Dansky. He adds, "Affymetrix integrates and shares this information to help the scientific community to collaborate in the challenge of turning genomic research into biological knowledge." With the Netaffx Analysis Center, users can search for all genes of interest simultaneously among the 15 to 20 databases currently in the system. Several additional vendors also offer tools for gene-expression data mining (see Table). These applications are, clearly, not created equal. "Data mining includes a whole suite of algorithms that are suited for different purposes," Ramakrishnan emphasizes. To use data mining effectively, "you need a good understanding of what you're looking for," he says. Beyond that, "Data mining is well-suited for descriptive analysis. Presented in a succinct form, it tells which questions to ask." And knowing what to ask, in any subject, is a very good first step. Gail Dutton (gaildutton@gi-wireless.net) is a freelance writer in Mayne Island, British Columbia. References 1. M.F. Ramoni et al., "Cluster analysis of gene expression dynamics," Proceedings of the National Academy of Sciences, 99:9121-6, July 9, 2002. 2. R. Nadon, J. Shoemaker, "Statistical issues with microarrays: Processing and analysis," Trends in Genetics, 15:265-71, May 2002. 3. L.A. Adamic et al., "A literature-based method for identifying gene-disease connections," available online at www.hpl.hp.com/shl/papers/genelit/index.html.

Gene Expression Data Mining

Interested in reading more?

Become a Member of