"This is just the tip of the iceberg," said bioinformatics specialist Atul Butte, MD, PhD, who is also a pediatrician at Lucile Packard Children's Hospital at Stanford. "Nearly 100 different diseases have been studied using microarrays, spanning all of medicine. This is a new way to explore this type of data. We can study virtually everything that's been studied." Butte is the first author of the study, which is published in the Jan. 6 online issue of Nature Biotechnology.
The advance comes with a caveat, however: clinically useful nuggets will be buried under the avalanche of data inundating international repositories each year unless scientists come up with a way to better classify their experiments and results.
"Libraries figured out a long time ago how to classify items using the Dewey decimal and other systems," said Butte, who estimates that the contents of the databases are more than doubling each year. "We need to write software now that will help scientists assign the proper concepts to each experiment."
Microarray experiments allow researchers to compare the expression patterns of tens of thousands of individual genes over time in diseased and healthy cells, or in many other experimental conditions. Each experiment generates thousands of pieces of data about the cell's genes. Although biologists use the technology routinely, focusing only on the few results pertinent t o their particular research topic, most scientific journals require that their authors submit all of their data to international databases for use by other researchers.
Butte and his Harvard co-author, Isaac Kohane, MD, PhD, used computer programs to automatically categorize the tens of thousands of microarray experiments in a single database based on the terms, or concepts, used by the submitter to describe the experiment. They then looked for findings shared by several experiments with similar concepts, such as tissue type, for example. Comparing results from many similar experiments allowed them to identify correlations that may not be statistically significant in just one experiment.
Butte and Kohane identified several previously unknown correlations: nine genes whose expression increased or decreased significantly with aging, two genes that are highly expressed in response to injury, and another gene in which the expression drops significantly in leukemic cells. They also confirmed these relationships by studying genes known to be associated with muscle tissue in both humans and mice.
Their classification system was stymied, however, when scientists included too much or too little information in the text annotations, or used imprecise words such as "pool," which can mean either a body of water or the action of combining the contents of two or more tubes.
"As a community, we've standardized the way the data itself is represented," said Butte, "but there are no formal requirements for the accompanying textual descriptions of this data. Sometimes people seem to almost copy and paste their entire scientific paper into the text box. We need to clean up our annotations because now we're showing that they have value."
Butte and Kohane favor using the existing Unified Medical Language System, which consists of more than 1 million biomedical concepts, to vastly simplify the computerized sorting of the thousands of microarray exper iments submitted to databases each year. Without such a system, valuable information will simply be lost as the results pile up. The National Institutes of Health recently funded the National Center for Biomedical Ontology, a consortium led by Stanford professor Mark Musen, MD, PhD, to develop ontologies to allow scientists to describe their data in standardized ways.
"All the answers are already there," said Butte. "We've reached a critical mass with this data. But unless we're careful, we're going to end up with a big mess."