December 15, 2003 | MIT researchers have developed a new algorithm -- GRAM (for Gene Regulatory Modules) -- that they say is more effective in discovering regulatory gene networks than traditional computational methods. Their report appears in the Letters section of the Oct. 12 online issue of Nature Biotechnology, and a Java implementation of the algorithm is freely available at psrg.lcs.mit.edu/GRAM/Index.html.
Many similar algorithms combine gene-expression data for transcription factors with information about shared DNA binding motifs to infer regulatory modules. The MIT effort goes a step further by using genomewide location analysis for DNA binding regulators to provide direct evidence of physical binding. GRAM also allows genes to belong to more than one module.
"Because expression and location analysis data provide complementary information, our goal was to develop an efficient computational method for integrating these data sources ... such an algorithm could assign groups of genes to regulators more accurately," write the authors, led by David Gifford of MIT's Computer Science and Artificial Intelligence Laboratory.
Shown above are regulatory modules in the yeast Saccharomyces cerevisiae, identified by GRAM analysis of genomewide data for 106 transcription factors profiled in more than 500 expression experiments. The algorithm identified 106 gene modules, containing 655 distinct genes, regulated by 68 of the transcription factors.
One problem GRAM overcomes is the need to use stringent P cutoff values (distribution measure) to screen out the inherent noise in expression data. This practice often misses relevant genes in its effort to reduce false positives and negatives. Researchers were able to reliably identify 40 percent more gene-regulator interactions by using a less stringent P value. – John Russell