You are here:
Comparison of De Novo Network Reverse Engineering Methods with Applications to Ecotoxicology
Citation:
BURGOON, L. AND S. W. EDWARDS. Comparison of De Novo Network Reverse Engineering Methods with Applications to Ecotoxicology. Presented at 18th Annual International Conference on Intelligent Systems in Molecular Biology, Boston, MA, July 09 - 13, 2010.
Impact/Purpose:
Although the results from the DREAM network construction competitions make clear that methods incorporating other knowledge (e.g., knockouts) fair better in specific circumstances, this additional information are lacking in ecotoxicology studies. Here, we compare the performance of correlation and mutual information network reverse engineering algorithms on synthetic data from AGN.
Description:
The DREAM competitions for network modeling comparisons have made several points clear: 1) incorporating knowledge beyond gene expression data may improve modeling (e.g., data from knock-out organisms), 2) most techniques do not perform better than random, and 3) more complex methods tend to perform well under specific circumstances. However, many of these more complicated techniques may not be suitable for ecotoxicology studies where additional data beyond gene expression are relatively scarce, requiring focus on techniques that only utilize gene expression data. Thus, we compared five network modeling methods that utilize only gene expression data, including correlation (Spearman and Pearson), and mutual information (CLR, MRNET, and ARACNE) methods using 36 publicly available synthetic small world networks. Our results demonstrate that the Spearman, MRNET, CLR, and Pearson methods scored the best amongst the five methods with area under the receiver operating characteristic (ROC) curves ranging from 0.75-0.85, while the ARACNE algorithm consistently underperfonned. We will also discuss our current work in the area of network intelligence fusion to improve overall network reconstruction performance by fusing together data from multiple modeling techniques. These results suggest that simpler correlation-based methods may perform equally well compared to the more complicated mutual information methods, and that gene expression data may be sufficient to achieve satisfactory network reconstructions. This abstract has been reviewed and approved for release by the Environmental Protection Agency but does not necessarily reflect the views ofthe Agency. Mention of trade names or commercial products does not constitute endorsement or recommendations for use. All reviewer requested minor revisions have been addressed and resolved