摘要：

Recent progress in high-throughput technology has resulted in a significant data overload. Determining howto obtain valuable knowledge from such massive raw data has become one of the most challenging issues inbiomedical research. As a result, bioinformatics researchers continue to look for advanced data analysistools to analysis and mine the available data. Correlation network models obtained from various biologicalassays, such as those measuring gene expression levels, are a powerful method for representing correlatedexpression. Although correlation does not always imply causation, the correlation network has been shownto be effective in identifying elements of interest in various bioinformatics applications. While these modelshave found success, little to no investigation has been made into the robustness of relationships in thecorrelation network with regard to vulnerability of the model according to manipulation of sample values.Particularly, reservations about the correlation network model stem from a lack of testing on the reliabilityof the model. In this work, we probe the robustness of the model by manipulating samples to create sixdifferent expression networks and find a slight inverse relationship between sample count and networksize/density. When samples are iteratively removed during model creation, the results suggest that networkedges may or may not remain within the statistical parameters of the model, suggesting that there is roomfor improvement in the filtering of these networks. A cursory investigation into a secondary robustnessthreshold using these measures confirms the existence of a positive relationship between sample size andedge robustness. This work represents an important step toward better understanding of the critical noiseversus signal issue in the correlation network model.

展开