Several developments regarding the analysis of gene co-expression profiles using complex network theory have been reported recently. Such approaches usually start with the construction of an unweighted gene co-expression network, therefore requiring the selection of a suitable threshold defining which pairs of vertices will be connected. We aimed at addressing such an important problem by suggesting and comparing five different approaches for threshold selection. Each of the methods considers a respective biologically-motivated criterion for electing a potentially suitable threshold. A set of 21 microarray experiments from different biological groups was used to investigate the effect of applying the five proposed criteria to several biological situations. For each experiment, we used the Pearson correlation coefficient to measure the relationship between each gene pair, and the resulting weight matrices were thresholded considering several values, generating respective adjacency matrices (co-expression networks). Each of the five proposed criteria was then applied in order to select the respective threshold value. The effects of these thresholding approaches on the topology of the resulting networks were compared by using several measurements, and we verified that, depending on the database, the impact on the topological properties can be large. However, a group of databases was verified to be similarly affected by most of the considered criteria. Based on such results, it can be suggested that when the generated networks present similar measurements, the thresholding method can be chosen with greater freedom. If the generated networks are markedly different, the thresholding method that better suits the interests of each specific research study represents a reasonable choice.

中文翻译：

---

阈值对基因共表达网络拓扑结构的影响

最近已经报道了有关使用复杂网络理论分析基因共表达谱的一些进展。此类方法通常从构建非加权基因共表达网络开始，因此需要选择合适的阈值来定义将连接哪些顶点对。我们旨在通过建议和比较五种不同的阈值选择方法来解决这一重要问题。每种方法均考虑用于选择潜在合适阈值的各自的生物学动机标准。来自不同生物学组的一组21个微阵列实验用于研究将5个提议的标准应用于几种生物学情况的效果。对于每个实验，我们使用Pearson相关系数来衡量每个基因对之间的关​​系，并考虑多个值对所得的权重矩阵进行阈值处理，从而生成各自的邻接矩阵（共表达网络）。然后应用五个建议标准中的每一个，以选择相应的阈值。通过使用几次测量比较了这些阈值方法对生成的网络拓扑的影响，并且我们验证了根据数据库的不同，对拓扑属性的影响可能很大。但是，已验证一组数据库也受到大多数考虑标准的类似影响。基于这样的结果，可以建议，当生成的网络呈现相似的测量结果时，可以更大的自由度选择阈值方法。