Figure 1.
SANTA can be used both to quantify the strength of association between networks and sets of node weights (using Knet) and to prioritise genes for follow-up analyses (using Knode). Different node colour intensities represent different node weights.
Figure 2.
Application of the Knet-function to two gene sets.
Example input: (A) S. cerevisiae GI map and (B) gene sets obtained from GO (‘GO:0044451: nucleoplasm part’ and ‘GO:0070011: peptidase activity’). (C, D) Network annotated with each gene set. From visual inspection, it appears that the gene set in (C) clusters more significantly than the gene set in (D). SANTA allows us to assess this clustering objectively. (E, F) The K net-function is computed for the observed gene sets (red and blue lines) and for a large number of permutations (yellow area). (G, H) In order to quantify the significance of the clustering, the area under the Knet-function curve (AUK) is computed for the observed gene set (red and blue lines) and for each permutation (grey histogram). An empirical p-value is calculated using a Z-test. For GO:0044451, p = 5.680×10−30 and for GO:0070011, p = 0.174, demonstrating objectively that the gene set in (C) does cluster more significantly.
Figure 3.
Application of Knetto simulated networks.
Scale-free networks containing clusters of high-weight nodes of various strengths were generated. The smaller the distance cutoff used to generate the cluster, the greater the strength of the clustering. (A) 1000 trials were completed for each distance cutoff. As expected, the most significant clustering was measured by the Knet-function when smaller distance cutoffs were used. (B) Q-Q plot of the p-values observed in the simulation study trials in which no distance cutoff was used and the p-values expected under the uniform distribution. The high-weight nodes were distributed homogeneously when no distance cutoff was used. The observed p-values deviate little from the expected p-values, demonstrating that the Knet-function does not detect clustering when clustering is not present.
Figure 4.
Comparison of Knet and Compactness.
Example of the difference between the Knet and the Compactness functions. Red circles represent hits on the network. P-values were computed for both functions using 1000 permutations. Only the Knet-function incorporates the global structure of the network and therefore only it identifies a more significant association between set 2 and the network.
Figure 5.
Comparison of Knode with BioNet.
Comparison of the ability of the Knode-function and BioNet to identify high-weight nodes contained within multiple clusters on a single simulated network. Across 1000 trials, the Knode-function identified a greater proportion of the high-weight nodes when they were distributed across 3 or 4 clusters.
Figure 6.
Applications of Knet to real networks.
(A) Comparison of the functional content of a network of raw GIs and a network representing correlation in GI profile. GO terms are associated more strongly with the GI-correlation network, indicating that this network is functionally more informative. (B) Comparison of the functional content of the untreated and MMS-treated GI networks. GO terms associated with the response to DNA damage were enriched within the treated network. (C) Comparison of the functional content of the untreated and UV-treated GI networks. GO terms associated with cell cycle progression were enriched within the treated network.
Figure 7.
Knet identifies the most functionally informative network.
Association of genes essential in the proliferation of cancer cell lines with a network of curated physical interactions (IntAct) and a functional network created using 21 data sources (HumanNet). Association was stronger between colon and ovarian cancer cell line RNAi hits and the functional network, indicating that the functional network provides more information about the mechanisms that drive cancer cell line maintenance.