Fig. 4

Degree bias adjustment analysis: combinations of transition matrix types and degree bias adjustment methods compared on 3 tasks: ranking of functionally related proteins, correlation between diffusion scores and degree, and retention of high-seed-value nodes in the context of AMEND. For all figures, the restart probability is set to 0.5 for RWR. A–C Empirical cumulative probabilities of ranks. Using Reactome Mus musculus pathways and a Mus musculus PPI network of physical and functional interactions, a fivefold cross validation procedure was implemented in which pathways are split into folds, nodes belonging to the training folds of a pathway are used as seeds in RWR, and the ranks of diffusion scores of nodes in the test fold are collected (higher score equals lower rank). Ranks shown in figures include ranks from all test folds from all pathways. D Correlation between diffusion scores and node degree. For each of 5 human gene expression datasets, -log₁₀-transformed p-values from differential expression analysis were diffused on a human PPI network of functional and physical interactions. Pearson correlation coefficients were averaged across datasets. E Average difference in empirical cumulative probabilities between seed values and degree for module nodes returned from AMEND. For each of 5 human gene expression datasets, seed values are -log₁₀-transformed p-values from differential expression analyses assigned to nodes such that seed value and degree are perfectly negatively correlated. Results are averaged across datasets