Fig 1.
A diagram of the conceptual entities of the proposed fusion (model guided co-expression networks) combining two schemes—parametric interaction models and co-expression networks. The most central methodological benefits (green checkmarks) and deficiencies (red X marks) are listed method-specifically.
Fig 2.
A schematic representation of the logical structure underlying the provided algorithmic implementation. The first panel from the top represents the empirical density function of the residual vector estimated from the main effect model. Here Q(⋅) is a quantile function of this distribution and a ∈ [0.5, 1[ is a user-defined cut-off point. The second panel from the top illustrates how the differential networks are computed from the low and high networks using the sign-adjusted version. Different edge colors represent the edge weights of opposite signs, but equal magnitudes, for simplicity. The bottom panel demonstrates how different sign-combinations can be interpreted as type I or II interactions in the underlying parametric model. Moreover, the most important arguments of the provided model_diffnet R-function are explained in the right panel to indicate their role in this procedural flow.
Table 1.
Averaged areas under the truncated and non-truncated ROC curves (AUCs) over ten replicates in the simulated scenarios without simulated main effects (Model A) and with additional main effects (Model B). These datasets are analyzed using the proposed dPCCN procedure and dCCN method as well as the exhaustive search and GGM model based approaches as reference methods.
Fig 3.
Estimated differential networks.
Estimated differential networks with the proposed residual- and sign-adjusted dPCCN approach in the DREAM9-challenge protein expression dataset using the patients’ non-censored survival times (142 observations) as response. The interaction types I and II are separated by green (type I) and red (type II) edges (plotted separately). A hard-thresholding was used to provide sparsity into the network structures such that the number of network edges is less than 70 for both interaction types. The estimated network structures are displayed only for connected nodes (with MiMI names) and the highlighted nodes indicate which network structures are discussed in detail.
Fig 4.
Survival analysis using type I interactions.
Kaplan-Meier curves and the 95% confidence intervals for three different classifiers in the TCGA dataset (173 observations). The first two panels represent the (AKT1, CCNE1) interaction term based classifier and the (ASNS, GAB2) interaction term based classifier. Red and blue survival curves correspond to high and low expression values. The last panel is a combined classifier; the low-risk group patients based on both interaction terms (AKT1, CCNE1) and (ASNS, GAB2) are classified into the combined low-risk group (purple curve) and into the combined high-risk group (orange curve) with the same logic. The p-values of the corresponding log-rank tests are also reported.
Fig 5.
Survival analysis using type II interactions.
Results from survival analysis including survival curves (and 95% confidence intervals) and statistics using GEPIA software with 20% and 80% cut-off points to classify high- and low expression ratio groups (21 individuals in both groups). In each panel, p-values of the log-rank tests are reported for hypothesis tests of no differences between groups.