Level 2: Interaction Networks
SWAN 35
SWAN incorporates prior knowledge network into the cutoff selection process for correlation networks. This is a hybrid inference/extraction algorithm that redefines the inference task as defining a cutoff threshold such that the agreement with prior information is maximized.
SWAN works by first constructing a correlation network from the data. The network is then filtered to remove edges that are not statistically significant. The remaining edges are then ranked according to their strength. Prior interaction networks can be easily integrated with inferred correlation networks. SWAN then selects a cutoff for the network based on prior knowledge. To calculate the correlation, SWAN uses shrinkage partial correlation based on the GeneNet algorithm – although this approach can be generalized to any correlation metric. The overlap is measured using Fisher’s exact test p-value, which indicates the agreement between the calculated correlation network and prior knowledge. The optimal cutoff is defined as the point where the overlap is maximal.
SWAN was tested on pan-cancer data of 26 cancer types extracted from The Cancer Genome Atlas (TCGA). The network was able to identify enriched genes (OG) in the elevated pathways and suppressed genes (TSG) within suppressed pathways with a p-value < 0.05. This result was compared with the Gene Set Enrichment Analysis (GSEA) and revealed that SWAN outperformed GSEA. To check if SWAN can study race-specific CNA patterns, ovarian cancer samples from an African American population were collected, and non-Hispanic white patients were used as control. SWAN identified that the cytokine pathway was elevated in the former population which can be mapped to the overall poor prognosis in these patients. Furthermore, SWAN was also able to figure out the effect of the knockdown of metallothionein 2A which led to an increase in formation of ɤH2AX foci.
GLRP 36
Graph Layer-wise Relevance Propagation (GLRP) is a novel method that extends the Layer-wise Relevance Propagation (LRP) technique to Graph Convolutional Neural Networks (Graph-CNN). LRP is an existing technique that explains the decisions made by deep learning models. The primary goal of GLRP is to explain the classification results of various omics data and molecular networks which could facilitate the decision-making processes in personalized medicine.
This is a unimodal, hybrid phenotype prediction and explanation/extraction algorithm that aims to ground predicted graphs to known protein-protein interaction networks. GLRP interprets the classification output by leveraging the molecular network and also produces patient-specific subnetworks that can be used to explain clinical outcomes and therapeutic vulnerabilities.
GLRP was trained on gene expression datasets of breast cancer and human umbilical vein endothelial cells (HUVECs). Their predictive performance was evaluated using the 10-fold cross-validation method. In the breast cancer study, GLRP was used to classify patients into metastatic and non-metastatic groups. The results were compared with the classification performance of random forest and glmgraph models as well as weighted gene co-expression network analysis. GLRP outperformed the other models, and the developed patient-specific subnetworks identified meaningful features in breast cancer samples.