Table 1.
Clinical characteristics of patients in the LCI and LEC cohorts at the time of surgery.
Figure 1.
Schematic overview of the iSubgraph algorithm.
The flow chart illustrates the transformation of microarray data into graph representation (left); followed by the graph mining-based method to identify significant miRNA-gene co-modules (top right) and tumor subclassification by the mixture model (bottom right).
Figure 2.
Graph Mining steps for a small dataset.
(A) Normalized gene and miRNA expression levels in tumor and adjacent nontumor tissues. (B) Steps of correlation analysis to decide actual target genes of miRNAs. (C) Template bipartite graph representing miRNA-gene interactions. (D) Graph patients constructed for all patients. Threshold for UP and DOWN tags is . (E) All closed frequent subgraphs for a support threshold of 2 patients.
Figure 3.
Mixture model for cancer subgroup discovery using plate notation.
Circles indicate random variables: Subgroup of patient, , gene expression level,
, and miRNA expression level,
. Shaded circles denote observed values. Outer rectangles indicate fixed paramters: Class mixing parameter,
, the mean expression level,
, and shared standard deviation,
. Directed edges show dependencies between variables and parameters. Capital letters represent the size of parameters (vector or matrix) and plate repetition.
Figure 4.
Schematic overview of the study design.
Significant genes and miRNAs for HCC were identified separately for the LCI and Hong Kong cohorts using the graph mining approach of iSubgraph. The included genes and miRNAs were used as features in the subsequent clustering step of iSubgraph for the LCI and LEC cohorts. Finally, computational and biological analyses were carried out on the selected genes/miRNAs and cancer subgroups.
Figure 5.
Frequency and graph size density of frequent subgraphs.
(A) Density histograms of the frequent subgraphs from the LCI cohort () and (B) Hong Kong cohort (
) with respect to frequency and graph size. The frequency of a subgraph equals to the number of patients who have that subgraph in their patient graph.
Table 2.
Number of genes and miRNAs after each graph mining step.
Figure 6.
Overlapping genes and miRNAs in subgraphs of the LCI and Hong Kong cohorts.
(A) Euler Diagrams showing the overlap between the genes and (B) miRNAs occurring in frequent subgraphs from the LCI and Hong Kong datasets. A single miRNA probe in the Rosetta platform (Hong Kong cohort) may correspond multiple probes in the OSU-CCC platform (LCI cohort). The numbers in parentheses show the probe counts in the LCI data. P-values were computed from hypergeometric tests. The number of common genes and miRNAs indicate solely those transcripts that were present in the TargetScan database.
Figure 7.
Regulatory Networks associated with HCC.
(A) MiRNA-mRNA networks from the LCI and (B) Hong Kong datasets. The networks were constructed by merging vertices and edges of all frequent subgraphs from those datasets. The node color represents the transcript and regulation type. The node size indicates the number of connections. Unlike the current layout of the nodes for visualization purposes, both networks are originally bipartite graphs.
Figure 8.
Clinical properties and heatmap of subgraph occurrence in the LCI cohort.
(A) Occurrence of all subgraphs identified in the LCI and (B) Hong Kong cohorts. The order of patients in subgraph occurrence heatmap (bottom) is arranged according to class assignments (top) and clinical properties (middle). Subgraph occurrence in a patient is defined as the ratio of common nodes between the subgraph and the patient graph to all nodes of that subgraph. Abbreviations: AFP, Alpha-fetoprotein; BCLC, Barcelona clinic liver cancer.
Figure 9.
Combined Kaplan-Meier curves of patients subgrouped using the LCI subgraphs.
The first row shows the survival and recurrence characteristics of the LCI cohort () and the second row shows those of the LEC cohort (
). The recurrence information was not available for some patients of the LEC cohort. From left to right, the columns indicate the overall survival (survival rates in the first 5 years), early survival (survival rates in the first 2 years), overall recurrence (disease-free survival rates in the first 5 years), and early recurrence (disease-free survival rates in the first 2 years) curves.
-values were calculated by the Log-rank (Mantel-Cox) test.
Table 3.
The most significant upstream regulators and corresponding -values calculated by Ingenuity Pathway Analysis.
Table 4.
The most significant canonical pathways and corresponding -values calculated by Ingenuity Pathway Analysis.
Figure 10.
Kaplan-Meier curves of the LCI cohort subgrouped by SNMNMF modules.
Survival curves in the first 5 years are shown for three different subgrouping of the LCI cohort (). The patients are divided into two equal size groups with respect to their signals in the co-module basis vectors. Shown are the three co-modules which exhibit the most significant differences in their clinical parameters among all co-modules.
-values were calculated by the Log-rank (Mantel-Cox) test.