Page 15 - Read Online

Page 4 of 12 Nishioka et al. Hepatoma Res 2018;4:1 I http://dx.doi.org/10.20517/2394-5079.2017.46

of FIB-4 scores was highly skewed and did not fit a normal or bimodal distribution to provide a logical
location for the cut-off point. A FIB-4 cut-off was therefore prospectively chosen based on review of
[16]
previous literature regarding FIB-4 scores and their prognostic value. A study conducted by Chan et al. ,
which aimed to determine an optimal cut-off point for diagnosing and prognosticating advanced liver
fibrosis after curative liver resection in HCC patients found that a FIB-4 index of 2.87 optimized both
sensitivity and specificity. As a result, samples were dichotomized based on a FIB-4 score of 2.87.

GSEA
GSEA was used to test the hypothesis that gene expression profiles corresponding to a priori defined gene
[25]
sets differ between samples belonging to 2 distinct phenotype classes . Using a Java-based implementation
of the GSEA algorithm (GSEA v3.0, Broad Institute, Boston, MA), the enrichment of gene sets of interest
within tumors corresponding to a given clinical phenotype were sought. To perform significance testing
against a null-hypothesis, permutation testing was performed to compute enrichment scores for 1000
random phenotype assignments. A FDR of less than 0.25 was used to indicate significant enrichment and
prompt further inquiry about tumor biology using biomedical literature referenced in the GSEA output.

Current versions (v6.0) of curated collections of gene sets were downloaded from an online database MSigDB
(MSigDB, Broad Institute, Boston, MA) from within the GSEA Java application. The Hallmarks collection
(comprised of 50 gene sets composed of coherently expressed genes reflecting well-defined biological states
or processes) and the chemical and genetic perturbations (CGP) collection (comprised of 2675 gene sets
reflecting gene signatures derived from published biomedical literature) were used for this study. The CGP
collection includes gene signatures reflecting genetic and chemical perturbations from a broad variety of
diseases. To estimate the number of HCC-related gene sets in the CGP collection, a query for “hepatocellular
carcinoma” was performed using the search mechanism of the mSigDB online portal (http://software.
broadinstitute.org/gsea/msigdb/index.jsp). This revealed 107 gene sets within the CGP collection related to
HCC that were supported by literature from Medline-indexed journals. These gene sets included multiple
published gene signatures for HCC molecular classification [8,10,11,26] and prognostication [12,27] .

Statistical methods
Differences involving normally distributed variables were assessed by t-test or analysis of variance. Post hoc
multiple pair wise comparisons were performed by the Steel-Dwass test. Comparisons among categorical
or dichotomized variables were assessed using Fisher’s exact test. Kaplan-Meier analysis was used to
compare overall survival rates post-surgical resection in patients stratified by AFP > 400 ng/mL and AFP ≤
400 ng/mL, and by combined S1 and S2 tumor subclasses vs. S3 subclass. Differences in survival curves
were assessed using the Log-Rank tests. Cox proportional hazard ratios were also computed for the effects
of AFP level differences and tumor class differences on overall survival post-surgical resection. Adjustments
to proportional hazards regression models were made only if multiple significant univariate predictors of
overall survival were identified. All statistical analyses were carried out using SAS version 9.3 (SAS Institute,
Cary, NC).

RESULTS
Patient clinical characteristics and demographics (n = 40) are summarized in Table 1. There were no
significant differences in various clinical parameters including age, gender, HBV infection, HCV infection,
significant alcohol use, Edmondson-Steiner grade, or proportion of high FIB4 scores between the AFP >
400 ng/mL and AFP ≤ 400 ng/mL groups of patients [Table 2].

Tumor classification
The number of tumors mapped into tumor sub-class S1, S2, and S3, were 12, 4, and 23 respectively. Only
one tumor could not be classified based on a FDR < 0.05. The remaining sub-class assignments were also