Prognostic Value of an Integrin-Based Signature in Hepatocellular Carcinoma and the Identification of Immunological Role of LIMS2

Objective Evidence proves that integrins affect almost every step of hepatocellular carcinoma (HCC) progression. The current study aimed at constructing an integrin-based signature for prognostic prediction of HCC. Methods TCGA-LIHC and ICGC-LIRI-JP cohorts were retrospectively analyzed. Integrin genes were analyzed via univariate Cox regression, followed by generation of a prognostic signature with LASSO approach. Independent factors were input into the nomogram. WGCNA was adopted to select this signature-specific genes. Gene Ontology (GO) enrichment together with Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway analysis were conducted to explore the function of the dysregulated genes. The abundance of tumor microenvironment components was estimated with diverse popular computational methods. The relative importance of genes from this signature was estimated through random-forest method. Results Eight integrin genes (ADAM15, CDC42, DAB2, ITGB1BP1, ITGB5, KIF14, LIMS2, and SELP) were adopted to define an integrin-based signature. Each patient was assigned the riskScore. High-riskScore subpopulation exhibited worse overall survival, with satisfying prediction efficacy. Also, the integrin-based signature was independent of routine clinicopathological parameters. The nomogram (comprising integrin-based signature, and stage) accurately inferred prognostic outcome, with the excellent net benefit. Genes with the strongest positive interaction to low-riskScore were primarily linked to biosynthetic, metabolic, and catabolic processes and immune pathways; those with the strongest association with high-riskScore were principally associated with diverse tumorigenic signaling. The integrin-based signature was strongly linked with tumor microenvironment components. Among the genes from this signature, LIMS2 possessed the highest importance, and its expression was proven through immunohistochemical staining. Conclusion Altogether, our study defined a quantitative integrin-based signature that reliably assessed HCC prognosis and tumor microenvironment features, which possessed the potential as a tool for prognostic prediction.


Introduction
Liver cancer remains the sixth most commonly diagnosed cancer together with the third most deadly cancer, with an estimated 906,000 new cases as well as 830,000 deaths globally [1]. Asian countries have the highest incidence of primary liver cancer cases, reporting approximately 72.5% of the world's cases [2]. Hepatocellular carcinoma (HCC) contributes 80% of all cases worldwide [3]. Elderly male together with Asian populations are still the highest risk groups for HCC. The preferred therapy of HCC remains surgery, which is the only method to achieve long-term survival and even a cure [4]. Radiofrequency ablation is the treatment of choice for malignancies that are extremely early in their stages as well as tumors that are early in their stages but cannot be removed surgically [5]. Ultrasound is well poised to address this need due to its low cost, portability, safety, and excellent temporal resolution. The role of ultrasound for HCC screening has been well established and supported by multiple international guidelines. Nonetheless, HCC patients are generally in intermediate or advanced stages. Transcatheter arterial chemoembolization is the standard of care for patients with intermediate HCC, resulting to the median survival of 25-30 months [6]. Molecular-targeted agents, sorafenib, etc. have been developed against advanced HCC [7]. Regrettably, liver toxicity and weak anti-tumor effect contribute to treatment failure and low survival benefit. Recently, immune-based therapies have generated notable improvement in clinical outcome of HCC [8]. Despite this, current immunotherapies only induce durable response for minority of HCC patients. Altogether, it is of significance to select potent therapeutic targets, and determine more reliable tools for stratifying HCC patients together with prognostic prediction.
Integrins cross the plasma membrane as well as connect the extracellular matrix (ECM) to the cytoskeleton, as elementary cell adhesion receptors mediating cellular and tissue functions [9]. Altered expression of integrins is commonly detected in HCC [10]. They profoundly affect almost every step of HCC progression from primary tumor development to metastases [10,11]. Additionally, integrins correlate to the acquisition of drug resistance and immune escape [12,13]. In addition to tumor cells, integrins are present in components within tumor microenvironment, which critically regulate their contributions to tumor progression [14]. For instance, SPON2 facilitated the recruitment of M1-like macrophages as well as mitigates HCC metastases through integrin signaling [10]. Cancer-associated fibroblasts facilitate vascular invasion of HCC through lowering integrin β1 [15]. Blockade of integrin signaling can attenuate HCC progression through hindering key signaling events in tumor microenvironment and tumor cells. Hence, integrins together with integrin-dependent functions have been regarded as attractive therapeutic targets against HCC [16]. In addition to this, integrins may become imaging biomarkers for evaluating the efficacy of anti-angiogenic or antitumor agents [17]. Moreover, integrin-targeted nanoparticles with varying anti-tumoral payloads are a definitely promising research field to lower toxicity linked to systemic radio-or chemotherapy [18]. To date, the now prognostic model based on integrin-related genes were rarely reported. Based on accumulated evidence, the current study conducted a comprehensive analysis of multidimensional integrin-relevant genomic data across HCC, and defined a quantitative integrin-based signature that may evaluate HCC prognostic outcome together with tumor microenvironment traits, which might open up a novel insight into improving HCC outcomes together with determining patients' therapeutic regimens.

Materials and Methods
2.1. Data Acquisition. Transcriptome data and clinical information of HCC patients were acquired from TCGA-LIHC as the training cohort. Under removal of patients with incomplete survival data, 343 HCC patients were included. Another RNA-seq dataset ICGC-LIRI-JP with 229 HCC samples obtained from the ICGC database were adopted for verification.

Collection of Integrin Gene
Set. The Molecular Signatures Database offers the annotated gene sets that involve biochemical pathways, signaling cascades, and expression profiling from published research together with other biological concepts [19]. We collected 128 integrin genes from this popular database, involving four biological process terms (integrin activation, integrin-mediated signaling pathway, positive regulation of integrin activation, and regulation of integrin activation) together with one cellular component terms (integrin complex).

Definition of an
Integrin-Based Signature. Prognostic significance of integrin genes was firstly evaluated. Through adopting univariate-cox regression approach, integrin genes with p < 0:05 were selected, and input into least absolute shrinkage and selection operator (LASSO) [20]. This analysis was conducted utilizing glmnet package [21]. The regression coefficient was computed utilizing multivariate-cox regression. The integrin-based signature-derived riskScore was generated through combination of regression coefficient together with transcript level of each integrin gene in this signature. With the median riskScore, patients were classified as low-and high-riskScore subpopulations. This classification was verified through PCA and tSNE approaches. Overall survival (OS) analysis was implemented with Kaplan-Meier (K-M) method together with log-rank test. Area under the receiver operating characteristic curve (AUC) was computed with "timeROC" package. Unitogether with multivariate-cox regression methods were adopted for inferring the independency of the integrinbased signature as a prognostic parameter. Through the use of subgroup analysis, we were able to deduce the sensitivity of this signature in prognostic prediction.

Nomogram Construction.
Nomogram was generated through incorporating independent risky factors (riskScore and stage) via adopting rms package. ROC curves were utilized for reflecting the predictive capability of the nomogram. Concordance index (C-index) was employed to estimate the nomogram discrimination through bootstrap approach with 1000 resamples. Calibration curve was graphically assessed through drawing the actual OS rate against the probability predicted by this nomogram, with the 45°l ine for the ideal prediction. Decision curve analysis was employed to evaluate the net benefit of the nomogram, routine clinicopathological parameters, and riskScore.

Weighted Gene Coexpression Network Analysis
(WGCNA). WGCNA package [22] was adopted to select the riskScore-specific modules. The transcriptome profiling was utilized as input for WGCNA, and riskScore was computed as well as defined as the clinical traits. A signed scale-free coexpression gene network was guaranteed via setting an appropriate power β value and scale-free R 2 value as the soft threshold parameters. Afterwards, we constructed a coexpression matrix in accordance with β value, and the input gene expression matrix for classifying genes with similar expression pattern into the same gene module, thus producing a coexpression module. Association of module 2 Disease Markers Eigengenes with riskScore were estimated with Eigengenes function. Heatmap was generated for visualizing the association of each coexpression module with riskScore. Modules with the strongest association with riskScore were selected as the riskScore-specific modules.
2.6. Functional Enrichment Analysis. Gene Ontology (GO) enrichment together with Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway analysis were conducted through adopting clusterProfiler package [23]. For preventing high false discovery rate (FDR) in multiple tests, q -value was inferred for FDR control. A gene set was regarded as significantly enriched if a p < 0:05 and false discovery rate <0.025.

Estimation of Tumor Microenvironment Components.
Seven computational approaches were employed to infer components within tumor microenvironment. Tumor Immune Estimation Resource (TIMER) adopts deconvolution approach to estimate the level of six tumor-infiltrating immune subsets from gene expression profiling [24]. CIBERSORT applies transcriptome profiles with a predefined immune signature matrix to calculate the deconvolution of 22 tumor-infiltrating immune cells in a given sample on the basis of support-vector regression [25]. quant-TIseq quantifies the fraction of 10 immune cell types utilizing bulk RNA-sequencing data [26]. MCPcounter quantifies the absolute abundance of 8 immune together with 2 stromal cell subsets within heterogeneous tissues through transcriptomic profiles [27]. XCELL infers 64 immune, and stromal cell types via adopting gene signature-based approach [28]. EPIC estimates the proportion of immune and cancer cells utilizing bulk gene expression profiling [29]. Associations of riskScore and genes in the integrin-based signature with the abundance of tumor microenvironment components were estimated with Spearman's correlation test.
2.8. Random-Forest Analysis. The relative importance of genes in the integrin-based signature was ranked via implementing random-forest analysis, and the gene with the highest importance was determined. LIMS2 transcript level was compared between low-and high-riskScore subpopulations. Association of LIMS2 transcript level with riskScore was inferred with Spearman's correlation test. Immunohistochemical staining of LIMS2 in HCC and normal tissue was acquired from the Human Protein Atlas.
As a result, low-riskScore patients signally distinguished from those with high-riskScore at the transcriptome level (Figures 3(c) and 3(d)).
Next, the current study observed whether the integrinbased signature generalized to other cohorts. Similarly, we computed riskScore of patients from ICGC-LIRI-JP cohort, which were then classified as low-and high-riskScore subpopulations. As expected, poorer OS outcome was proven in high-riskScore subpopulation (Figure 3(e)). In addition, AUC values of OS at one, two together with three years were all over 0.7 (Figure 3(f)). PCA and tSNE demonstrated the arresting discrepancy between subpopulations (Figures 3(g) and 3(h)).
The integrin-based signature is independent of routine clinicopathological parameters.

Disease Markers
Next, the present study estimated the associations of riskScore and routine clinicopathological parameters with OS outcome across TCGA-LIHC utilizing univariate-cox regression approach. Consequently, riskScore together with stage were linked with poor OS outcome (Figure 4(a)).
Multivariate-cox regression approach was adopted to infer whether riskScore was independent of routine clinicopathological parameters. As illustrated in Figure 4

Generation of an
Integrin-Based Signature-and Stage-Based Nomogram into HCC Clinical Practice. Two independent risky factors (riskScore together with stage) were selected for generating a nomogram for HCC prognostic prediction ( Figure 5(a)). Firstly, points for riskScore and stage were derived in TCGA-LIHC cases. Total points were acquired through adding the points of two risky factors, and the corresponding location of the point of each patient was observed in the line of total points. At last, the probability of one-, three-together with five-year OS for HCC was referred through plotting a straight line on the bottom three rows. ROC curves and C-indices were adopted for evaluating the prediction accuracy of the nomogram. AUC values of OS at one, two together with three years were all over 0.7 ( Figure 5(b)), and the C-indices were over 0.7 for shortand long-term OS outcomes ( Figure 5(c)). In addition, calibration curve illustrated that the one-, three-together with five-year OS probability predicted by this nomogram was consistent with the actual OS rate ( Figure 5(d)). Above evidence proved the excellent prediction efficacy of this nomogram. Decision curve analysis curves at one-, three-together with five-year OS displayed the potential for clinical application as well as better net benefits (Figures 5(e)-5(g)).

Selection of Integrin-Based Signature-Specific Genes.
WGCNA was employed for identifying integrin-based signature-specific genes across TCGA-LIHC. Transcriptome data and clinical trait (low-and high-riskScore) were input into WGNCA. The first power value when the index of scale-free topologies was up to 0.90 was set as the optimal soft threshold power (β) for establishing a scale-free network, and genes with similar expression patterns were assigned to the same coexpression module utilizing dynamic tree cut approach, thus generating 12 coexpression modules (Figures 6(a)-6(c)). Afterwards, associations of coexpression modules with low-and high-riskScore were evaluated. Tur-quoise module exhibited the strongest positive interaction to low-riskScore ( Figure 6(d)). In addition, yellow module displayed the strongest positive association with high-riskScore. Thus, genes in turquoise and yellow modules were regarded as integrin-based signature-specific genes.
Next, biological implication of integrin-based signaturespecific genes was assessed. Genes in turquoise module were primarily linked to biosynthetic, metabolic, and catabolic processes together with immune pathways (Figures 6(e) and 6(f)). Genes in yellow module primarily correlated to diverse tumorigenic signaling (Figures 6(g) and 6(h)).

Interactions of the Integrin-Based Signature with
Components within Tumor Microenvironment. Diverse computational approaches were adopted for inferring the interactions of the integrin-based signature with components within tumor microenvironment across TCGA-LIHC. Overall, high-riskScore exhibited higher abundance of immunosuppressive cells, and the riskScore was positively correlated to immunosuppressive cells (Figures 7(a) and 7(b)). In addition, genes from the integrin-based signature (ADAM15, CDC42, DAB2, ITGB1BP1, ITGB5, KIF14, SELP, and LIMS2) were strongly linked with the abundance of components within tumor microenvironment (Figures 7(c)-7(i)).
3.6. The Importance of LIMS2 from the Integrin-Based Signature in HCC. Random-forest approach was adopting for assessing the relative importance of genes in the integrin-based signature. Consequently, LIMS2 possessed the highest importance (Figure 8(a)). In contrast to low-riskScore subpopulation, lower transcript level of LIMS2 was observed in high-riskScore subpopulation (Figure 8(b)). In addition, LIMS2 transcript level was negatively linked to riskScore (Figure 8(c)). Immunohistochemical staining demonstrated that LIMS2 protein displayed low expression level in normal tissue, without detection in HCC tissue (Figures 8(d) and 8(e)). Finally, we performed RT-PCR and found that LIMS2 expression was distinctly decreased in HCC specimens compared with nontumor specimens (Figure 8(f)

Discussion
Despite the notable improvement in HCC research, patients' outcome remains depressing [30]. Hence, it is imperative to search for novel tools for HCC prognostic prediction. Evidence demonstrates that integrins affect almost every step of HCC progression [10]. Herein, eight integrin genes (ADAM15, CDC42, DAB2, ITGB1BP1, ITGB5, KIF14, LIMS2, and SELP) were selected and adopted to define an integrin-based signature. High-riskScore subpopulation displayed worse OS, with satisfying prediction efficacy. In addition, the integrin-based signature was independent of routine clinicopathological parameters. To facilitate clinical practice, we produced the integrin-based signature-and stage-based nomogram that accurately inferred prognostic outcome, with the excellent net benefit.
Accumulated evidence proves the significance of genes from the integrin-based signature in HCC. For instance, ADAM15 metalloproteinase, a multidomain disintegrin protease, is linked to prognostic outcome, infiltration of immune cells together with apoptosis in HCC [31]. CDC42 stimulates tumor growth, angiogenesis together with metastatic potential of HCC [32]. DAB2 mitigates tumor growth and metastasis of HCC [33]. ITGB1BP1 induces HCC