Raw single-cell transcriptome profiling data for ten HCC patients from two relevant sites, primary tumor (HCC01T, HCC02T, HCC03T, HCC04T, HCC05T, HCC06T, HCC07T, HCC08T, HCC09T, and HCC10T) and adjacent non-tumor liver (HCC03N, HCC04N, HCC05N, HCC06N, HCC07N, HCC08N, HCC09N, and HCC10N), was achieved from GSE149614 dataset in Gene Expression Omnibus (GEO) database (https://www.ncbi.nlm.nih.gov/geo/). Bulk RNA-seq data of HCC patients were collected as follows, TCGA-LIHC dataset from The Cancer Genome Atlas (https://portal.gdc.cancer.gov/repository) database, HCCDB18 dataset from Integrative Molecular Database of Hepatocellular Carcinoma (http://lifeome.net/database/hccdb/download.html) database, and GSE datasets (GSE76427, GSE54236, GSE36376, GSE69715, GSE121248, GSE107170, and GSE45267) from GEO database. 110 normal liver samples RNA-seq data were retrieved from Genotype-Tissue Expression (GTEx, https://gtexportal.org/home/) database. The samples information in all data sets can be seen in Supplementary Table S1.

We used “Seurat” packages (Hao et al., 2021) in R software to process the single cell data. Gene number, relative hemoglobin, and mitochondrial and ribosomal abundance (Supplementary Figure S1), indicating that the cellular readouts were comparable between samples and no transcriptional batch effects were observed. Cells with <1,500 or >20,000 detected genes containing mitochondrial genome >5% were excluded. Indicating that the cellular readouts were comparable between samples and no transcriptional batch effects were observed. Cells with <2,500 or >20,000 detected genes containing mitochondrial genome >4% were excluded. Next, single-cell data were normalized, and variable genes were hunted by the “SCTransform” method. After, 20 most powerful principal components were found by PCA analysis (Supplementary Figure S2). Further dimension reduction of those principal components was proceeded by the UMAP method to visualize cell distribution. Cell types of principal components were annotated by “CellMarker” (http://xteam.xbio.top/CellMarker/) and “PanglaoDB” (https://panglaodb.se/) databases. Cell-cell communication were conducted by “CellChat” R package. In addition, Bulk RNA-sequencing expression datasets were normalized by log2 (FPKM+1) through “limma” package in R.

The verified pathways containing gene sets provided in the Molecular Signatures database were utilized by Gene Set Variation Analysis (GSEA) (Mo et al., 2003) to analyze the biological functions of single-cell data. Gene-related differential expression biological processes in transcriptome data were analyzed by single-sample Gene Set Variation Analysis (ssGSEA) (Rooney et al., 2015; Zhang et al., 2021b). Metascape (a gene annotation and analysis resource, https://metascape.org/) and molecular complex detection (MCODE) (Osterman et al., 2020) methods analyzed oxidative stress-related genes’ interactions and functional pathways.

CellChat (Jin et al., 2021) toolkit in R enables the analysis and interpretation of cell-cell communication within complex biological systems. Using CellChat, the gene expression profiles are analyzed to identify ligand-receptor pairs between different cell types. CellChat provides a comprehensive database of known ligand-receptor interactions, which is used to identify potential communication channels between cell types. It evaluates the significance of ligand-receptor pairs in cell-cell interactions using correlation analysis (Cowen et al., 2017; Ronen and Akalin, 2018; Szklarczyk et al., 2019). Once the ligand-receptor pairs are identified, CellChat analyzes the signaling pathways associated with these interactions. We utilize information from the secreted signaling pathways database (Jin et al., 2020) in CellChat to determine the ligand-receptor interactions in intercellular communication.

Due to the relatively small number of non-tumor liver tissue samples in TCGA-LIHC dataset compared to HCC samples, we used “Sva” package to merge the normalized GTEX dataset with the TCGA dataset and remove batch effects. The two datasets after removal of batch effects are shown in Supplementary Figure S3. Model building was based on the oxidative stress-related genes expression profile of merged dataset. The samples in the merged dataset were randomly split into training sets and test sets in the proportion 7:3, respectively. First, we applied machine learning algorithms, including extreme gradient boosting (XGB) (Li et al., 2023), support vector machine (SVM) (Huang et al., 2023), generalized linear model (GLM) (Wang et al., 2023), and random forest (RF) (Su et al., 2023), to create models that screened for significant oxidative stress-related gene feature variables. All machine learning methods used default settings to generate models. The grid search function in the “caret” R package was used to modify model parameters automatically, which were tested using 5-fold cross-validation. Secondly, we interpreted the four machine learning models obtained above using the explaining features function of the “Dalex” R package. The explaining features, including model performance analysis and variable importance analysis, were then used to analyze the four models described above. Model performance analysis visualized the performance distribution of the four models. The models’ cumulative residual and box plot distribution diagrams were drawn, respectively. The variable importance analyzes the relative importance of different variables in the model to the model’s prediction. It explains the impact of missing the variable on the prediction value of the response variable through the root mean square error loss function. Thirdly, the area under the curve was obtained by receiver operating characteristic curves (ROC) analysis to identify the best machine-learning model. The top six significant variable features from the optimal model were considered diagnostic gene markers for predicting the occurrence of HCC. Finally, for clinical usability, we used these diagnostic gene markers to conduct nomogram model construction and to judge their prediction error from the actual situation through the calibration curves. Additionally, the prediction performance of the diagnostic gene markers was examined by seven external independent cohorts, including GSE76427, GSE54236, GSE36376, GSE69715, GSE121248, GSE107170, and GSE45267.

We further analyzed six oxidative stress-related gene markers (GPX4, HMOX1, PRDX1, FOS, PRDX5, and TXN) with the diagnostic value, which from the machine learning analysis steps. To judge whether they also play a role in predicting the prognosis of HCC patients. Univariate and multivariate Cox regression analyses were performed for these six genes based on survival and gene expression information of HCC patients in the TCGA cohort. After screening, PRDX1 was correlated with the prognostic survival status of HCC patients. The patients with HCC were then classified into high and low-expression groups by median cut-off value, using Kaplan-Meier survival estimate (KM) and time-dependent ROC analysis to assess whether PRDX1 was a prognostic factor. The survival analysis results were finally validated using an independent HCCDB18 cohort.

To explore the correlation of PRDX1 with the immune microenvironment in HCC. We used six algorithms [CIBERSORT (Chen et al., 2018), CIBERSOR-ABS (Wang et al., 2020), QUANTISEQ (Plattner et al., 2020), MCP-counter (Zhang et al., 2021a), XCELL (Aran et al., 2017), and TIMER (Li et al., 2017)] for immune cell infiltration analyses to assess HCC individuals with the different expression level of PRDX1. Differences in immune function between different PRDX1 expression groups were then analyzed using the “GSVA” package.

The analyses of our study were all employed by R 4.1.1 software. Analyses used a false discovery rate (FDR) < 0.05 from the p-value, which was adjusted by the Bonferroni method. Figure 1 displays a flow chart outlining our study.

The GSE149614 dataset was utilized for single-cell analysis in our study. After quality control for scRNA-seq data, we filtered out 28,687 cells in tumor-adjacent noncancerous tissues (control) and 34,414 cells in HCC tissues for further study. We characterized the cellular landscape using cell classification marker genes identified in previous studies (Ronen and Akalin, 2018) (Supplementary Figure S4). Following, seven cell types, including dendritic cells, gamma delta T cells, mucosal−associated invariant T cells, T memory cells, B cells, and hepatocytes or malignant hepatocyte cells, were identified from the landscape (Figure 2A). When we compared cell type proportions (Figure 2B) between non-tumor liver tissues and HCC tissues, we discovered that the ratio of mucosal-associated invariant T cell, T memory cell, and gamma delta T cell in HCC tissues were significantly decreased. In contrast, the proportion of B cells increased in HCC tissues.

To investigate the oxidative stress characteristics of non-malignant samples and HCC samples, oxidative stress AUC (oxidative stress scoring for each cell) was generated based on the expression profile of genes in oxidative stress response collected from the Molecular Signatures database, which summarized genes involved in oxidative stress from validated experimental studies (Supplementary Table S2). The AUCell R package was used to determine each cell line’s oxidative stress AUC activity based on the differential expression level of oxidative stress-related genes in each cell type. Cells expressing higher oxidative stress AUC values in HCC samples were mainly in T memory cells, dendritic cells, and malignant hepatocytes, colored in yellow (Figures 3A, B). As mucosal-associated invariant T cell, T memory cell, and gamma delta T cell clusters were remarkably decreased in the non-tumor group, colored in purple. It can be seen that the oxidative stress AUC of the cells was generally higher in the HCC samples than in the non-tumor liver tissues of the control group in the boxplot (Figure 3C). This means that cells in the HCC microenvironment experienced more severe oxidative stress than non-tumor tissues.

Gene set enrichment analyses were performed to explore the biological processes changes of these cell clusters. In the group of HCC, antigen processing via the MHC class II route and extrinsic apoptosis of T memory cells were triggered. Immune cell receptor pathways were blocked in T memory cells Supplementary Figure S5A). In mucosal-associated invariant T cells from the tumor group, peptidase activity, catabolic process, and cellular response to stress were elevated, and innate immune response, adaptive immune response, and defense response were suppressed (Supplementary Figure S5B). The biological function analysis results revealed that the microtubule-based process, apoptotic process, and response to heat were remarkably activated, while the immune response regulating cell surface receptor signaling pathway and cell recognition were inhibited in the gamma delta T cells of the HCC group (Supplementary Figure S5C). The metabolic process was enhanced in dendritic cells of the HCC group, but the response to growth factor, regulation of lymphocyte activation, and cell activation were repressed (Supplementary Figure S5D). The response to endothelial reticulum stress and monocarboxylic acid metabolic process were greatly stimulated in the B cells of the tumor group, whereas cytoplasmic translation and biosynthetic process were significantly suppressed (Supplementary Figure S5E). Compared to non-tumor hepatocytes, malignant hepatocytes have markedly active cell proliferation and differentiation pathways, including ribonucleoprotein complex subunit organization, electron transport chain, and aerobic respiration. In addition, malignant hepatocytes exhibited suppressed lymphocyte activation, immunological response, and defensive response (Supplementary Figure S5F).

Since there were apparent differences in the levels of oxidative stress between HCC cells and adjacent non-tumor tissue cells, we further compared the situations of intercellular communication in the two tissues. According to the preliminary clustered cell-cell communication network analysis results, the interaction strength and net number of interactions between malignant hepatocytes cells and immune cells communication were superior to those between non-tumor hepatocytes cells and immune cells (Figures 4A, B; Supplementary Figure S6). Next, we performed a Laplacian clustering analysis of signaling pathways involved in intercellular communication based on functional and structural similarities (Figures 4C, D). Herein we focus on the functional similarity (two signaling pathways or two ligand-receptor pairs with similar roles) of intercellular communication. Inflammation and immune cell chemotaxis signaling pathways (such as IFN−II, TNF, IL16, IL10, and CXCL) were included in groups 1 and 2. In group 3, MIF, MK, FN1, and Galectin pathways for communication and migration between immune cells and malignant hepatocytes cells. Group 4 mainly involved the intra and exocrine pathways (ANGPTL, GDF) in HCC. Our analysis of signaling pathways in cell populations identified three synergistic patterns of efferent (Figure 4E) and afferent signaling (Figure 4F). The inflammation pathways were the synergistic efferent signaling pathways of T memory cells of adjacent non-tumor tissue and gamma delta T cells, T memory cells, mucosal-associated invariant T cells, and B cells from both tissues. In efferent immune inflammatory response pathways, dendritic cells of adjacent non-tumor tissue and T memory cells of HCC tissue communicate synergistically. Inflammatory pathways were coordinately afferent in T memory cells of adjacent non-tumor tissue, B cells, gamma delta T cells, and mucosal−associated invariant T cells. Dendritic cells of adjacent non-tumor tissue and T memory cells of HCC tissue had synergistic immune chemotactic afferent pathways. Tumor growth and differentiation pathways were coordinately afferent in malignant hepatocytes cells.

Further, by comparing outgoing and incoming interaction strength in non-tumor and HCC tissues, we identified significant changes in the signals sent or received by T memory cells, B cells, and malignant hepatocytes (Figure 5A). The memory T cells and B cells in the control group mainly served as signal receivers in cell communication, while the memory T cells and B cells in HCC mainly served as signal transmitters. In HCC, malignant hepatocytes send a higher intensity of intercellular communication signals than non-neoplastic hepatocytes in the control group. The dot plot visually represents the dominant senders and receivers of intercellular communication. The X and Y-axes correspond to the total outgoing and incoming communication probabilities, respectively, for each cell group. The size of the dots indicates the number of inferred links associated with each cell block, considering both outgoing and incoming interactions. Larger dots indicate a higher number of links. The colors of the dots differentiate between different cell groups, allowing for identification and comparison. Interestingly, we discovered that only the MIF pathway was significantly enriched in non-neoplastic and neoplastic contexts (Figure 5B). Each cell cluster exhibited significant MIF pathway communication in incoming (Figure 5C) and outgoing (Figure 5D) modes under non-tumor and HCC conditions. In light of the above findings, we further investigated the communication between HCC cells and hepatocytes or immune cells through receptor-ligand interaction studies (Figure 5E). The results indicated that malignant hepatocytes and immune cells, including B cells, dendritic cells, gamma delta T cells, mucosal−associated invariant T cells, and T memory cells interact via CD74-CXCR4. Second, HCC cells interact with dendritic cells, gamma delta T cells, mucosal−associated invariant T cells, and T memory cells through CD74⁻CD44. However, no relevant communication between HCC cells and adjacent non-tumor hepatocytes has been found in MIF signaling. Strikingly, MIF communication was significantly more potent in the HCC environment than in the non-tumor environment (Figures 6A, B). The width of the edges in the graph is determined by the number of interactions, which reflects the number of ligand-receptor pairs involved in the communication between two interacting cell clusters. A wider edge indicates a higher number of interactions between the cells. In the HCC state of the MIF signaling pathway (Figure 6C), malignant hepatocytes serve as signal senders. Gamma delta T cells, mucosal−associated invariant T cells, and T memory cells were recipients. Dendritic cells and mucosal−associated invariant T cells act as mediators. B cells and mucosal−associated invariant T cells were the influencers. In contrast, non-neoplastic hepatocytes in the non-tumor state communicate significantly much weaker with immune cells (Figures 6D–F).

Next, differentially expressed oxidative stress response-related genes (DEGs) of each cell type between non-tumor samples and HCC samples were marked in the volcano and bar plot (Supplementary Figures S7A, B; Supplementary Table S3). These DEGs were further subjected to MCODE analysis to screen hub genes of oxidative stress response in the pathogenesis of HCC. The MCODE analysis determined that HMOX1, POR, SOD1, PRDX6, P4HB, GPX4, PRDX3, PRDX5, JUN, PRDX2, TXN, FOS, and PRDX1 were the oxidative stress-related hub genes among DEGs (Figure 7A). Metascape analysis further revealed that the genes HMOX1, POR, JUN, and FOS play a role in regulating the p53 pathway as hub genes. Other hub genes regulate oxidative stress response through the ROS pathway. All the DEGs were associated with the ROS pathway, TNFA signaling via NFKB, apoptosis, hypoxia, and mTORC1 signaling pathway (Figure 7B). This analysis provided evidence of the involvement of these genes in critical biological processes of HCC. Using cBioPortal (Cerami et al., 2012; Gao et al., 2013) tools in the TCGA cohort, we further investigated the mutational landscape of these 13 hub genes in HCC. Chromosome mapping revealed that chromosomes 1 and 19 contain the most significant number of dysregulated oxidative stress-related hub genes in HCC. In contrast, no X and Y chromosomes contain dysregulated oxidative stress-related hub genes (Figure 7C). We analyzed the somatic cell mutation spectrum of 366 patients with HCC to determine the mutation rate of these hub genes. Among the hub genes in HCC patients, PRDX6, P4HB, and POR showed the most significant amplified genetic alteration rates of 9%, 5%, and 1.9%, respectively. PRDX1 displayed a high structural variation rate of 5%. It was observed that FOS had a deep deletion rate of 1.3%. The remaining hub genes exhibited minimal mutation rates below 0.5% in HCC patients (Figure 7D). These findings shed light on the mutational landscape of the hub genes associated with oxidative stress response in HCC, highlighting specific genes with significant genetic alterations.

Since the oxidative stress response level was significantly different between non-tumor tissues and HCC tissue cells, based on the bulk transcript data, we further analyzed the diagnostic value of oxidative stress response-related hub genes by four machine learning models (RF, SVM, XGB, GLM). The RF and SVM models performed the lowest residual values (Figures 8A, B). Meanwhile, the top 10 ranked gene variables in the four machine learning models were filtered out by root mean square error (Figure 8C). By visual ROC analysis, the RF and SVM models also had the highest AUC (0.981) in 5-fold cross-validation (Figure 8D). Then, in conjunction with these outcomes, six oxidative stress response-related gene markers (GPX4, HMOX1, PRDX1, FOS, PRDX5, and TXN) with the highest diagnostic value were chosen from the RF and SVM model (Supplementary Table S4). Further, we constructed a diagnostic nomogram based on these six gene markers to test their clinical decision-making efficiency in predicting the early occurrence of HCC (Figure 9A). The calibration curve analysis showed that the accuracy of the nomogram model in predicting the occurrence risk of HCC was very close to the actual sample probability (Figure 9B). The decision curve analysis reveals that this nomogram is highly accurate with a considerable net benefit (Figure 9C). Additionally, the effectiveness of our nomogram model and gene markers has been verified in seven external datasets by ROC analysis (GSE76427, AUC: 0.935; GSE54236, AUC: 0.828; GSE36376, AUC: 0.963; GSE69715, AUC: 1.000; GSE121248, AUC: 0.974; GSE107170, AUC: 0.883; GSE45267, AUC: 0.974; Training cohort, AUC: 0.993) (Figure 9D, Supplementary Figure S8). The differences in single-cell levels of these six genes between non-neoplastic and malignant hepatocytes were consistent with the transcriptome analysis (Figures 9E, F). Additionally, we investigated the expression of these genes in immunohistochemical pathological sections from The Human Protein Atlas (www.proteinatlas.org/) and found that they were highly expressed, further supporting their potential as biomarkers (Supplementary Figure S9). These results suggested that these six gene diagnostic markers can effectively identify early-stage HCC individuals.

We further performed survival analysis for the six oxidative stress-related diagnostic gene markers from above steps to observe their associations with HCC prognostic progression. Among them, the expression levels of HMOX1, PRDX1, and TXN were significantly related to overall survival in univariate Cox regression analysis of the TCGA-LIHC cohort (p = 0.004, p < 0.001, p = 0.029, respectively). Meanwhile, in the HCCDB18 cohort, PRDX1 and PRDX5 were markedly linked with HCC individuals’ overall survival rate by univariate cox regression analysis (p = 0.018, p = 0.018, respectively). Based on multivariate cox regression in two cohorts (Supplementary Table S5), PRDX1 was identified to be the independent prognostic predictor for HCC (TCGA-LIHC: p < 0.001, Hazard ratio = 1.715, 95%Confidence interval = 1.326–2.217; HCCDB18: p = 0.018, Hazard ratio = 1.787, 95%Confidence interval = 1.104–2.892). Thus, PRDX1 was selected for further research.

Depending on median cut-off values, HCC patients in TCGA-LIHC and HCCDB18 were divided into high and low PRDX1 expression groups. The overall survival of HCC patients in the high PRDX1 group was considerably worse than that of patients in the low PRDX1 group, as shown by Kaplan-Meier analysis (Figures 10A, B). We also found that HCC patients with high expression of PRDX1 have much worse progress-free intervals and disease-specific survival status (Supplementary Figure S10). This tendency of HCC patients was also visible in the survival plot (Figures 10C, D). In univariate cox analysis with clinicopathological characteristics, the PRDX1 was significantly related to the overall survival of HCC patients (TCGA-LIHC: Hazard ratio = 2.449, 95%Confidence interval = 1.452–4.129, p < 0.001; HCCDB18: Hazard ratio = 2.988, 95%Confidence interval = 1.462–6.098, p = 0.003) (Figures 10E, G). Multivariate cox regression determined that the PRDX1 was an independent predictive factor for patients’ prognoses among clinicopathological characteristics (TCGA-LIHC: Hazard ratio = 2.395, 95%Confidence interval = 1.394–4.115, p = 0.002; HCCDB18: Hazard ratio = 2.413, 95%Confidence interval = 1.193–4.881, p = 0.014) (Figures 10F, H). At the same time, the prognostic prediction power (AUC) for the PRDX1 in 1 year, 2 years, and 3 years in the ROC analysis of two cohorts was relative satisfaction (Figures 10I, J). Furthermore, a nomogram was created to evaluate the prognosis status of HCC patients by incorporating clinicopathological characteristics, including grade, stage, age, gender, and the expression profile of PRDX1 (Figure 10L). The ideal fitting degree of the calibration curves to the observed values validated the accuracy of nomogram in predicting HCC patients’ outcomes (Figures 10M, N). Moreover, we found the expression level of PRDX1 was consistently high in most HCC cell lines by Cancer Cell Line Encyclopedia (https://sites.broadinstitute.org/ccle) (Figure 10K). As a result, as an independent prognostic factor, PRDX1 has high clinical value in assessing the progression of HCC patients.

Gene set enrichment analyses were conducted in HCC, and PRDX1 was found to be mainly involved in digestion, ribonucleoprotein complex biogenesis, ribosome biogenesis, gene silencing by RNA, and smooth muscle cell differentiation (Supplementary Figures S11A, B). Furthermore, we explored the PRDX1-related specific biological process activation status in HCC patients. In the bubble plot, the activated PRDX1-related pathways in HCC patients mainly have ribosome biogenesis, ribonucleoprotein complex biogenesis, ncRNA processing, cellular glucuronidation, uronic acid metabolic process, snRNA processing, and regulation of programmed necrotic cell death (Supplementary Figure S11C). In contrast, PRDX1-related pathways, including vascular-associated smooth muscle cell differentiation, smooth muscle cell differentiation, IL-6 mediated signaling pathway, endosome to plasma membrane protein transport, gene silencing by RNA, and negative regulation of transporter activity, were suppressed in HCC patients.

By integrating multiple immune cell infiltration analysis methods (Timer, Cibersort, Cibersort−abs, Quantise, Mcpcounter, and Xcell), the abundances of immune cells, incorporated CD4⁺ T memory resting cell, CD4⁺ T central memory cell, Tregs, neutrophils, myeloid dendritic cell, and NK cells, were markedly enriched in the high-PRDX1 group than the low-PRDX1 group. Non-immune cells, such as endothelial cells, and cancer-associated fibroblast, were also significantly enriched in the high-PRDX1 group (Figure 11A; Supplementary Figure S12). From ESTIMATE algorithm analysis results, the immunological score and estimated score (tumor purity) of HCC patients with PRDX1 high expression were relatively high than those with low expression (Figure 11B). Afterward, the group with high PRDX1 expression held significant T cell exclusion and microsatellite instability (MSI) (Figures 11C, D). Meanwhile, the level of PRDX1 expression linked positively with HCC tumor mutational burden (TMB) (Figure 11E). Moreover, using Human Protein Atlas database (http://www.proteinatlas.org/), we discovered immunohistochemical staining intensity of PRDX1 of the HCC sample was considerably higher than the normal liver sample (Figure 11F). Immune functions, including APC co-inhibition, HLA, MHC class I, and para-inflammation, were active in the HCC patients with high PRDX1 expression. At the same time, type II IFN response was more active in the group with low PRDX1 expression (Figure 12A).

Following, we explored the relationships between immune checkpoints and HCC patients in different PRDX1 expression groups. HCC patients in the high-PRDX1 group had elevated immune checkpoints, incorporating PDCD1LG2, CTLA4, TIGIT, LAIR1, CD86, CD47, TNFRSF9, TNFRSF18 and TNFRSF4 than the patients in the low-PRDX1 group (Figure 12B). Increased expression levels of immune checkpoints were positively correlated with the expression of PRDX1 (Figure 12C). It implied that HCC patients with high PRDX1 expression might benefit more from immune checkpoint inhibitor therapies. To explore suitable drugs for HCC patients with high expression of PRDX1, we used the oncoPredict R package in testing the sensitivity of drugs between two groups with differing PRDX1 expression levels, based on Cancer Therapeutics Response Portal (http://portals.broadinstitute.org/ctrp.v2.1/) and Genomics of Drug Sensitivity in Cancer (https://www.cancerrxgene.org/) databases (Figure 12D). From the results, HCC patients in the high PRDX1 expression group exhibited greater sensitivity to twelve drugs, including Sorafenib (multi-kinase inhibitor), Linsitinib (IGF-1R inhibitor), JAK inhibitor (AZ960), ERK inhibitor (ERK 2440), Mitochondrial inhibitor (Dihydrorotenone), IKK inhibitor (BMS-345541), IRAK4 inhibitor (IRAK4_4710), SYK inhibitor (Entospletinib), TGF-β Receptor I/II inhibitor (LY2109761), MCL-1 inhibitor (AZD5991), KRAS (G12C) inhibitor, and I−BET inhibitor (I-BET-762), than patients with low PRDX1 expression group. These pharmaceuticals could be beneficial in treating HCC patients with high PRDX1 expression.