Identification of a novel platelet-related prognostic model for hepatocellular carcinoma

Abstract

Hepatocellular carcinoma (HCC) is a highly lethal malignancy with high invasiveness and metastasis. Despite progress in its treatment, the high mortality rate persists due to poor prognosis. In this study, we aimed to develop a prognostic model for HCC using platelet-related genes. We downloaded and preprocessed data from the Cancer Genome Atlas (TCGA) and Gene Expression Omnibus (GEO) databases. Through differential gene expression analysis, univariate Cox regression, Least absolute shrinkage and selection operator (LASSO) machine learning, and multivariate Cox regression, we constructed a prognostic model consisting of six genes (KIF18A, HRG, TUBA4B, MAFF, SELP, and IGF1). The survival analysis and ROC curve evaluation conducted on both the training and validation sets showed excellent predictive performance of the model. In addition, the model was also associated with the malignancy and risk of metastasis of tumors. Additionally, analysis of half maximal inhibitory concentration (IC50) values between high- and low-risk groups for various drugs showed significant differences, suggesting potential differences in predicted drug sensitivity between risk groups. We identified potential stem-like cell subpopulations in HCC cells, which are mostly in the late stage of malignant cell differentiation. In conclusion, we successfully constructed a novel platelet-related prognostic model for HCC.

Supplementary Information

The online version contains supplementary material available at 10.1007/s12672-026-04976-4.

Introduction

Hepatocellular carcinoma (HCC) is one of the deadliest malignancies in the world, with high invasiveness and metastasis [1]. Significant progress has been made in the monitoring, diagnosis, and management of HCC. For example, some patients with HCC can be treated clinically through immunotherapy with immune checkpoint inhibitors (ICIs) such as anti-CTLA-4 and anti-PD-1. In addition, targeted therapies such as sorafenib and rapatinib have also shown progress in treating advanced HCC. However, the mortality rate of HCC remains high [2–5] mainly due to the poor prognosis. Thus, there is an urgent need to construct an efficient prognostic model to guide the selection of appropriate treatment strategies and medications for HCC.

Platelets, as a cellular component in the blood, play an important role in many physiological and disease processes [6]. A growing number of studies have shown that platelets can affect the invasion, metastasis, and shedding of tumor cells [7–9]. In addition, platelets can protect the survival of tumor cells in circulation and promote the migration of tumors in the body [10]. Nataša Pavlović et al.. reported that activated platelets can promote HCC cell proliferation and facilitate the unfavorable tumor-stroma interaction, which promotes tumor progression [11]. A study of Lu et al.. suggested that platelets could enhance the metastasis of primary HCC through the induction of autophagy in cancer cells by TGF-β1 [12]. Gao et al.. reported that platelets accelerated HCC metastasis through the TLR4/ADAM10/CX3CL1 axis [13]. Recently, fewer studies have reported on a platelet-related prognostic model for HCC [14–16]. However, these signatures are mainly associated with the immune microenvironment or responsiveness to immunotherapy with ICIs. The prediction capability of these signatures in predicting the malignancy and metastatic risk of HCC is unclear.

In this study, using single-cell and bulk RNA sequencing analysis, we identified six platelet markers for HCC. Based on these markers, a novel prognostic model for HCC was constructed. After evaluation and verification, the prognostic model proved to be effective in the diagnosis and prognosis assessment of HCC. Meanwhile, the prognostic model could predict the malignancy and the metastatic risk of HCC. In addition, this study identified that potential stem-like cells are mostly cells in the late stage of the malignant cell differentiation process. The flow chart is presented in Fig. 1.

Fig. 1.

Schematic diagram of this study

Materials and methods

Data download and preprocessing

The transcriptome data related to HCC were obtained from the Cancer Genome Atlas (TCGA) database and were retrieved through the UCSC Xena platform (https://xenabrowser.net/), consisting of RNA-seq count data and FPKM expression data. The corresponding clinical characteristics and survival follow-up information were also downloaded simultaneously. During the data preprocessing stage, to avoid including the same patient multiple times in the analysis, the first 12 digits of the TCGA barcode were used as the unique patient identifier. Sample screening was conducted at the patient level, and only primary solid tumor samples (Primary Solid Tumor, sample type code 01 A) and corresponding normal tissue samples (Solid Tissue Normal, sample type code 11 A) were retained. When a patient had multiple tumor sequencing samples, the duplicated function in R language was used to retain only one sample to ensure that each patient was included only once in the analysis. In the end, the training set included a total of 419 samples, including 369 HCC tumor samples and 50 normal samples (paracancerous samples). This dataset was used as the training set for differential expression analysis and the construction of the prognostic model (Table 1).

Table 1.

Datasets of mRNA expression profiles in hepatocellular carcinoma (HCC)

Dataset ID	Platform	HCC	Normal
TCGA	–	369	50
GSE14520	GPL3921	221	–
GSE76427	GPL10558	115	52

The HCC datasets GSE14520 and GSE76427 from the Gene Expression Omnibus (GEO) database (https://www.ncbi.nlm.nih.gov/geo/) were obtained as the validation sets. The downloaded data were the log2-transformed signal intensity values standardized by the RMA method. During the data preprocessing, the probes were first mapped to the corresponding genes, and the empty probes without gene annotations were excluded; if multiple probes corresponded to the same gene, the average expression value of their expression levels was taken as the final expression level of that gene. The GSE14520 and GSE76427 datasets were used to conduct external validation of the stability and predictive performance of the prognostic model (Table 1).

According to the previous research methods [17], based on the Gene Set Enrichment Analysis (GSEA) database (https://www.gsea-msigdb.org/gsea/index.jsp), and the “platelet” as the search keyword, relevant gene sets were systematically screened, and finally, 480 platelet-related genes were sorted out. The single-cell dataset is from the study by Sun et al.., including 4 samples of distant metastasis (early recurrent HCC) and 12 samples of non-distant metastasis (primary HCC) [18].

Screening differentially expressed genes (DEGs)

The difference analysis of TCGA data was performed using the package DEseq2 (Version 1.42.1) in R (Version 4.2.0). Adj.pvalue < 0.05 corrected by Benjamini-Hochberg (FDR) and |log2FC|>1 were selected as thresholds to obtain DEGs between HCC and control. Meanwhile, a heatmap depicting the expression of DEGs was drawn using the R software package pheatmap (version 1.0.12).

Candidate gene collection and expression data

The DEGs and platelet-related genes were investigated jointly, and the genes that appeared twice were screened as the objects of investigation; the candidate gene set was obtained. The expression profiling data of candidate gene clusters in TCGA were selected to construct and analyze the prognostic model in the next step.

Univariate cox regression analysis screening for candidate genes associated with HCC prognosis

Univariate Cox proportional risk regression analysis was performed based on the candidate gene expression profile data and the survival information of the training set samples. The number of samples with both expression profile and survival information was 363 (R package survival, Version 3.5-8). The cutoff of the univariate Cox was set to p < 0.05. The genes were screened by PH hypothesis test. Then, the resulting genes are used for a subsequent least absolute shrinkage and selection operator (LASSO) machine learning algorithm to further screen candidate genes.

LASSO machine learning algorithm further screening for candidate genes

Based on the gene expression profile obtained by univariate Cox analysis and the survival data of the training set samples, LASSO machine learning algorithm (R-package glmnet, Version 4.1-8) was further used to screen candidate genes. Specifically, the glmnet software package was used to conduct LASSO-Cox proportional hazards regression analysis on the prognostic-related genes, and the optimal penalty parameter λ was determined through 10-fold cross-validation. During the model construction process, the cv.glmnet function (cv.glmnet(x, y, family = “cox”, maxit = 1000)) was employed to perform cross-validation on the Cox proportional hazards model, with model performance evaluated using partial likelihood deviance. During the cross-validation process, the samples were randomly divided into 10 subsets, and 9 of these subsets were used for model training while the remaining 1 subset was used for validation. This process was repeated in a cycle until all folds were completed. Finally, the λ value that resulted in the minimum average cross-validation bias for the model was selected as the optimal penalty parameter. At this parameter value, the model fitting performance was the best. Subsequently, genes with non-zero regression coefficients were further screened based on the optimal penalty parameter for the subsequent construction of the prognostic model.

Constructing the prognostic model of platelet-related HCC using multivariate cox regression analysis

To further construct a prognostic model, according to the results of LASSO analysis, a multivariate Cox regression analysis on the expression profiles, expression levels and survival information of candidate genes was constructed and the Step function was used to screen model genes.

Evaluation and validation of a prognostic model

To assess the prognostic value of the model constructed, the risk value of each sample in the training set through the gene expression profile and the coefficient of the prognostic model were calculated. Then, the samples were divided into two groups (high- and low-risk groups) according to the median risk value. Survival analysis and Receiver Operating Characteristic (ROC) curve (R package timeROC, Version 0.4) prognosis prediction were performed for the high- and low-risk groups. If there was a significant difference in survival between high- and low-risk groups in the results (p < 0.05) and the Area Under Curve (AUC) value of the ROC curve was greater than 0.6, it indicated that the prognostic model constructed in this study could well predict the prognosis of the samples. To further evaluate the discriminative ability and stability of the model, this study calculated Harrell’s concordance index (C-index) and its 95% confidence interval to measure the overall discriminative ability of the model for patient survival ranking. In addition, to evaluate the internal stability and potential overfitting risk of the model, bootstrap resampling was performed on the training set for internal validation (repeated 1000 times). The degree of optimistic bias of the model was evaluated by comparing the performance of the original model with the performance after resampling correction. To test the independent prognostic value of the risk score, the risk score and clinically recognized important prognostic variables (including tumor stage, AFP level, tumor size, and age) were included in a multivariate Cox proportional hazards regression model for analysis.

Furthermore, to evaluate the predictive performance of the platelet-related HCC prognostic model constructed in this study, we selected previously published platelet-related HCC prognostic models as controls for comparative analysis. The control models were derived from publicly available literature [19]. Through decision curve analysis (DCA), the potential net benefits and application values of different models in clinical decision-making were systematically compared.

To explore the potential biological functions of the core genes and their associations with platelet-related pathways, the modEnrichr online analysis platform (https://maayanlab.cloud/modEnrichr/) was used to conduct functional enrichment analysis of the model genes. The enrichment analysis results were output using the default parameters of modEnrichr.

Analysis of difference in IC50 between high- and low-risk groups

The R package oncePredict (Version 1.2) was used to analyze the half maximal inhibitory concentration (IC50) values of each sample in the predicted training set based on two databases: Genomics of Drug Sensitivity in Cancer (GDSC) and Cancer Therapeutics Response Portal (CTRP). The rank sum test was used to compare the differences in IC50 values between high- and low-risk groups.

Single-cell analysis

Single-cell analysis was performed employing the Seurat package (Version 5.0.3). First, the original single-cell expression matrix was imported, and the sequencing depth (nCount_RNA), the number of detected genes (nFeature_RNA), and the proportion of mitochondrial gene expression for each sample were visually evaluated using violin plots. Subsequently, the data were preprocessed and quality controlled following the standard procedure: when constructing the Seurat object, only genes expressed in at least 3 cells were retained (min.cells = 3), and cells with a detected gene number of no less than 200 were selected (min.features = 200). Further, cells were evaluated based on the proportion of mitochondrial gene expression. The results showed that the proportion of mitochondrial genes in all cells in this dataset was below 5%, which met the conventional quality control standards. Therefore, no additional cells were excluded. Subsequently, the data were standardized using the ScaleData function, and the cells were annotated using the annotation information of the original data set. After the data processing was completed, subsequent analysis was performed.

Secondary clustering on malignant tumor cells was performed. First, the stemness genes in a previous study [20] were used to calculate the stemness gene scores of malignant cells. The Monocle software package (version 2.30.0) was used to conduct pseudo-time series analysis on malignant cells to reconstruct their differentiation trajectories and evaluate the potential association between the malignancy of the tumor and the platelet risk score. Given that the trajectory inference of Monocle2 relies on ordered genes to depict the continuous transitions of cell states, in this study, 2000 highly variable genes were selected as ordered genes for pseudo-time analysis. Highly variable genes are more likely to capture the key transcriptional dynamic changes within the cell population, which helps to improve the stability and reliability of trajectory reconstruction.

Results

Systematic screening of platelet-related prognostic genes and construction of risk model

By conducting differential expression analysis on the training set, we obtained 3295 DEGs in HCC, among which 2256 were up-regulated genes and 1039 were down-regulated genes (|log2FC| > 1 and Padj < 0.05) (Fig. 2A, B). Further, we conducted intersection analysis of these DEGs with the platelet-related gene set to obtain 102 intersection genes (Fig. 2C). Subsequently, these intersection genes were used as candidate genes and a series of regression analyses was employed to identify genes significantly related to HCC prognosis. In the training set, univariate Cox regression analysis was performed on the above candidate genes, and we selected 47 characteristic genes significantly related to HCC overall survival (OS) (p < 0.05) (Supplementary Table 1). After the proportional hazards assumption test, 28 genes were retained (Fig. 2D). Then, LASSO Cox regression was used for feature compression, and 11 key genes were selected under the optimal penalty parameter (lambda.min = 0.034) (Fig. 2E, F). Finally, these 11 genes were included in the multivariate Cox regression analysis, and a platelet-related HCC prognostic model was constructed using stepwise regression. The model consisted of six genes: KIF18A, HRG, TUBA4B, MAFF, SELP, and IGF1. The regression coefficients and risk contributions are shown in Fig. 2G and H. KIF18A, TUBA4B, and MAFF were identified as risk genes (HR > 1), while HRG, SELP, and IGF1 were identified as protective genes (HR < 1) (Fig. 2G, Supplementary Table 2).

Fig. 2.

Construction of a Prognostic Model. A: Volcano map of differentially expressed genes (DEGs). (Note: The vertical axis represents -log10 (padj), and the horizontal axis represents the difference fold log2 (Fold change). Each point in the figure represents a gene; green represents up-regulated DEGs; orange represents down-regulated DEGs, and gray represents non-significantly DEGs). B: Heat map of DEGs. (Note: Each column represents a sample, and each row represents the expression level of a gene in different samples (displaying the top 40 genes in |FC|); the color of the heat map represents the expression level of the gene in the sample, and the higher the expression level, the darker the color (green for high expression, orange for low expression); green represents the control group, and red represents the disease group; violin plots and boxplots represent the expression of genes in different samples). C: Venn diagram of candidate genes. D: Forest plot of univariate Cox analysis. (The left side represents the gene and the corresponding p value and HR value; the red square on the right side indicates that the HR value is greater than 1, the green square indicates that the HR value is less than 1, and the line segments on both sides of the square are the 95% confidence interval (CI) of the HR value). E and F: LASSO regression analysis. (Note: The vertical axis in (E) represents the coefficient of the gene, and the vertical axis in (F) represents the partial likelihood deviation). G: Forest plot of multivariate Cox analysis. (Note: The left side represents the gene and the corresponding p value and HR value; the red square on the right side indicates that the HR value is greater than 1, the green square indicates that the HR value is less than 1, and the line segments on both sides of the square are the 95% CI of the HR value). H: Coefficients of prognostic genes

The model formula is as follows:

Evaluation of prognostic model

This prognostic model was evaluated using TCGA data of HCC-related transcriptomes as the training set. Firstly, we calculated the risk score for each sample based on the expression levels and coefficients of the six genes in the prognostic model. Subsequently, using the median risk score (0.3977) as the threshold, the patient samples were divided into high-risk and low-risk groups (Fig. 3A). The survival analysis of the training set showed a significant difference in OS rates between the high-risk group and the low-risk group (p < 0.05). Compared to the low-risk group, the mortality rate of patients in the high-risk group was higher, and the survival rate was significantly lower, suggesting a poorer prognosis (Fig. 3B, C). Further evaluation of the model’s predictive performance through the ROC curve showed that the AUC values of the model at 1 year, 2 years, and 3 years were 0.65, 0.64, and 0.64, respectively, all higher than 0.6, indicating that this prognostic model has stable and acceptable predictive efficacy (Fig. 3D). Additionally, the expression heatmap of risk-related genes showed that the risk genes KIF18A, TUBA4B, and MAFF were significantly overexpressed in the high-risk group, while the protective genes HRG, SELP, and IGF1 had higher expression levels in the low-risk group (Fig. 3E), which was consistent with the results of model construction.

Fig. 3.

Evaluation of prognostic model. A Statistical scatter plot of RiskScore for each sample in the training set. B Statistical scatter plot of survival and death status for each sample in the training set. C Kaplan-Meier curve for survival analysis. (Green represents the high-risk group and orange represents the low-risk group). D ROC curve based on the patients’ risk scores. E Heat map on the expression of prognostic genes between high- and low-risk groups in the training set

Validation of prognostic models

To further evaluate the robustness of the prognostic model, we used the HCC cohort from the GEO database as an external validation set. In the GSE14520 dataset, the risk score of each sample was calculated based on the expression levels of the six genes in the model and the regression coefficients. The patients were divided into high-risk and low-risk groups using the median value of the risk score (2.6882) (Fig. 4A). The Kaplan–Meier survival analysis showed that patients in the high-risk group had a significantly shorter OS period compared to the low-risk group, indicating a worse prognosis (Fig. 4B-C). The ROC curve analysis indicated that the model had good predictive performance in the GSE14520 dataset, with AUC values of 0.76, 0.70, and 0.69 for 1-year, 2-year, and 3-year periods, respectively (Fig. 4D). Additionally, the gene expression heatmap showed that the risk genes KIF18A and MAFF were continuously highly expressed in the high-risk group, while the protective genes HRG, SELP, and IGF1 had higher expression levels in the low-risk group; TUBA4B showed no significant difference between the two groups (Fig. 4E). The model performance was further validated in the independent validation cohort GSE76427. The Kaplan-Meier survival analysis also showed that the prognosis of patients in the high-risk group was significantly worse than that in the low-risk group (p < 0.05) (Figure S1A). The ROC analysis results showed that the model had AUC values of 0.54, 0.63, and 0.82 for 1-year, 3-year, and 5-year periods in this cohort (Figure S1B). Among them, the one-year AUC was relatively low, indicating that the model has limited discriminative ability in short-term survival prediction.

Fig. 4.

Validation of prognostic model. A Statistical scatter plot of RiskScore for each sample in the validation set GSE14520. B Statistical scatter plot of survival and death status for each sample in the validation set GSE14520. C Kaplan-Meier curve for survival analysis. (Green represents the high-risk group and orange represents the low-risk group). D ROC curve based on the patients’ risk scores. E Heat map on the expression of prognostic genes between high- and low-risk groups in the validation set GSE14520

To comprehensively evaluate the overall discriminative ability of the model, we further calculated Harrell’s C-index and its 95% confidence interval. The results showed that the C-index of the training queue was 0.747 (95% CI: 0.697–0.797), indicating that the model had good discriminative ability in the training set. The C-index of GSE76427 validation queue was 0.551 (95% CI: 0.424–0.678), indicating that the overall discriminative ability of the model in this external queue was relatively limited. Considering the differences in patient source, clinical feature distribution, and detection platform between GSE76427 and the training queue, it may lead to data distribution bias, thereby affecting the generalization performance of the model. In addition, model calibration evaluation was conducted based on 1000 bootstrap resampling in the GSE76427 queue. The results showed that the model had a slight optimistic bias in the high-risk interval for predicting 1-year and 3-year survival rates, but the overall predicted probability still maintained good consistency with the actual observed values, and the calibration error of 3-year survival prediction did not significantly increase compared to 1-year (Figure S1C). Multivariate Cox regression analysis showed that after adjusting for clinical covariates, risk score remained an independent predictor of overall survival (HR = 2.702, 95% CI: 1.678–4.352, P < 0.01) (Figure S1D). Based on the above results, the model exhibits good discriminative ability in the training queue, and there is a certain degree of performance fluctuation in the external validation queue. Overall, it still maintains a certain survival ranking ability and acceptable predictive stability.

Furthermore, we compared the platelet-related HCC prognostic model constructed in this study with previously reported related models [19]. The DCA results indicated that the model in this study was overall comparable to the previous models in terms of predictive performance, showing certain potential for clinical decision-making benefits (Figure S1E). These results suggest that the model constructed in this study maintains stable predictive performance while providing new gene combinations and analytical approaches for the prognostic assessment of platelet-related HCC.

To explore the potential biological functions of the model genes and their association with platelet-related pathways, a functional enrichment analysis was performed on the 6 model genes (KIF18A, HRG, TUBA4B, MAFF, SELP, and IGF1). The results showed that these genes were significantly enriched in platelet-related pathways (Figure S1F).

Drug sensitivity analysis and its mechanism association with risk scoring

To systematically evaluate the potential value of the platelet-related HCC risk model in predicting drug responses, we analyzed the predicted IC50 of different drugs in the training set based on the GDSC and CTRP databases. The IC50 distributions of each drug are presented in Supplementary Table S3 (for GDSC) and Supplementary Table S4 (for CTRP). Among the HCC-related drugs, various targeted therapeutic drugs that have been widely used in clinical or preclinical studies were identified, such as Sorafenib, Lenvatinib, and Regorafenib (Supplementary Table 4), which exert anti-tumor effects mainly by inhibiting signaling pathways such as VEGFR, PDGFR, and RAF/MEK/ERK. Subsequently, the Wilcoxon rank sum test was used to compare the IC50 differences between the high-risk group and the low-risk group, and the statistical results are summarized in Supplementary Table S5-S6 (for GDSC) and Supplementary Table S7-S8 (for CTRP). The results showed that there were significant IC50 differences between the two groups for 86 drugs in the GDSC database and 81 drugs in the CTRP database.

Furthermore, based on the significance level, the significant drugs were ranked, and the top six drugs were visualized to analyze the correlation between their IC50 and risk scores. In the GDSC database, the high-risk group showed higher predictive sensitivity for MK-1775, Sepantronium bromide, Daporinad, Wee1 inhibitor, ML323, and Paclitaxel (Fig. 5A), and the IC50 values of these drugs were significantly negatively correlated with the risk score (Fig. 5B). In the CTRP database, the high-risk group showed higher predictive sensitivity for Navitoclax, Zebularine, NSC632839, GSK-J4, YM-155, and Bortezomib (Fig. 5C), and their IC50 values were also significantly negatively correlated with the risk score (Fig. 5D).

Fig. 5.

Drug Sensitivity Analysis. A and B: Intergroup differences in the IC50 of six drugs in the Genomics of Drug Sensitivity in Cancer (GDSC) database (A) and their correlations with risk scores (B). C and D: Intergroup differences in the IC50 of six drugs in the Cancer Therapeutics Response Portal (CTRP) database (C) and their correlations with risk scores (D)

In addition, we focused on analyzing the sensitivity of the high-risk and low-risk groups to some drugs that have been used in HCC clinical studies. In sorafenib, the predicted IC50 between the high-risk group and the low-risk group did not reach statistical significance (p > 0.05) (Figure S2A), suggesting that the risk model has limited ability to distinguish the efficacy of sorafenib. However, we observed an extremely significant difference in the predicted IC50 of 5-Fluorouracil (5-FU) between the two groups (p < 0.001) (Figure S2B), suggesting that the risk model may reflect differences in sensitivity to 5-FU and other chemotherapy drugs. However, as these results were based on computational inference from pharmacokinetic databases, they were exploratory findings that still require further experimental and clinical validation.

Furthermore, to further clarify the potential biological mechanism between the risk score and drug sensitivity, we systematically analyzed the associations between the risk model and the pathways related to drug responses. Based on the msigdbr database, we integrated multiple gene sets related to the drug action mechanisms, including vascularization-related pathways (VEGF signaling, MAPK signaling, Angiogenesis) and immune microenvironment-related pathways (IFN-γ response, T cell exhaustion, Antigen presentation). Subsequently, the ssGSEA algorithm in the GSVA package was used to calculate the pathway activity scores for each sample, and the relationship between these scores and the drug IC50 was evaluated through Spearman correlation analysis (Figure S2C-D). The results showed that the predicted IC50 of 5-FU was significantly negatively correlated with the activities of the above multiple pathways. That is, the samples with higher pathway activity had stronger predicted sensitivity to 5-FU (lower IC50) (Figure S2D).

Overall, the risk model reveals differences in predicting chemotherapy drug sensitivity among different risk groups and provides preliminary clues for understanding their potential biological associations with drug response-related pathways. However, it should be noted that the above results are preliminary exploratory analyses and still need further validation before clinical application.

Quality control and annotation of single-cell data

After completing the quality control and standardization processing, a systematic analysis was conducted on the single-cell RNA sequencing data of 16 HCC samples. The quality control results showed that there was some heterogeneity in sequencing depth and the number of gene detections among different samples, but the overall data quality was good (Figure S3A). Further correlation analysis indicated that the proportion of mitochondrial genes was significantly negatively correlated with the total UMI count (r = -0.74), while the number of detected genes was moderately positively correlated with the total UMI count (r = 0.65) (Figures S3B–C), which was consistent with the typical quality characteristics of single-cell sequencing data. Based on the annotation information provided by the original study and combined with the expression characteristics of marker genes, all cells were labeled with their types. The UMAP dimensionality reduction results showed that a total of 10 major cell types were identified, including B cell, Endothelial, Epithelial, HSC, Malignant cell, Myeloid cell, NK cell, pDC, Plasma cell, and T cell (Fig. 6A). The marker gene expression patterns of each cell type were clear, further verifying the reliability of the annotation results (Fig. 6B). Subsequently, a comparative analysis of the cell composition between the distant metastasis group (early recurrent HCC) and the non-distant metastasis group (primary HCC) was conducted. The results showed that compared with the distant metastasis group, the proportion of malignant cells in the non-distant metastasis group was significantly lower, while the proportion of immune-related cells, especially T cells and Myeloid cell, was relatively higher (Fig. 6C). These results suggest that HCC cells with different metastasis statuses have significant differences in tumor cell composition and immune microenvironment characteristics.

Fig. 6.

Cell type annotation and inter-group composition difference analysis of single-cell RNA sequencing data. A UMAP dimensionality reduction diagram showing the clustering and annotation results of all cells, with different colors representing different cell types. B Dot plots of marker genes for each cell type. The size of the dots represents the proportion of cells expressing the gene, and the color shade indicates the average expression level of the gene. C Stacked bar charts of the proportion of each cell type in the distant metastasis group (early recurrence HCC) and the non-distant metastasis group (primary HCC)

Pseudo-temporal trajectory of malignant tumor cells and dynamic changes of genes related to risk model

To explore the relationship between the differentiation trajectory of tumor cells and disease progression, we conducted a pseudo-temporal trajectory on malignant tumor cells. The results showed that the malignant tumor cells differentiated from the branching point 2 in two directions: the left upper branch maintained a later pseudo-time state (blue, high pseudo-time value), while the right lower branch evolved towards an earlier pseudo-time state (dark cyan, low pseudo-time value) (Fig. 7A). In the early stage of differentiation and throughout one of the differentiation directions, there were distributions of metastatic cells, while in the other differentiation direction, mainly non-metastatic cells were distributed (Fig. 7B). The assessment of risk scores revealed that the distribution of risk scores was relatively consistent with the distribution trajectory of non-metastatic cells (Fig. 7C).

Fig. 7.

Pseudo-time analyses of malignant cells. A Diagram of the pseudo-time process of malignant cells. B Grouping of pseudo-time changes of cells. C. Changes of riskScore value in the pseudo-time process. D Changes of prognostic gene expression in the pseudo-time process

Further analysis of the dynamic expression patterns of key prognostic genes in the risk model during the pseudo-time process revealed that most genes exhibited distinct phased changes (Fig. 7D). Among them, HRG was highly expressed in the early stage of pseudo-time and then gradually decreased; while IGF1, KIF18A and MAFF significantly increased with the extension of pseudo-time; SELP and TUBA4B maintained a relatively low expression level throughout the development process. These results suggest that these genes may be involved in driving the transformation of tumor cells to a high-differentiation and high-metastatic potential state.

Identification of potential stem-like cell subpopulations in HCC cells

We selected malignant cells from the data and then performed secondary clustering. Four malignant cell subtypes were obtained (cluster 0–3). The distribution of four malignant cell subtypes is shown in Fig. 8A. We then used the stemness genes in a previous study [20] to calculate the stemness score for malignant cells through GSVA (1.50.0). The rank sum test was employed to compare the differences in the stemness scores of the four cell subtypes (Fig. 8B). The analysis results showed that there were significant differences in the stemness scores among the four cell clusters (clusters 0–3), with the order being cluster 1 > cluster 0 > cluster 2 > cluster 3 (Fig. 8B). Pseudo-time analysis of the malignant cell subpopulations revealed that different cell clusters exhibit distinct phased distribution characteristics along the pseudo-time (Fig. 8C). At the early differentiation stage, the cells were mainly composed of cluster 0; as the pseudo-time progressed, the proportion of cluster 1 gradually increased and remained dominant in multiple differentiation branches. In contrast, cluster 2 was mainly enriched in the later pseudo-time stages, while cluster 3 was relatively less distributed throughout the trajectory. Notably, cluster 1 was distributed throughout the pseudo-time sequence, suggesting that it may represent a malignant cell subpopulation with persistent stemness characteristics.

Fig. 8.

Potential stem-like cell subsets in malignant cells. A The UMAP distribution of four malignant cell subtypes. B The differences in the stemness scores of the four malignant cell subtypes. *** represents p < 0.001; **** represents p < 0.0001. C: Pseudo-temporal analyses of malignant cell subpopulations

Furthermore, we analyzed the correlation between the model genes and the stem cell-like cell subpopulations. The results showed that IGF1 (R = 0.43) and KIF18A (R = 0.37) were moderately positively correlated with the stemness score, HRG was weakly negatively correlated (R = -0.18), and TUBA4B, SELP, and MAFF were weakly positively correlated. All correlations were statistically significant (p < 0.05) (Figure S4), suggesting that these genes may be involved in regulating the stemness characteristics of cells.

Discussion

HCC is a malignant tumor with consistently high incidence and mortality rates worldwide. Its high heterogeneity and tendency to metastasize pose significant challenges in prognosis assessment and treatment decision-making [21]. Although targeted therapy and immunotherapy have been continuously developing in recent years, there is still a lack of stable, reliable and biologically interpretable molecular markers in clinical practice for patient risk stratification and treatment guidance [22]. Increasing evidence suggests that platelets not only participate in the processes of hemostasis and coagulation but also play a crucial role in the progression of HCC by regulating tumor cell invasion, immune escape, and the metastatic microenvironment [12, 13]. Furthermore, the pro-tumor effect of platelets is not isolated but closely collaborates with the tumor-related vascular system. A previous study has shown that the structure and functional state of liver blood vessels are closely related to the postoperative outcome and long-term survival of HCC, suggesting that the vascular-platelet-related pathways may form an important regulatory axis in the malignant progression of tumors [23]. However, most of the existing studies on platelets have focused on evaluating the efficacy of immunotherapy or correlating with the characteristics of the immune microenvironment, mainly based on static prognostic stratification analysis. There is still a lack of research that systematically integrates platelet-related molecular characteristics from the perspective of tumor malignant evolution and metastasis risk to depict the dynamic changes in the biological state of HCC [14–16]. Based on this, this study focused on platelet-related genes, systematically constructed and validated a platelet-related HCC prognostic risk model, and further combined drug sensitivity analysis and single-cell transcriptome data to reveal the potential clinical value of this model in prognostic assessment, treatment response prediction, and tumor stemness evolution at multiple levels, providing potential guidance for precision stratification and individualized therapy in HCC.

This study employed a multi-step regression analysis strategy to select six key genes, namely KIF18A, HRG, TUBA4B, MAFF, SELP, and IGF1, from the DEGs related to platelets. Based on these genes, a stable prognostic prediction model was constructed. This model demonstrated excellent predictive performance in the TCGA training set and multiple external validation cohorts from GEO, suggesting its strong robustness and generalization ability. From a biological perspective, these genes have clear or potential functional associations in tumor progression. KIF18A, as a cell cycle and mitosis regulatory factor, has been reported to be associated with the enhanced proliferation and invasion ability of various tumors [24–26]. In HCC, KIF18A is highly expressed and can drive tumor cell proliferation, invasion, and metastasis by promoting the cell cycle process and activating transfer-related signaling pathways, thereby accelerating the malignant progression of the tumor [27]. MAFF participates in oxidative stress and transcriptional regulation and may affect the adaptability of tumor cells to environmental changes [28, 29]. MAFF is found to promote the recruitment and infiltration of neutrophils by upregulating the expression of CXCL1, thereby accelerating the liver metastasis process [30]. At the same time, some research suggests that MafF may exert tumor-suppressive effects in HCC through the ITCH/miR-224-5p axis [31], indicating that it may have dual biological effects under different regulatory contexts. TUBA4B is a molecule closely related to tumor occurrence and development, and has been proven to participate in the regulation of cell proliferation, invasion, and apoptosis in various tumors, including breast cancer and colorectal cancer, and is considered a potential prognostic biomarker [32–34]. In glioma, the low expression of TUBA4B is significantly associated with poor clinical prognosis [35], but its specific function in HCC has not been systematically studied. HRG plays an important protective role in angiogenesis regulation, immune response maintenance, and tumor microenvironment homeostasis. Genetic deficiency of HRG can exacerbate immune escape and vascular abnormalities, promoting tumor growth and metastasis [36]. Correspondingly, overexpression of HRG can significantly inhibit HCC cell proliferation and promote cell apoptosis [37]. The IGF1-related signaling pathway is an important regulatory axis for HCC cell growth, migration, and drug resistance formation. IGF-1R antagonists combined with sorafenib treatment are shown to produce a synergistic effect, effectively inhibiting HCC cell migration [38]. Additionally, SELP, as a key platelet adhesion molecule, may participate in the fine regulation of platelet-tumor cell interactions [39]. Previous studies suggest that SELP mediates the interactions between platelets, endothelial cells, immune cells, and tumor cells, connecting multiple processes such as cancer-related thrombosis formation, inflammatory response, tumor metastasis establishment, and tumor immune regulation, playing an important bridging and promoting role in tumor occurrence and development, and gradually becoming a potential therapeutic target [40]. In summary, the integration of these genes in the same prognostic model suggests that the clinical outcome of HCC is not dominated by a single platelet-related factor, but rather reflects the coordinated imbalance of multiple biological processes such as platelet-mediated pro-tumor and anti-tumor signals, cell cycle regulation, and remodeling of the tumor microenvironment.

In the context of precise treatment, predicting the response of patients to different drugs is a key concern in clinical practice. This study further evaluated the relationship between platelet-related risk scores and drug sensitivity and found significant differences in the predicted IC50 values of various drugs between the high-risk and low-risk groups. Sorafenib, as the most widely used first-line multi-target tyrosine kinase inhibitor in HCC treatment, mainly exerts anti-tumor and anti-angiogenic effects by inhibiting the RAF/MEK/ERK signaling pathway and vascular generation-related receptors such as VEGFR and PDGFR [41–43]. However, in this study, the predicted sensitivity of sorafenib did not show significant differences between the different risk groups, suggesting that the platelet-related risk characteristics may not be the main determinant of the efficacy of sorafenib. In contrast, the classic antimetabolite chemotherapy drug 5-FU showed significant differences in predicted IC50 between the high-risk and low-risk groups. 5-FU mainly exerts cytotoxic effects by inhibiting thymidine kinase, interfering with DNA synthesis, and inducing cell cycle arrest. Its efficacy has been proven to be closely related to the proliferation status of tumor cells, vascular supply, and the immune microenvironment [44, 45]. Further pathway correlation analysis showed that the predicted sensitivity of 5-FU was significantly negatively correlated with the activity of angiogenesis-related pathways, MAPK signaling pathways, and multiple immune-related pathways, suggesting that the platelet-related risk model may indirectly affect the response of tumors to chemotherapy drugs (especially 5-FU) by reshaping tumor vascular status, regulating key signaling pathways, and the immune microenvironment. The above results indicate that the platelet-related risk model not only has prognostic evaluation value but also provides preliminary evidence for understanding the potential association between platelet-related molecular features and predicting chemotherapy sensitivity. However, due to the fact that drug reactions are estimated based on predicted IC50 values from cell line databases, the above results are exploratory analyses and cannot be directly used for clinical treatment decisions.

To further elucidate the biological basis of the platelet-related risk model, this study combined single-cell RNA sequencing data to conduct an in-depth analysis of the heterogeneity and evolutionary trajectory of malignant tumor cells. The time-series analysis revealed that malignant cells exhibited distinct phased changes along different differentiation directions. The distribution of risk scores was highly consistent with the evolutionary trajectory of non-metastatic cells, suggesting that the platelet-related risk model may reflect the differences in the differentiation status of tumor cells. On this basis, we further analyzed malignant cells in combination with the stemness score. We observed that a portion of malignant cells, mainly distributed in the late differentiation stage, exhibited obvious stem cell-like characteristics. Given the crucial role of cancer stem cells (CSCs) in tumor occurrence, recurrence, and metastasis [46], understanding the characteristics and sources of these potential CSCs is of great significance for optimizing the precision treatment strategy for HCC [47–49]. We further identified subpopulations of malignant cells with significant differences in stemness characteristics and found that cluster 1, with a higher stemness score, persisted throughout the time-series analysis, suggesting that it may represent a cell population with stable stemness characteristics and potential tumor-initiating ability. Correlation analysis further indicated that the expression levels of model genes were significantly correlated with the stemness score, suggesting that the platelet-related genes are not only closely related to the overall prognosis of patients but may also participate in the invasion and recurrence process of HCC by regulating the stemness state of tumor cells, providing new single-cell-level evidence for understanding the role of platelets in the progression of HCC.

This study still has some limitations. Firstly, this study mainly relies on retrospective public database data. Further validation of the model’s clinical applicability is needed in a prospective clinical cohort. Secondly, the drug sensitivity analysis is based on predicted IC50 and further verification of the model’s guiding value for actual treatment responses through in vitro and in vivo experiments is required. Additionally, the sample size of single-cell analysis is relatively limited. In the future, more HCC single-cell data can be combined to enhance the robustness of the conclusion. Moreover, the specific molecular mechanism of the model genes in the platelet-tumor interaction still needs further experimental research to be clarified.

Conclusion

In summary, this study constructed a platelet-related HCC prognostic model, and by integrating bulk transcriptome data, drug sensitivity analysis, and single-cell data, it systematically revealed the potential value of this model in prognosis assessment, treatment response prediction, and regulation of tumor stemness. This model not only provides new molecular evidence for the involvement of platelets in HCC progression, but also offers new research ideas for precise stratification and individualized treatment strategies of HCC.

Author contributions

(I) Conception and design: Hong Pan(II) Administrative support: Da Li(III) Collection and assembly of data: Hong Pan, Xi Cheng, Da Li(IV) Data analysis and interpretation: Hong Pan, Xi Cheng, Yong Dong(V) Manuscript writing: All authors(VI) Final approval of manuscript: All authors.

Funding

Zhejiang Provincial Natural Science Foundation, No.LY24H160004. Clinical Medical Research Special Fund of Zhejiang Medical Association, No.2022ZYC-D08.

Data availability

UCSC database (UCSC Xena, [https://xenabrowser.net/](https:/xenabrowser.net) ), GEO database ( [https://www.ncbi.nlm.nih.gov/geo/](https:/www.ncbi.nlm.nih.gov/geo) ), accession number: GSE14520.The datasets generated and analysed during the current study are available in the Zenodo repository, at https://dx.doi.org/10.5281/zenodo.17563572.

Declarations

Ethics approval and consent to participate

Not applicable.

Competing interests

The authors declare no competing interests.