A Single-Cell-Derived CD8+ T-Cell State Plasticity Score for Prognostic Stratification and Characterization of Immune Responsiveness in Hepatocellular Carcinoma

Abstract

Background

T cells play a central role in antitumor immunity. Emerging evidence indicates that T-cell plasticity and differentiation states are associated with clinical outcomes and responses to cancer immunotherapy. CD8⁺ T cells that retain differentiation potential are thought to sustain effective antitumor immunity. However, robust quantitative metrics for assessing tumor-infiltrating CD8⁺ T-cell state plasticity and defining its clinical significance in hepatocellular carcinoma remain lacking.

Methods

We integrated publicly available single-cell RNA-sequencing datasets from HCC cohorts to characterize CD8⁺ T-cell states using unsupervised pseudotime reconstruction and supervised ordinal transcriptional modeling. We used genes showing progressive changes along the inferred state axis and enriched in CD8⁺ T cells to develop a T-cell State Plasticity score using Cox proportional hazards and LASSO regression in the training cohort. The model was evaluated in independent bulk transcriptomic cohorts.

Results

Supervised psupertime analysis validated the predefined Tcm–Tem–Tpex–Ttex ordinal axis in held-out cells, supporting the existence of an ordered CD8⁺ T-cell state continuum in HCC. The CD8⁺ T-cell State Plasticity score robustly stratified clinical outcomes across multiple large HCC cohorts. High-risk scores, indicative of reduced CD8⁺ T-cell state plasticity, were associated with adverse clinical outcomes and an immunosuppressive tumor immune microenvironment.

Conclusion

CD8+ T-cell states are associated with tumor immune characteristics and clinical outcomes in HCC. This single-cell–derived T cell plasticity scoring system provides a quantitative framework for prognostic prediction and immune landscape evaluation, with potential implications for clinical decision-making.

Introduction

Primary liver cancer is the sixth most commonly diagnosed malignancy worldwide.¹ Many patients with HCC are diagnosed at advanced stages, and curative surgical treatment is often no longer feasible.² Major risk factors include chronic hepatitis B or hepatitis C virus infection, excessive alcohol consumption, aflatoxin exposure, and metabolic dysfunction-associated steatotic liver disease. These factors promote HCC through persistent hepatic injury, inflammation, fibrosis, and cirrhosis.³ Immune checkpoint inhibitors (ICIs) have provided a new therapeutic option for advanced HCC and improved survival in a subset of patients.^4,5 However, treatment responses remain heterogeneous, and only a subset of patients derive durable clinical benefits.⁶ These findings highlight the crucial role of the tumor immune microenvironment.^7,8 In particular, the states and functional properties of tumor-infiltrating T cells may affect responses to immunotherapy and patient outcomes.⁹

The tumor microenvironment (TME) comprises immune cells, stromal cells, extracellular matrix components, cytokines, chemokines, and metabolic signals that collectively regulate the recruitment, activation, maintenance, differentiation, and exhaustion of tumor-infiltrating T cells.¹⁰ Accordingly, T-cell abundance, states, and functional properties may link tumor immune heterogeneity with tumor control, responses to immune checkpoint blockade, and patient outcomes. Most previous studies have focused on T-cell infiltration levels or the abundance of specific subsets in relation to patient outcomes. However, increasing evidence indicates that T cells do not exist in fixed or discrete states.^11–13 Under persistent antigen stimulation and tumor microenvironmental regulation, they undergo dynamic transitions along continuous transcriptional trajectories.¹⁴ This dynamic property, often referred to as T-cell state plasticity, reflects the capacity of T cells to maintain, acquire, or transition between distinct transcriptional and functional states in response to environmental cues. Rather than representing unrestricted developmental potential, T-cell state plasticity describes the relative flexibility of immune cell states along a biologically defined continuum. Therefore, approaches based solely on cell abundance or individual functional markers may fail to capture the underlying immune potential of T cells. Single-cell transcriptomic and lineage-tracing studies have further shown that tumor-infiltrating T cells can transition from memory-like or plastic states with self-renewal capacity toward effector states and ultimately terminal exhaustion.^15–17

Within this spectrum, CD8⁺ T cells represent a particularly informative model for studying state plasticity in antitumor immunity because their differentiation continuum from memory-like/progenitor states toward terminal exhaustion has been extensively characterized. A subset of CD8⁺ T cells express memory- and stemness-related markers, such as TCF7 and IL7R.¹⁸ These cells maintain strong self-renewal capacity and can develop into effector cells or terminally exhausted cells.¹⁹ These cells are commonly referred to as progenitor-exhausted T cells (Tpex).^20,21 Many studies have shown that immune checkpoint blockade preferentially acts on progenitor-like T cells rather than terminally exhausted cells. Their expansion and differentiation contribute to durable treatment effects.^22–25 In contrast, excessive progression toward terminal exhaustion and loss of plasticity are associated with reduced responses to immunotherapy.^12,13,26

Although CD8⁺ T-cell state heterogeneity and plasticity are increasingly recognized as clinically relevant, studies in HCC remain largely descriptive at the single-cell level. Quantitative tools for measuring this feature across large patient cohorts remain lacking. Bulk RNA sequencing is widely used in clinical research but cannot resolve the internal states of individual T-cell populations.²⁷ Translating single-cell-defined CD8⁺ T-cell state programs into a quantitative bulk transcriptomic score may therefore enable evaluation of their associations with the tumor immune landscape and clinical outcomes.

In this study, we used a publicly available HCC single-cell RNA-sequencing discovery cohort to define a transcriptionally inferred CD8⁺ T-cell state continuum and derive a plasticity-differentiation axis. In this study, “CD8⁺ T-cell state plasticity” refers to the relative capacity of a CD8⁺ T-cell population to retain memory- or progenitor-like transcriptional programs and to occupy intermediate states along an inferred transcriptional continuum toward effector or exhausted states. This concept reflects a transcriptomically inferred state property rather than a direct measurement of functional self-renewal capacity, lineage transition, or developmental potential. Genes associated with the inferred axis and enriched in CD8⁺ T cells were used to construct the CD8⁺ T-cell State Plasticity (TCSP) score, which was evaluated in independent retrospective HCC cohorts. We further examined its associations with clinical outcomes, tumor immune characteristics, and pathological responses to neoadjuvant anti-PD-1 therapy. The TCSP score should therefore be interpreted as a bulk transcriptomic surrogate of the single-cell-derived CD8⁺ T-cell state program.

Materials and Methods

Data Sources

Single-cell transcriptomic profiles analyzed were derived from a previously published HCC cohort.²⁸ For prognostic model development and validation, bulk RNA sequencing datasets with matched clinical annotations were retrieved from three independent public repositories: GSE14520 (GEO),²⁹ TCGA-LIHC (TCGA),³⁰ and LIRI-JP (ICGC).³¹ Additionally, scRNA-seq data from an independent PD-1–treatment HCC cohort (GSE206325) were incorporated to evaluate CD8⁺ T-cell subset distribution and therapeutic response.²²

Preprocessing of scRNA-Seq Data

Raw 10× Genomics expression matrices were processed in R (version 4.4.2) using the Seurat package (version 4.4.0). Potential doublets were identified and removed using the scDblFinder package, and only cells classified as singlets were retained for downstream analyses. After doublet removal, cells were filtered according to predefined quality-control criteria. Specifically, cells were retained when they met all of the following thresholds: number of detected genes (nFeature_RNA) >100 and <7500; total unique molecular identifier (UMI) counts (nCount_RNA) >1000 and <50,000; mitochondrial transcript proportion <50%; hemoglobin gene proportion <1%; MALAT1 transcript proportion <5%; and ribosomal protein gene proportion <20%.

Following quality control, datasets were normalized using the NormalizeData function, followed by identification of highly variable genes using the FindVariableFeatures function. A total of 3000 highly variable genes were selected for downstream analyses. The expression matrix was subsequently scaled using ScaleData, and principal component analysis (PCA) was performed based on the selected highly variable genes. Batch effects among samples were corrected using the Harmony algorithm with the original sample identity as the batch variable. The corrected Harmony embeddings were used for downstream dimensionality reduction, clustering, and visualization. The first 50 principal components were used for principal component-based clustering and UMAP visualization.

Cell identities were annotated based on a combination of computationally identified marker genes and established canonical markers. Differentially expressed marker genes were identified using the COSG algorithm and differential expression analysis, and cluster identities were further evaluated using DotPlot and FeaturePlot visualization. Major cell populations were annotated as T cells, myeloid cells, hepatocytes, epithelial cells, fibroblasts, B cells, and neutrophils.

To characterize T-cell heterogeneity, T cells were extracted from the integrated dataset and subjected to independent subclustering and annotation using the same analytical workflow. T-cell subsets were identified based on established lineage markers and differential expression patterns, including CD8⁺ T cells, CD4⁺ T cells, T/NKT cells, and γδ T cells.

CD8⁺ T cells were subsequently isolated for high-resolution transcriptional-state analysis. CD8⁺ T-cell subclusters were annotated according to canonical marker expression and transcriptional signatures, including exhausted T cells (Ttex), tissue-resident memory T cells (Trm), cycling T cells, effector memory T cells (Tem), central memory T cells (Tcm), progenitor exhausted T cells (Tpex), activated CD8⁺ T cells, and T/NKT cells.

Pseudotime Trajectory Analysis and Supervised Transcriptional State Ordering

CD8⁺ T-cell transcriptional state trajectories were reconstructed using Monocle3 based on selected CD8⁺ T-cell subsets, including CD8⁺ Tcm, CD8⁺ Tem, CD8⁺ Tpex, CD8⁺ Ttex, CD8⁺ Trm, activated CD8⁺ T cells, cycling CD8⁺ T cells, and T/NKT cells. The Seurat object was converted into a Monocle3 cell dataset using the raw RNA count matrix. Data preprocessing was performed using preprocess_cds with 50 dimensions (num_dim = 50) and log normalization, followed by batch correction using align_cds with sample identity (orig.ident) as the alignment variable. Cell clustering was performed using the Leiden algorithm with k = 30, and principal graphs were learned using learn_graph with use_partition = TRUE, close_loop = FALSE, graph pruning enabled (prune_graph = TRUE), a minimum branch length of 10, and nn.k = 40.

CD8⁺ Tcm cells were selected as the root population for pseudotime ordering based on their characteristic memory phenotype and their early position in the inferred transcriptional state continuum. Specifically, the partition containing the largest number of CD8⁺ Tcm cells were selected, and the 200 CD8⁺ Tcm cells closest to the median UMAP position of this population were used as root cells. The inferred pseudotime trajectory was used to characterize transcriptional state transitions among different CD8⁺ T-cell subsets.

To assess whether CD8⁺ T-cell transcriptional profiles were compatible with the predefined Tcm–Tem–Tpex–Ttex ordering, we performed supervised transcriptional state ordering using psupertime. CD8⁺ T cells annotated as Tcm, Tem, Tpex, or Ttex were retained, and 6500 cells were randomly sampled from each state to reduce bias arising from unequal cell abundance. RNA expression values were log-normalized using the LogNormalize method in Seurat with a scale factor of 10,000, and the top 3000 highly variable genes were used for model fitting.

The four states were specified as an ordered factor in the sequence Tcm < Tem < Tpex < Ttex. Psupertime was fitted using a proportional-odds ordinal regression model with L1 regularization. Gene-expression values were scaled before model fitting, internal smoothing was disabled, and genes detected in fewer than 1% of cells were excluded. The regularization parameter was selected using the one-standard-error criterion in five-fold cross-validation, with cross-entropy used as the scoring metric. Ten percent of cells were withheld before model fitting and retained as a held-out test set. The random seed was set to 1234.

Each cell was assigned a continuous psupertime score from the fitted model. Agreement between psupertime scores and the predefined ordinal state labels in the held-out test set was evaluated using Spearman’s rank correlation. Score distributions were summarized using medians and interquartile ranges (IQRs). Reproducibility was assessed by comparing training and held-out cells. Model-derived ordinal cut points and agreement between annotated and predicted states were examined using density distributions and a confusion matrix, respectively.

Identification of Genes Associated with the Plasticity–Differentiation Axis

Gene dynamics along the CD8⁺ T-cell plasticity-differentiation axis were evaluated across ordered Tcm, Tem, Tpex, and Ttex states. For each gene, mean raw expression counts were calculated within each subset and converted to counts per million (CPM). The four subsets were assigned ordered state indices of 1 to 4, respectively. Spearman rank correlation was used to assess the association between mean expression and the ordered state index; no minimum absolute correlation coefficient threshold was imposed. A generalized additive model was then fitted using the mgcv package with expression as a function of state index (k = 4). Predicted expression values from the fitted curve were regressed against the ordered state index to estimate the overall trend slope. A gene was considered progressively upregulated or downregulated when the slope P-value was <0.05 and the slope direction was consistent with the sign of the Spearman correlation coefficient. CD8⁺ T-cell-enriched genes were identified using the Wilcoxon rank-sum test implemented in Seurat’s FindMarkers function, with logfc.threshold = 0.25 and min.pct = 0.10. Genes with an average log2 fold change >1.5 and a Benjamini-Hochberg-adjusted p-value <0.05 were retained. These enriched genes were intersected with genes showing ordered-state-associated downregulation, yielding 66 candidate genes. The term “CD8⁺ T-cell-enriched” indicates higher expression in the predefined CD8⁺ T-cell population than in other major cell populations and does not imply exclusive expression in CD8⁺ T cells.

Construction of the Prognostic Model

The GSE14520 cohort was used as the training cohort for model training. The 66 candidate genes were first evaluated using univariate Cox proportional hazards regression, and genes with P < 0.10 were subsequently entered into LASSO-Cox regression. The model was fitted with alpha = 1 using tenfold cross-validation with a random seed of 100. The penalty parameter was selected at lambda.min, corresponding to the minimum cross-validated partial-likelihood deviance. The TCSP score was defined as the sum of the products of the standardized expression value of each selected gene and its corresponding LASSO-Cox coefficient. Expression values were converted into Z-scores using gene-specific means and standard deviations estimated from the GSE14520 training cohort. Positive coefficients increase the TCSP score with increasing gene expression, whereas negative coefficients decrease the score. In the current model definition, a higher TCSP score represents higher predicted mortality risk and corresponds biologically to lower transcriptomically inferred CD8⁺ T-cell state plasticity. The resulting additive score represents a statistical linear predictor on the log-hazard scale:

External validation was performed in the TCGA-LIHC and ICGC-LIRI-JP cohorts. The coefficients and gene-specific standardization parameters derived from GSE14520 were applied unchanged to each external cohort, and the prognostic model was not refitted. Patients were stratified into high- and low-TCSP-score groups using the cohort-specific median TCSP score. Overall survival and recurrence-free survival, where available, were assessed using Kaplan-Meier analysis and two-sided Log rank tests. Time-dependent receiver operating characteristic (ROC) curves were evaluated at 1, 3, and 5 years. Multivariable Cox regression models included the TCSP score and available clinical covariates, including age, sex, and tumor stage. Hazard ratios (HRs) and 95% confidence intervals (CIs) were reported.

Tumor Immune Microenvironment Analysis

TCGA-LIHC TPM expression values were log2-transformed and normalized before analysis. The expression of immune-related genes and ssGSEA-estimated immune cell infiltration were compared between TCSP groups.

Assessment of the Prognostic Value of TCSP Independent of Total T-Cell Abundance

To determine whether the prognostic value of TCSP was independent of overall T-cell abundance, total T-cell infiltration in the TCGA-LIHC cohort was estimated using MCP-counter based on log2-transformed TPM expression data. The association between TCSP and estimated total T-cell infiltration was assessed using Spearman correlation analysis. Patients with complete data for overall survival, TCSP, estimated total T-cell infiltration, age, sex, and AJCC stage were included in subsequent analyses. TCSP, total T-cell infiltration, and age were standardized before model fitting. A multivariable Cox proportional hazards regression model was then constructed with TCSP, estimated total T-cell infiltration, age, sex, and AJCC stage as covariates to evaluate whether TCSP provided prognostic information independent of total T-cell abundance and established clinical factors. Hazard ratios for continuous variables were reported per one-standard-deviation increase. To further assess the incremental prognostic contribution of TCSP, nested Cox models with and without TCSP were compared using a likelihood ratio test, and model discrimination was evaluated using the concordance index (C-index). Variance inflation factors (VIFs) were calculated to assess potential multicollinearity among covariates.

Clinical Samples and RT-qPCR Validation

Paired primary HCC tumor and adjacent non-tumor liver tissues were obtained from Guangzhou Medical University (approval no. GYZL-2024-KY31). All procedures involving human specimens were conducted in accordance with the institutional guidelines for human research. Written informed consent was obtained from all patients for the use of their clinical specimens for medical research. Total RNA was extracted using TRIzol Reagent (Life Technologies, CA), and reverse transcription was performed using an Advantage RT-for-PCR Kit (Takara, JPN) according to the manufacturer’s instructions. For qPCR analysis, aliquots of double-stranded cDNA were amplified using a SYBR Green PCR Kit (Vazyme, CHN). ΔCt was calculated as Ct(target gene) - Ct(reference gene), with higher ΔCt values indicating lower relative transcript abundance. Tumor and adjacent non-tumor tissues were compared as paired samples. Primer sequences are provided in Supplementary Table 1.

Statistical Analysis

All statistical analyses were performed using R. Unless otherwise stated, all statistical tests were two-sided, and P values < 0.05 were considered statistically significant. Benjamini-Hochberg correction was applied when multiple genes or immune features were tested. Continuous variables were summarized as mean and standard deviation or median and interquartile range, as appropriate, and categorical variables were summarized as counts and percentages.

Results

Transcriptionally Inferred CD8⁺ T-Cell State Continuum in Hepatocellular Carcinoma

We analyzed publicly available single-cell RNA sequencing datasets to characterize T-cell heterogeneity in HCC.²⁸ We identified seven major cell populations, including hepatocytes, epithelial cells, fibroblasts, T cells, B cells, myeloid cells, and neutrophils (Supplementary Figure 1A and B).³² We therefore focused on tumor-infiltrating CD8⁺ T cells. CD8⁺ T cells were identified based on CD3D, CD3E, CD8A, and CD8B expression and annotated using canonical marker genes. These cells were further resolved into transcriptionally distinct subsets,^33,34 including central memory T cells (Tcm), effector memory T cells (Tem), progenitor exhausted T cells (Tpex), terminally exhausted T cells (Ttex), tissue-resident memory T cells (Trm), proliferating T cells, and activated T cells (Figure 1A and B; Supplementary Figure 1C).

Figure 1.

Single-cell analysis of the transcriptionally inferred CD8⁺ T-cell state continuum in hepatocellular carcinoma. (A) UMAP visualization of tumor-infiltrating CD8⁺ T cells. (B) Dot plot showing canonical marker genes across CD8⁺ T cell subsets. Dot size represents the percentage of cells expressing each gene, and color intensity represents the average expression level. (C) Monocle3-derived differentiation-related state ordering. Left: inferred pseudotime distribution; right: subset annotations projected onto the inferred continuum. Ttex cells are highly represented in the terminal region. (D) Distribution of psupertime scores among Tcm, Tem, Tpex, and Ttex cells. Violin plots show cell-level score distributions, boxes indicate IQRs, and black dots indicate medians. (E) Heatmap of genes with monotonic expression changes across the ordered state axis. (F) KEGG enrichment of axis-associated genes.

Monocle3 revealed a branched transcriptional state continuum of tumor-infiltrating CD8⁺ T cells. NKT-like cells were largely excluded from the main branch of the trajectory. One branch extended from Tcm through Tem toward Ttex, whereas another branch connected Tpex with the Ttex-dominated terminal region (Figure 1C). Tem and Tpex cells occupied intermediate regions along the inferred trajectory. The large terminal Ttex population displayed terminal-exhaustion-associated transcriptional features and occurred predominantly at late inferred pseudotime. Because the analysis was based on cross-sectional transcriptomic data, this ordering does not establish longitudinal state transitions or direct lineage relationships.

Before model fitting, we specified the Tcm–Tem–Tpex–Ttex order a priori on the basis of established subset biology and the transcriptional programs used for cell-state annotation. Tcm and Tem were placed at the memory-associated end of the axis, with Tcm preceding the more effector-differentiated Tem state. Under persistent antigen stimulation, cross-study analyses of tumor-infiltrating T cells have proposed transitions from effector-memory-like states toward Tpex and Tex states. Moreover, lineage and adoptive-transfer studies have shown that TCF1⁺ progenitor-exhausted cells can self-renew and generate more differentiated, TCF1⁻ terminally exhausted progeny. Thus, this ordering was used as a biologically motivated ordinal hypothesis linking conventional memory and exhaustion-associated states.

We next used supervised psupertime analysis to assess whether the transcriptional profiles of these populations were compatible with the predefined ordinal axis. In the 2600 cells withheld from model fitting, median psupertime scores increased from Tcm to Tem, Tpex, and Ttex cells (−2.46, 0.006, 0.81, and 1.78, respectively). Psupertime scores were strongly associated with the predefined ordinal state labels in the held-out cells (Spearman’s ρ = 0.726, P < 0.001; Figure 1D). Similar median scores and IQRs were observed in the training and held-out subsets (Supplementary Figure 1D).

The held-out score distributions shifted progressively along the predefined axis but remained substantially overlapping, particularly among Tem, Tpex, and Ttex cells (Supplementary Figure 1E). Concordant state assignment was higher for Tcm (70.8%) and Ttex (64.9%) than for Tem (41.3%) and Tpex (38.5%), and most discordant assignments occurred between neighboring states (Supplementary Figure 1F). These findings indicate that the transcriptional profiles reproducibly recapitulated the predefined ordinal axis while retaining considerable overlap between adjacent states.

Based on the predefined ordinal state axis and its transcriptomic compatibility in the psupertime analysis, we examined gene-expression changes across the ordered Tcm, Tem, Tpex, and Ttex populations. Two groups of genes showed monotonic expression changes along the state axis (Figure 1E). Genes showing ordered-state-associated downregulation were enriched in pathways related to early-state maintenance and signal transduction, whereas progressively upregulated genes were enriched in pathways associated with chronic antigen stimulation, immunosuppression, and metabolic stress (Figure 1F). Thus, the axis represents an ordered transcriptional organization of CD8⁺ T-cell states rather than a direct measurement of functional plasticity, self-renewal, or lineage transition.

To examine whether individual CD8⁺ T-cell subsets were associated with clinical outcomes, we performed univariate Cox regression analyses using their relative proportions across tumor samples (Supplementary Figure 1G). A higher proportion of activated CD8⁺ T cells was associated with worse overall survival (OS; HR = 1.420, 95% CI 1.128–1.788, P = 0.003), whereas higher proportions of Tpex and Tem cells showed protective associations. These heterogeneous subset-level associations suggested that no single subset proportion fully captured the broader transcriptional variation underlying CD8⁺ T-cell states. In the independent GSE206325 neoadjuvant anti-PD-1 cohort, pathological responders showed higher Tpex cell proportions than pathological non-responders (P = 0.024), whereas the other CD8⁺ T-cell subsets showed no significant differences (Supplementary Figure 2A–C). These findings are consistent with an association between progenitor-like CD8⁺ T-cell states and pathological response to anti-PD-1 therapy.

Development of the TCSP Score from the CD8⁺ T-Cell State Axis

We identified genes showing monotonic expression changes along the inferred CD8⁺ T-cell state axis. To enrich the candidate set for CD8⁺ T-cell-associated signals, axis genes showing ordered-state-associated downregulation were intersected with genes expressed at higher levels in CD8⁺ T cells than in other major cell populations (Figure 2A and B). This intersection yielded 66 CD8⁺ T-cell-enriched candidate genes (Supplementary Table 2).

Figure 2.

Construction of the TCSP prognostic model. (A) Volcano plot of differential gene expression between CD8⁺ T cells and other cell types. (B) Intersection of CD8⁺ T-cell-enriched genes with progressively downregulated state-axis genes, yielding 66 candidate genes. (C) Univariate Cox regression analysis of candidate genes in the training cohort assessing associations with overall survival. (D) LASSO-Cox regression coefficient trajectories across penalty parameters (λ). (E) Tenfold cross-validation curves of partial likelihood deviance, with λ_min indicating the optimal penalty parameter. (F) Selected genes included in the final model and their corresponding LASSO coefficients. (G) Kaplan–Meier survival curves comparing overall survival between high- and low-risk groups stratified by the median TCSP score. (H) Time-dependent ROC curves showing prognostic performance of the TCSP score. (I) Multivariable Cox regression analysis including age, sex, and tumor stage, demonstrating independent association of TCSP with survival.

In the training cohort, the 66 candidate genes were first evaluated using univariate Cox regression to identify genes and assess associations with overall survival. To retain potentially relevant variables, a threshold of P < 0.10 was applied for initial feature screening (Figure 2C). LASSO-Cox regression was subsequently performed for feature selection and coefficient estimation. The final model was developed using the optimal penalty parameter (λ_min) selected by cross-validation (Figure 2D and 2E). For each patient, the TCSP score was calculated as the weighted sum of standardized expression values of the selected genes and their corresponding regression coefficients (Figure 2F). Patients were stratified into high- and low-TCSP groups using the cohort median score. Kaplan–Meier analysis revealed significantly shorter OS among patients with high TCSP scores compared with those with low scores (Figure 2G). Time-dependent ROC analysis further supported the prognostic value of the model, yielding AUCs of 0.643, 0.743, and 0.721 at 1, 3, and 5 years, respectively (Figure 2H). After adjustment for age, sex, and tumor stage, multivariate Cox regression analysis verified that the TCSP score remained independently associated with overall survival. Specifically, a high TCSP score remained significantly associated with worse OS (Figure 2I).

The TCSP Score Showed Robust Prognostic Stratification Capacity in Independent HCC Cohorts

In the TCGA-LIHC cohort, patients with high TCSP scores had significantly shorter overall survival (P = 0.00011) (Figure 3A). Time-dependent ROC analysis demonstrated the prognostic performance of the TCSP score (Figure 3B). Multivariate Cox regression analysis confirmed that the TCSP score remained independently associated with overall survival after adjustment for age, sex, and tumor stage (Figure 3C). Similar results were observed in the ICGC-LIRI-JP cohort, where high TCSP scores were also associated with shorter overall survival (Figure 3D). Time-dependent ROC analysis further confirmed the discriminative performance of the TCSP score (Figure 3E). Multivariate Cox regression analysis demonstrated that the TCSP score remained independently associated with OS after adjustment for clinical variables available in the ICGC-LIRI-JP cohort (Figure 3F).

Figure 3.

External validation of TCSP-associated prognostic stratification. (A) Kaplan–Meier survival analysis in the TCGA-LIHC cohort stratified by the median TCSP score. (B) Time-dependent ROC curves in the TCGA-LIHC cohort. (C) Multivariable Cox regression analysis in TCGA-LIHC demonstrating independent association of TCSP with overall survival. (D and E) Kaplan–Meier and time-dependent ROC analyses in the ICGC-LIRI-JP cohort showing survival stratification by TCSP. (F) Multivariable Cox regression in the ICGC-LIRI-JP cohort showing independent association of TCSP with survival.

We further evaluated whether the TCSP score could stratify recurrence-free survival (RFS). Patients with high TCSP scores had significantly shorter RFS in both the GSE14520 and TCGA-LIHC cohorts (Supplementary Figure 3A and C). Time-dependent ROC analysis indicated moderate discriminative ability for RFS at multiple follow-up time points (Supplementary Figure 3B and D).

Association of the TCSP Score with the Tumor Immune Microenvironment and Pathological Response

We further characterized the tumor immune microenvironment across TCSP-defined groups by comparing immune-related gene expression profiles. Immune receptors, costimulatory molecules, and MHC-related genes were generally more highly expressed in low-TCSP tumors, whereas their expression was reduced in the high-TCSP tumors. Chemokines and their receptors were also expressed at higher levels in low-TCSP tumors (Figure 4A). We then examined immune cell infiltration using immune deconvolution analysis. The low-TCSP group showed higher estimated infiltration of several antitumor immune cell types, including memory T cells, natural killer (NK) cells, activated B cells, activated CD8⁺ T cells, and dendritic cells. The high-TCSP group showed lower infiltration of these immune cell populations (Figure 4B). These findings indicate that lower TCSP scores are associated with a more immune-active tumor immune microenvironment.

Figure 4.

Association of the TCSP score with tumor immune microenvironment characteristics and immunotherapy response. (A) Heatmap of immune-related gene expression comparing high- and low-TCSP groups. (B) Immune cell deconvolution analysis comparing high- and low-TCSP groups. (C) Functional immune scores (IPS, TLS, CYT) by TCSP group. (D) Proportions of pathological responders and pathological non-responders according to TCSP stratification in an independent neoadjuvant anti-PD-1-treated HCC cohort. (*P < 0.05, **P < 0.01, ***P < 0.001, ****P < 0.0001 and ns, not significant.).

We also compared several composite immune scores between the TCSP risk groups. The low-TCSP group showed higher immune phenotype scores (IPS), tertiary lymphoid structure (TLS) scores, and cytolytic activity (CYT) scores than the high-TCSP group (Figure 4C). We then analyzed an independent neoadjuvant anti-PD-1-treated cohort. After stratification by the TCSP score, the proportions of pathological responders were compared between low- and high-TCSP groups. The low-TCSP group showed a numerically higher proportion of pathological responders than the high-TCSP group (41.67% versus 16.67%); however, this difference was not statistically significant (Fisher’s exact test, P = 0.37; Figure 4D). These findings suggest a numerical trend toward a higher pathological response rate in patients with low TCSP scores, which requires validation in larger independent cohorts.

To determine whether TCSP score merely reflected total T-cell abundance, we performed an additional analysis in the TCGA-LIHC cohort. The TCSP score showed only a weak correlation with MCP-counter-estimated total T-cell abundance (Spearman ρ = −0.137, P = 0.011). After adjustment for total T-cell abundance, age, sex, and AJCC stage, each one-standard-deviation increase in the TCSP score remained associated with higher mortality risk (HR = 1.531, 95% CI 1.277–1.835, P < 0.001), whereas total T-cell abundance was not independently associated with OS in the same model (HR = 0.995, P = 0.960). Adding the TCSP score increased the model C-index from 0.622 to 0.671 (Supplementary Figure 4). These findings indicate that the TCSP score contains prognostic information beyond estimated total T-cell abundance, although the bulk-tissue score cannot be interpreted as a purely CD8⁺ T-cell-intrinsic measure.

Finally, tissue-level expression of CD247, PRKCQ, and PIP4K2A was examined by RT-qPCR in 20 paired HCC tumors and adjacent non-tumor liver tissues from an independent clinical cohort. HCC tissues had significantly higher ΔCt values for CD247 (P < 0.01) and PRKCQ (P < 0.05), indicating lower relative expression of both genes than in paired adjacent non-tumor tissues. In contrast, PIP4K2A had a significantly lower ΔCt value in HCC tissues (P < 0.01), indicating higher relative expression (Supplementary Figure 5). These findings support differential tissue-level expression of selected TCSP-associated genes, but do not validate TCSP prognostic performance.

Discussion

Tumor plasticity is widely regarded as a central driver of malignant tumor progression.^35,36 In recent years, the concept of plasticity has been extended beyond tumor cells to include immune cells within the tumor microenvironment. Not only the extent of T-cell infiltration but also the specific cellular states of these T cells are critical determinants of antitumor immune responses and patient prognosis.⁸ However, molecular markers capable of comprehensively capturing T-cell states and predicting clinical outcomes remain limited.^37,38 In this study, we developed a single-cell-derived TCSP score that captures a transcriptional program associated with the inferred CD8⁺ T-cell state continuum and demonstrated its prognostic relevance across independent HCC cohorts. Accordingly, TCSP should be interpreted as a bulk transcriptomic surrogate of this single-cell-derived state program rather than a direct measure of functional plasticity, self-renewal, or lineage transition. Higher TCSP scores were consistently associated with shorter OS and RFS across independent retrospective HCC cohorts.

The combined analyses support an ordered transcriptional organization of tumor-infiltrating CD8⁺ T-cell states. Monocle3 provided an unsupervised representation of the state continuum, whereas supervised psupertime analysis showed that held-out transcriptional profiles reproducibly recapitulated the predefined ordinal state axis. The substantial overlap and frequent neighboring-state assignments observed for Tem, Tpex, and Ttex cells further suggest that these populations are not separated by sharp transcriptional boundaries. Together, these results connect the inferred CD8⁺ T-cell state program with immune heterogeneity and clinical outcomes without establishing direct functional or lineage-level plasticity. A key strength of this study is the translation of a single-cell-derived state program into a score that can be evaluated in clinically annotated bulk cohorts. High TCSP scores were associated with adverse survival outcomes, whereas low-TCSP tumors showed higher expression of immune-related molecules, greater estimated infiltration of several immune cell populations, and higher IPS, TLS, and CYT scores. These findings suggest that the state-associated program captured by TCSP is linked to broader immune activity within the tumor microenvironment.¹⁰ Because immune-cell abundance could confound this relationship, we additionally adjusted for MCP-counter-estimated total T-cell abundance. TCSP showed only a weak correlation with total T-cell abundance and remained associated with OS after adjustment for total T-cell abundance and clinical covariates. Thus, TCSP is not equivalent to total T-cell abundance and provides incremental prognostic information, although its derivation from bulk tissue means that some compositional influence remains. The association between TCSP and distinct immune landscapes suggests that reduced inferred CD8⁺ T-cell plasticity is linked to a less immunologically active tumor microenvironment, which may partly explain the adverse prognosis observed in high-TCSP tumors. The numerically lower proportion of pathological responders in the high-TCSP group is consistent with this interpretation but remains exploratory given the lack of statistical significance and the limited cohort size. The observed association between TCSP and RFS further suggests a potential link between CD8⁺ T-cell plasticity and recurrence risk in HCC.

RT-qPCR analysis of 20 paired HCC tumor and adjacent non-tumor liver tissues showed lower relative expression of CD247 and PRKCQ and higher relative expression of PIP4K2A in HCC tissues. CD247 and PRKCQ have negative coefficients in the TCSP formula, whereas the qPCR analysis independently showed lower tissue-level abundance of these transcripts in HCC. However, the coefficient signs represent conditional contributions to a multigene survival predictor, whereas the qPCR analysis compared tumor with adjacent non-tumor liver tissues. Their directional correspondence therefore does not establish a causal or mechanistic relationship. Likewise, higher PIP4K2A expression in HCC supports altered tissue-level expression but does not establish its prognostic role or cellular origin. Because bulk tissues contain malignant, immune, and stromal cells, this experiment cannot determine whether the observed differences originated from CD8⁺ T cells. It also does not validate TCSP prognostic performance or directly demonstrate functional state plasticity. Similarly, the negative LASSO-Cox coefficient of CCND3 (−0.3784) should not be interpreted as evidence that CCND3 functions as a tumor suppressor. In a penalized multivariable model, the sign of an individual coefficient represents its conditional association with the outcome after accounting for the other model genes and coefficient shrinkage. CCND3 expression may also arise from malignant cells, proliferating immune cells, or other tumor-microenvironment populations. The additive TCSP formula is therefore a statistical linear predictor on the log-hazard scale rather than a mechanistic representation of the nonlinear biological network controlling immune differentiation.

Several limitations of this study should be acknowledged. First, all analyses were based on public datasets, and potential platform heterogeneity may exist. Second, the TCSP score was calculated from bulk RNA sequencing data and may therefore incorporate transcriptional signals from non-CD8⁺ T-cell populations. Third, although single-cell analysis identified genes associated with CD8+ T cell plasticity, their functional roles remain to be experimentally validated.

Overall, this study provides a framework for interpreting tumor immune heterogeneity in HCC through the lens of CD8⁺ T-cell state plasticity and establishes a scoring system that enables assessment of CD8⁺ T-cell state programs and their associations with disease progression and clinical outcomes. These findings support the clinical relevance of CD8⁺ T-cell state plasticity in HCC and suggest that TCSP may serve as a practical framework for linking immune-state heterogeneity to prognosis and features of the tumor immune landscape.

Funding Statement

This research was supported by the National Natural Science Foundation of China (82372714). The open research funds from the Affiliated Qingyuan Hospital (Qingyuan People’s Hospital), Guangzhou Medical University (No. 20230328 to Linhai Li and Ming Liu).

Ethics Approval

The publicly available datasets analyzed in this study included GSE14520 and GSE206325 from the Gene Expression Omnibus, TCGA-LIHC from The Cancer Genome Atlas, and LIRI-JP from the International Cancer Genome Consortium. The controlled-access dataset was obtained from GSA-Human after approval of the data-access application and was analyzed in accordance with the GSA-Human Data Access Agreement. Only de-identified or anonymized data were analyzed, and no attempt was made to re-identify individual participants. The original studies and data providers were responsible for obtaining the applicable ethical approvals and informed consent for the original collection and generation of these data.

According to Article 32 of the Measures for Ethical Review of Life Science and Medical Research Involving Humans issued in China in 2023, research using legally obtained public data or anonymized information data may be exempted from ethical review when it does not cause harm to humans and does not involve sensitive personal information or commercial interests. Therefore, additional institutional ethical approval and informed consent were not required for this secondary analysis.

Paired primary HCC tumor and adjacent non-tumor liver tissues were obtained from Guangzhou Medical University. All procedures involving human specimens were conducted in accordance with the institutional guidelines for human research. Written informed consent was obtained from all patients for the use of their clinical specimens for medical research. All studies were approved by the review board of Guangzhou Medical University.

Author Contributions

All authors contributed to the conception and design of the study. Le Xing and Ming Liu initiated and conceived the project. Le Xing drafted the manuscript. Dan He, Dong-Mei Li, Chen Yue, and Zhong Liang were responsible for data collection and data analysis. Cheng-Yang Zhang provided guidance on data analysis. Ying-Chen Hou, Chen-Yu Jiang, and Guang-Ding Zhou contributed to data interpretation and manuscript revision. Hongjiao Yu, Yan-Ling Zhang, Yan-Xia Shi provided valuable comments and suggestions. Lin-Hai Li, and Ming Liu supervised the study and critically revised the manuscript. All authors made a significant contribution to the work reported, whether that is in the conception, study design, execution, acquisition of data, analysis and interpretation, or in all these areas; took part in drafting, revising or critically reviewing the article; gave final approval of the version to be published; have agreed on the journal to which the article has been submitted; and agree to be accountable for all aspects of the work.

Disclosure

Chen-Yu Jiang reports Support for the manuscript from data interpretation and manuscript revision, during the conduct of the study. All authors declared that there are no other conflicts of interest in this work.