Tumor subtype-specific transcriptional signatures predict recurrence risk in hepatocellular carcinoma Authors: Hyeran Shim, Hoang Bao Khanh Chu, Jiun Kim, Seohyeon Park, Subin Yoo, Jeongsil Kim-Ha, Young-Joon Kim Published: 2026 ABSTRACT AbstractHepatocellular carcinoma (HCC) displays significant molecular heterogeneity that clinical staging alone does not fully account for when assessing the risk of recurrence following curative treatment. To address this, we developed a multi-cohort survival modeling framework that translates bulk RNA-seq profiles into biologically meaningful tumor lineage scores. These scores are then used to stratify disease-free survival (DFS). Our study included 1,059 patients from four independent cohorts, where tumor immunogenic and proliferative lineage scores satisfied the proportional hazards assumptions and were integrated into a cohort-stratified Cox model. Each patient received a linear predictor (LP), and predefined cutpoints were established to categorize them into actionable risk groups. These groups exhibited consistent DFS separation in both the training (70%) and independent test (30%) sets. In ROC analyses, the LP demonstrated moderate overall discrimination (AUC 0.629, 95% CI 0.595-0.662). It showed higher discrimination for early recurrence (< 1 year; AUC 0.653, 95% CI 0.591-0.715) and lower discrimination for later recurrence (3-5 years; AUC 0.576, 95% CI 0.473-0.678). This approach establishes a biologically informed framework for stratifying recurrence risk based on RNA-seq data, potentially enhancing risk-adapted surveillance and postoperative management for HCC.Keywords: Hepatocellular carcinoma, Recurrence risk, Risk stratification, RNA-seq, Stratified Cox proportional hazards model INTRODUCTION Hepatocellular carcinoma (HCC) is a leading cause of cancer mortality, and recurrence following curative-intent treatment is common (1). While tumor stage and clinical factors are essential for prognosis, they do not fully capture the biological diversity that influences post-treatment relapse (2). RNA sequencing (RNA-seq) allows for the direct quantification of transcriptional programs. Recent research has demonstrated that cell-type–resolved marker sets—covering tumor epithelial lineages and key immune and stromal compartments—can simplify complex expression profiles into interpretable per-sample scores (3). We present a multi-cohort modeling framework that transforms bulk RNA-seq data into biologically interpretable tumor lineage scores and assesses their association with disease-free survival (DFS) through cohort-stratified Cox regression. Our approach involves: (i) calculating immunogenic and proliferative lineage scores for each sample using a rank-based single-sample scoring method, (ii) evaluating proportional hazards assumptions to support time-varying effects, (iii) modeling coefficients and baseline survival estimates in the training set, and (iv) establishing risk-group cutpoints to facilitate prospective application without the need for retraining. This work aims to develop a recurrence risk stratification tool that is applicable across various institutions. RESULTS Cohorts, endpoints, and data readiness We assembled a harmonized multi-cohort HCC dataset with DFS as the primary endpoint, defined by time-to-event and event indicators. To account for site heterogeneity and prevent information leakage, we used the cohort as a prespecified stratification factor (Fig. 1B-E). This stratification was incorporated into the subsequent survival analysis using a stratified Cox model. RNA-seq matrices were transformed into cell-type–based marker-expression scores, which included tumor-intrinsic lineage programs (tumor proliferative lineage and tumor immunogenic lineage). This transformation was achieved through a validated marker gene-set scoring approach to derive the scores (4). The scores were combined with clinical information (age, stage, etiology) to create both cohort analysis tables and a pooled, stratified Cox model table for modeling. The rank-based scoring system produced consistent cell-type scores across batches, and the standardized differences remained below conventional imbalance cutoffs. Overall analysis pipeline and clinical characteristics of analyzed cohorts. (A) Modeling pipeline overview. The workflow encompasses data collection, cell-type marker expression rank-based scoring from RNA-seq, proportional hazards verification, feature selection, and stratified Cox model training with clinical application for risk group definition. (B-E) Clinicopathological characteristics of four independent HCC cohorts. Donut charts display the distribution of key clinical variables: age at diagnosis (< 60 years vs. ≥ 60 years), disease-free survival status (recurrence: 0 = no event, 1 = recurrence/death), the American Joint Committee on Cancer (AJCC) stage (I, II, III, IV), and liver disease etiology (HBV = hepatitis B virus, HCV = hepatitis C virus, NBNC = non-HBV non-HCV including alcohol-related and metabolic dysfunction-associated steatohepatitis, ALCOHOL, Unknown). Sample sizes (n) and percentages are annotated for each category. (B) AMC cohort (n = 206). (C) CGRC cohort (n = 310). (D) ICGC cohort (n = 203). (E) TCGA cohort (n = 340). Proportional hazards assessment and survival-associated coefficient screening We conducted a systematic comparison of proportional hazards (PH) (5) and time-varying (TV) Cox models (6) for immunogenic and proliferation scores using 70% of HCC samples from four cohorts, resulting in a training sample size of n = 741. An automated selection procedure, combined with cross-validation and Schoenfeld testing (7), was used as a gatekeeper to determine that both markers were PH-adequate, thereby supporting a proportional hazards specification with constant coefficients. The final multivariable PH model exhibited greater statistical significance compared to the TV model. Both the immunogenic and proliferation scores yielded more significant P-values under the PH assumption (P < 0.05 for both markers in PH vs. P > 0.05 in TV), further validating the use of the simpler proportional hazards framework (Fig. 2A, B). Cox model development, validation, and risk stratification performance. (A) Forest plot indicating TV models for immunogenic and proliferation scores. Error bars represent 95% confidence intervals. The vertical line at HR = 1.0 indicates no effect. LE: less than or equal to, GT: greater than. (B) Final multivariable Cox PH model forest plot showing hazard ratios for the two cell-type marker scores in the Cox model. (C) Relationship between LP and predicted 1-year survival probability (S365) in the training cohort (n = 741). (D) Model discrimination by recurrence status in the training cohort. Left panel: box plot of differences in LP scores in patients experiencing recurrence versus non-recurrent patients. Right panel: box plot of differences in predicted 1-year survival probability scores in patients experiencing recurrence versus non-recurrent patients. (E) Kaplan-Meier survival curves for locked risk groups in the training cohort (n = 741) based on global cut-off. Patients were stratified into tertiles based on linear predictor values: Low risk (orange, n = 233), Mid risk (green, n = 290), and High risk (blue, n = 218). (F) Kaplan-Meier survival curves for locked risk groups in the test cohort (n = 318) based on global cut-off. Patients were stratified into tertiles based on linear predictor values: Low risk (orange, n = 100), Mid risk (green, n = 143), and High risk (blue, n = 75). Risk stratification using the combined score model revealed significant survival differences in the test sample set, which consisted of over 30% of the samples from the four cohorts (n = 318). This was evident for both binary classifications (high vs. low risk; log-rank P = 0.0133) (Supplementary Fig. 1A) and tertile classifications (low/mid/high risk groups) (Supplementary Fig. 1B), thereby validating the prognostic value of these cell-type scores. Model training and prediction analysis We developed a Cox proportional hazards model using validated survival coefficients and an optimized feature scaling methodology to enhance prognostic discrimination while maintaining the validity of statistical inferences. To address the insufficient variation in predicted survival probabilities seen with standard implementations, we employed a systematic feature scaling approach. Model performance was evaluated using the training cohort, revealing a significant relationship between the linear predictor (LP) values and the predicted 1-year survival probability, with a strong negative correlation (R2 = 0.97, P < 2.2 × 10−16) that aligns with theoretical predictions of the Cox model (Fig. 2C). The nearly perfect correlation between the linear predictor and survival probability validates the feature scaling method as an effective way to improve the model’s discriminatory power without sacrificing statistical validity or interpretability of coefficients. Patients with recurrence exhibited significantly higher linear predictor scores (median: 0.0, IQR: −0.3 to 0.3) compared to non-recurrent patients (median: −0.3, IQR: −0.7 to 0.1; Wilcoxon test, P = 4.6 × 10−10) (Fig. 2D, left panel). Additionally, the predicted 1-year survival probabilities were significantly lower in the recurrence group (median S365: 0.75, IQR: 0.66-0.82) than in the non-recurrence group (median S365: 0.80, IQR: 0.72-0.86; P = 4.6 × 10−10) (Fig. 2D, right panel). The model’s performance was further validated in the test cohort, where the linear predictor continued to show a significant association with the predicted 1-year survival probability (R2 = 0.97, P < 2.2 × 10−16), confirming the robustness of our risk stratification approach. Cut-point locking and risk-group separation Using the LP from the refitted model, we established fixed cut points—both global and, when applicable, cohort-local—to define clinically relevant risk groups. For the three-group partitioning, we conducted a grid search that enforced minimum group proportions and minimum events per group. The resulting groups exhibited clear and consistent survival differences across the entire cohort, with Kaplan–Meier curves demonstrating excellent ordering and significant stratified log-rank statistics for both the training (P = 1.4 × 10−8) and test datasets (P = 4.3 × 10−5) (Fig. 2E, F). Additionally, the cohort-local cut points revealed meaningful risk group stratifications across all four cohorts (Supplementary Fig. 2C-F). Final thresholds, group proportions, and event counts are summarized in Supplementary Table 1, 2, confirming the finalization of these cut points for deployment. Prospective deployment and independent audit To evaluate the temporal stability and clinical utility of the linear predictor across various follow-up intervals, we conducted a time-dependent receiver operating characteristic (ROC) analysis, stratified by the timing of recurrences (8) (Fig. 3). The model demonstrated moderate discriminatory capacity for predicting disease-free survival over various time periods. For all patients, the linear predictor achieved an area under the curve (AUC) of 0.629 (95% CI: 0.595-0.662), indicating moderate overall predictive accuracy. When stratified by recurrence timing, the model showed varying performance across different temporal subgroups: it exhibited optimal discrimination for early recurrence (< 1 year; AUC = 0.653, 95% CI: 0.591-0.715), consistent performance for intermediate recurrence (1-3 years; AUC = 0.650, 95% CI: 0.594-0.705), and reduced yet still meaningful discrimination for late recurrence (3-5 years; AUC = 0.576, 95% CI: 0.473-0.678). The progressive decline in AUC with increasing follow-up duration (from 0.653 to 0.576) suggests that the gene signature-based risk score more effectively captures the biological determinants of early aggressive recurrence, compared to late relapse events, which may be influenced by further tumor evolution, changes in the microenvironment, or treatment-related factors not reflected in baseline molecular profiles (9). The higher AUC for early recurrence patients is clinically significant, as these high-risk individuals may benefit the most from intensified surveillance protocols or adjuvant therapeutic interventions (10). These findings show that the model remains predictive throughout the follow-up period, with its strongest discriminatory power for identifying patients at imminent risk of disease progression. This supports its potential use in personalized, risk-adapted clinical management strategies. ROC curves for recurrence prediction. ROC curves demonstrate the discriminatory capacity of the linear predictor for disease-free survival across different follow-up intervals. The curves represent: all patients (black), patients with recurrence < 1 year (red), patients with recurrence 1-3 years (blue), and patients with recurrence 3-5 years (green). The diagonal reference line (gray) represents random chance (AUC = 0.5). DISCUSSION In this study, we developed and validated a cell-type–resolved molecular risk stratification model for hepatocellular carcinoma that integrates tumor-intrinsic transcriptional programs with rigorous survival modeling methodologies. By leveraging RNA-seq–derived marker-expression scores that represent immunogenic and proliferative tumor compartments, our approach distills complex cancer features into clinically actionable risk predictions while maintaining biological interpretability. The model demonstrated moderate discriminatory capacity across independent test cohorts, with the strongest performance observed for early recurrence prediction (AUC = 0.653 for < 1 year). This addresses a critical unmet need in identifying high-risk patients who are most likely to benefit from intensified surveillance or adjuvant therapy (11). Our methodological framework includes several design elements aimed at enhancing clinical translatability beyond traditional Cox regression methods. By implementing cohort-stratified splitting during model training, we explicitly prevent information leakage that can artificially inflate performance estimates—an issue often encountered when pooling multi-institutional data without considering site-level characteristics. We treat cohort as a stratification factor rather than a covariate, which allows our model to estimate hazard ratios averaged across sites while permitting baseline hazards to vary freely. This approach accommodates differences in patient selection, surgical techniques, and follow-up protocols across institutions. Additionally, our locked cutpoint approach, established through systematic optimization rather than post-hoc selection, offers reproducible risk group definitions that can be applied to prospective patients without needing to retrain the model. This contrasts sharply with studies that define risk groups using median splits or quartiles in discovery cohorts, as those methods often produce unstable classifications when applied to independent populations with varying risk distributions (12). Furthermore, our comprehensive audit suite, which includes AUC analysis and subgroup stratifications, offers transparent performance reporting across various patient characteristics. This allows clinicians to evaluate the reliability of the model for specific patient subpopulations encountered in practice. In conclusion, we present a biologically grounded and methodologically rigorous risk stratification model for HCC that translates cell-type–resolved transcriptional programs into clinically applicable survival predictions. The model incorporates internal cross-validation, automated proportional hazards verification, and locked cut-point deployment, ensuring its performance across independent cohorts while maintaining biological interpretability through immunogenic and proliferative marker scores. With superior performance in predicting early recurrence and consistent risk stratification across diverse patient populations, this approach serves as a promising tool for personalized postoperative management. By identifying high-risk patients who are most likely to experience rapid disease progression, our framework facilitates rational triage for adjuvant therapies and intensive surveillance strategies, ultimately advancing the implementation of precision oncology in HCC (13). Prospective clinical validation through randomized trials, integration with complementary biomarker modalities, and mechanistic analysis of molecularly defined risk groups will help refine this framework and establish its role in evidence-based HCC care pathways (14). MATERIALS AND METHODS Cohorts, endpoints, and candidate predictors We compiled a harmonized multi-cohort dataset consisting of 1,059 hepatocellular carcinoma samples from four independent cohorts: AMC (GSE124751 & GSE124752) (n = 206), CGRC (n = 310), ICGC (n = 203), and TCGA (n = 340). DFS was the primary endpoint, defined by the time to recurrence and recurrence status. RNA-seq data were transformed to log2(TPM+1) for each cohort prior to calculating the cell-type score using the singscore function (4). Clinical variables, including patient age, tumor stage, and etiology, were combined with cell type scores to create a comprehensive analytical dataset. Train-Test split and leakage prevention To ensure robust validation while accounting for inter-cohort heterogeneity, we employed a cohort-stratified proportional splitting strategy. We divided the samples into training (70%, n = 741) and testing (30%, n = 318) sets, ensuring that each cohort and their respective event rates were proportionally represented in both partitions. This method prevented information leakage by allowing all four cohorts to contribute to both the training and testing sets, resulting in the following final distributions: AMC (144 train/62 test), CGRC (217/93), ICGC (142/61), and TCGA (238/102). All subsequent model development, parameter tuning, and feature selection were conducted solely on the training set. Proportional hazards assessment and survival-associated coefficient screening To assess the temporal stability of predictor effects, we implemented an automated PH assessment using Schoenfeld residual tests on rank-transformed values: markers with P > 0.05 (indicating satisfied PH assumptions) were modeled with time-invariant coefficients, while those with P ≤ 0.05 required time-varying effects. Both the immunogenic and proliferation scores strongly met the PH assumptions, with Schoenfeld P-values of 0.97 and 0.976, respectively, allowing us to retain them with constant hazard ratios. Feature selection was performed using a 5-fold cross-validated concordance index optimization on the training set, with both markers included as forced-entry variables. The final multivariable Cox model was fitted using cohort-proportional stratification to address baseline hazard heterogeneity across study sites, and model coefficients were locked for subsequent risk prediction and validation analyses. Model construction and linear predictor (LP) score derivation After determining the coefficients, a final Cox proportional hazards model was constructed using the complete training dataset (n = 741) with the locked feature set and their corresponding regression coefficients. Patient-level risk scores (linear predictors) were calculated as the weighted sum of scaled marker values using the fixed coefficients: LP = β1 × (Immunogenic_scaled) + β2 × (Proliferation_scaled), where β1 and β2 are time-invariant regression coefficients obtained from prior proportional hazards assessments. Individual survival probabilities at 1 year (S365) were calculated based on the baseline cumulative hazard, with patient-specific adjustments derived from their linear predictor values: S365(individual) = S0(365)^exp(LP). The final model, which includes baseline survival curves and regression coefficients with P-values, was then applied to independent test samples. KM-based cut-point locking with cohort-aware options To convert continuous linear predictor values into clinically actionable risk categories, we used a cut-point optimization procedure on the training set (n = 741) to identify thresholds that best distinguished survival outcomes. For binary risk stratification (high vs. low risk), we applied a 5-fold cross-validation approach to evaluate candidate cut-points across quantiles from the 5th to the 95th percentiles. The final cut-point was chosen as the median threshold across folds that corresponded to the quantile position maximizing the mean cross-validated χ2 statistic. For tertile risk stratification (low/middle/high risk), we assessed all pairwise combinations of cut-points using multivariate log-rank tests to find the threshold pair that maximized survival separation. To ensure clinical viability and statistical robustness, we imposed minimum group size constraints (≥ 20% of total samples per group) and minimum event requirements (≥ 30 events per group for global cut-points). This method was applied to all training samples combined (n = 741) to determine global cut-points for universal application, as well as separately within each cohort to identify cohort-specific local cut-points that reflect site-specific score distributions. The optimized cut-points were then locked and validated in the independent test set (n = 318). Supplementary Material ACKNOWLEDGEMENTS This research was supported by the Bio & Medical Technology Development Program of the National Research Foundation (NRF) funded by the Ministry of Science & ICT (grant number: NRF-2017M3A9A7050614). Footnotes ETHICS APPROVAL This study was approved by the Institutional Review Board of Seoul National University Hospital (IRB No.: H-1602-016-739). CONFLICTS OF INTEREST The authors have no conflicting interests. REFERENCES Associated Data Supplementary Materials