Review Reports - Biomarker-Based Precision Prediction of Immunotherapy Response in Hepatocellular Carcinoma

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The perfect classification performance (AUC = 1.0) in both training and testing sets suggests possible overfitting, especially given the small combined sample size (n=35). More details on how the dataset was split for training/testing and whether stratified sampling or nested cross-validation was used would improve reliability.

HOXA10 and CNTN3 are discussed, PLPPR1, HAGLR, and ENPP3 receive limited mechanistic interpretation

A few typographical errors (“re- sponse”, “ComBat- corrected”, etc.) appear due to line-break artifacts. A proofread will resolve formatting issues.

Check uniformity of gene name capitalization (e.g., HOXA10 vs HAGLR).

Author Response

Comment 1:

Response 1:

We thank the reviewer for this constructive comment and agree that perfect classification performance obtained from a single train/test split should be interpreted cautiously, particularly given the limited sample size.

In the revised manuscript, we have clarified the data splitting procedure and added additional validation analyses to better assess model stability. The initial training/testing evaluation was performed using a predefined split to obtain an initial estimate of discriminative performance. To further reduce potential bias from a single split, we subsequently conducted repeated stratified cross-validation (5-fold cross-validation repeated 100 times) using a fixed, predefined Signature Score. Importantly, feature selection and coefficient estimation were not repeated during cross-validation, thereby minimizing the risk of information leakage.

Across repeated resampling, the model demonstrated highly consistent AUC values, supporting the robustness of the Signature Score within this cohort. These results are now presented in Supplementary Figure 1B–C, and the corresponding methodological details have been added to the Materials and Methods section. The Results, Discussion and Conclusions text has also been revised to avoid overstatement and to emphasize internal validation rather than generalizability.

The corresponding revisions have been incorporated into the Abstracts section (page 1).

Abstracts

Background: Hepatocellular carcinoma (HCC) remains a major global health challenge with limited treatment options for advanced disease. Although immune checkpoint inhibitors (ICIs) have shown clinical benefits, response rates remain low, emphasizing the need for reliable biomarkers to guide patient selection. Given the critical role of metabolic reprogramming in immune modulation, this study aimed to identify a metabolic gene signature predictive of immunotherapy response in HCC. Methods: Three independent transcriptomic datasets (GSE279750, GSE215011, and GSE235863) comprising 35 ICI-treated HCC samples were integrated after quality control and ComBat batch correction. Differentially expressed genes were identified using DESeq2 and limma, followed by integration of the meta-analysis results. Machine learning models, including LASSO regression and random forest algorithms, were applied for feature selection, and a logistic regression model was developed for predictive scoring. Results: A five-gene metabolic signature (PLPPR1, CNTN3, HOXA10, HAGLR, and ENPP3) demonstrated good discriminative ability between responders and non-responders, with consistent performance observed across internal validation analyses. Functional enrichment analysis revealed significant involvement of metabolic pathways, with HOXA10 linked to immune evasion and CNTN3 associated with immune activation. Conclusions: This five-gene signature represents a biologically interpretable biomarker panel with potential utility for immunotherapy response stratification in HCC. The integrative analytical framework provides preliminary evidence supporting its value, warranting further validation in larger, independent clinical cohorts before clinical translation.

The corresponding revisions have been incorporated into the Materials and Methods

section (page 4, lines 150-156).

“To further mitigate potential bias from a single data split, additional internal validation was conducted using repeated stratified cross-validation (5-fold cross-validation repeated 100 times). Importantly, feature selection and coefficient estimation were fixed prior to cross-validation, and only the predefined Signature Score was evaluated during resampling, thereby minimizing the risk of information leakage. Differences in Signature Score distributions between responders and non-responders were assessed using the Wilcoxon rank-sum test.”

The corresponding revisions have been incorporated into the Result section (page 10, lines 296–308, 327-353).

“To evaluate the predictive performance, we constructed a logistic regression model based on the five-gene signature and first assessed its discriminative ability using a predefined training and testing split. ROC analysis demonstrated strong discriminative performance in the initial training/testing evaluation (Figure S1A). Given the limited sample size, we further evaluated the stability of this result using repeated stratified cross-validation with a fixed Signature Score. Across resampling iterations, the model showed highly consistent discriminative performance, yielding consistently high AUC values (Figure S1B–C). Collectively, these results establish a robust five-gene predictive signature derived from dual-algorithm feature selection, supported by consistent performance across independent testing and repeated resampling, underscoring the internal robustness of the integrated machine learning framework for predicting immunotherapy response within the current cohort.

“3.5 The Five-Gene Signature Score Demonstrates Consistent Predictive Performance with Internal Stability

Based on the machine learning–based feature selection results, the top five genes (PLPPR1, CNTN3, HOXA10, HAGLR, and ENPP3) were incorporated into a logistic regression model. Using the integrated transcriptomic dataset, regression coefficients were optimized through maximum likelihood estimation to model discrimination between responders and non-responders. The resulting five-gene signature score was defined as follows:

Signature Score = (−194.836 × PLPPR1) + (147.927 × CNTN3) − (326.820 × HOXA10) − (2.582 × HAGLR) − (41.937 × ENPP3)

In this model, the positive coefficient of CNTN3 indicates a positive association between its expression and immunotherapy response, whereas HOXA10 exhibited the largest negative coefficient, identifying it as a major contributor to non-response. PLPPR1 and ENPP3 also showed negative associations with response, while HAGLR contributed a smaller effect. The absolute coefficient magnitudes reflect the relative contributions of individual genes to the composite score. Receiver operating characteristic (ROC) analysis showed that the five-gene signature score achieved complete separation between responders and non-responders in the integrated cohort, yielding an observed AUC of 1.0 (Figure 6A), exceeding individual gene performance (AUC range: 0.675–0.841). Consistent with this, box plot visualization demonstrated distinct signature score distributions between groups without overlap (p = 1.35 × 10⁻⁹, Wilcoxon rank-sum test; Figure 6B). To assess internal stability, bootstrap validation with 1,000 resampling iterations was performed (Supplementary Figure 1D). The bootstrap AUC distribution showed minimal variation, with a mean AUC of 1.0 and a 95% confidence interval of [1.0, 1.0], indicating stable performance within the current dataset. Given the limited sample size, these results should be interpreted as evidence of internal consistency rather than definitive generalizability.”

The corresponding revisions have been incorporated into the Discussion section (page 14, lines 387–394).

“Our study developed and internally evaluated a five-gene signature (PLPPR1, CNTN3, HOXA10, HAGLR, and ENPP3) for predicting immunotherapy response through integrative analysis of three independent GEO transcriptomic datasets. Within the combined cohort, the signature showed high discriminative ability and internally stable performance across resampling-based validation analyses. However, given the limited sample size and retrospective design, these findings should be interpreted with caution. Further validation in larger, independent, and prospectively collected clinical cohorts is required to establish robustness, generalizability, and clinical applicability.”

The corresponding revisions have been incorporated into the Conclusions section (page 14, lines 453–463).

“In conclusion, this study identified and internally evaluated a five-gene metabolic signature (PLPPR1, CNTN3, HOXA10, HAGLR, and ENPP3) associated with immunotherapy response in hepatocellular carcinoma through integrative analysis of multiple transcriptomic datasets. By combining meta-analysis with complementary machine learning approaches, we demonstrated internally consistent discrimination between responders and non-responders and provided biologically interpretable insights into metabolic and immune-related mechanisms. While these findings offer preliminary evidence supporting the potential utility of this signature for response stratification, external validation in larger, independent, and prospectively collected cohorts will be essential before clinical translation. Our study highlights the value of integrative multi-cohort and systems-level approaches for biomarker discovery in immuno-oncology.”

Comment 2:

HOXA10 and CNTN3 are discussed, PLPPR1, HAGLR, and ENPP3 receive limited mechanistic interpretation

Response 2:

We thank the reviewer for this important and insightful observation. This comment prompted us to examine the manuscript more carefully and highlighted that the original version placed greater emphasis on HOXA10 and CNTN3, whereas the mechanistic roles of PLPPR1, HAGLR, and ENPP3 were described in more limited detail.

To address this concern, we have substantially revised the Discussion to expand the biological interpretation of these three genes in the context of immunotherapy response. Specifically, PLPPR1 is now discussed as a regulator of lipid signaling and membrane dynamics, processes that are increasingly recognized as critical for immune receptor signaling, immune cell activation, and metabolic fitness within the tumor microenvironment. HAGLR (HOXD-AS1) has been elaborated as a long non-coding RNA involved in transcriptional regulation, epithelial–mesenchymal transition, and immune suppression, suggesting a role in shaping immunosuppressive tumor states that may limit immunotherapy efficacy. In addition, ENPP3 is now linked to extracellular nucleotide metabolism and purinergic signaling, particularly the balance between immunostimulatory ATP and immunosuppressive adenosine, a pathway known to impair antitumor immunity. Together with the expanded discussion of HOXA10 and CNTN3, these revisions provide a more comprehensive and mechanistically grounded framework linking all five genes to immunotherapy responsiveness. The corresponding changes have been incorporated into the revised Discussion section.

The corresponding revisions have been incorporated into the Discussion section (page 15, lines 404–423).

“Within this framework, HOXA10 emerged as the strongest negative predictor of immunotherapy response (coefficient: −326.820), consistent with its role as a master transcriptional regulator implicated in immune evasion and treatment resistance[31,32]. High HOXA10 levels may promote immunosuppressive microenvironments by regulating immune cell differentiation[33,34]. CNTN3, with positive coefficient (+147.927), encodes a cell adhesion molecule that may facilitate beneficial immune cell infiltration or enhance immune recognition mechanisms[35]. PLPPR1 encodes a membrane-associated phospholipid phosphatase–related protein involved in lipid signaling and membrane dynamics, processes increasingly recognized as critical for immune receptor signaling, immune cell activation, and metabolic fitness within the tumor microenvironment[36,37]. HAGLR (also known as HOXD-AS1), a long non-coding RNA, has been implicated in transcriptional regulation linked to tumor progression, epithelial–mesenchymal transition, and immune suppression, suggesting a role in shaping immunosuppressive tumor states that may limit immunotherapy efficacy[38,39]. ENPP3 has recently been identified as an extracellular cGAMP hydrolase that acts as an innate immune checkpoint by attenuating cGAMP–STING signaling[40]. Genetic or functional disruption of ENPP3 enhances antitumor immunity in a STING-dependent manner, implicating ENPP3 in immune regulation within the tumor microenvironment[40]. Collectively, these genes converge on metabolic regulation, transcriptional control, and immune modulation, key biological processes underlying immunotherapy responsiveness.”

Comment 3:

A few typographical errors (“re- sponse”, “ComBat- corrected”, etc.) appear due to line-break artifacts. A proofread will resolve formatting issues.

Response 3:
Thank you for pointing this out. We have carefully proofread the revised manuscript and corrected all typographical and formatting issues caused by line-break artifacts, including those affecting word hyphenation (e.g., “response” and “ComBat-corrected”).

Comment 4:

Check uniformity of gene name capitalization (e.g., HOXA10 vs HAGLR).

Response 4:

We thank the reviewer for this comment. We have carefully checked gene name capitalization throughout the entire manuscript and ensured consistent formatting according to standard human gene nomenclature conventions. All human gene symbols (e.g., HOXA10, HAGLR) are now presented in uppercase consistently across the main text, figures, tables, and supplementary materials.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

The manuscript titled “Biomarker-Based Precision Prediction of Immunotherapy Response in Hepatocellular Carcinoma” presents a strong and timely contribution to precision oncology. The integration of transcriptomic datasets, the application of robust machine learning approaches, and the identification of a biologically plausible five-gene metabolic signature provide clear scientific value. Before acceptance, a few minor revisions are recommended to further strengthen clarity and reproducibility: (1) provide a brief justification for choosing the three specific datasets and describe their clinical characteristics more clearly; (2) include additional details on hyperparameter tuning and model evaluation procedures, particularly for the LASSO and random forest models; (3) clarify whether AUC = 1.0 may indicate potential overfitting and discuss this briefly in the limitations section; (4) improve the description of the biological mechanisms linking each of the five genes with immunotherapy response; and (5) enhance the readability of figures by standardizing color schemes and improving axis labels. Addressing these minor points will refine the manuscript and enhance its overall impact.

Author Response

Comments:

The manuscript titled “Biomarker-Based Precision Prediction of Immunotherapy Response in Hepatocellular Carcinoma” presents a strong and timely contribution to precision oncology. The integration of transcriptomic datasets, the application of robust machine learning approaches, and the identification of a biologically plausible five-gene metabolic signature provide clear scientific value. Before acceptance, a few minor revisions are recommended to further strengthen clarity and reproducibility: (1) provide a brief justification for choosing the three specific datasets and describe their clinical characteristics more clearly; (2) include additional details on hyperparameter tuning and model evaluation procedures, particularly for the LASSO and random forest models; (3) clarify whether AUC = 1.0 may indicate potential overfitting and discuss this briefly in the limitations section; (4) improve the description of the biological mechanisms linking each of the five genes with immunotherapy response; and (5) enhance the readability of figures by standardizing color schemes and improving axis labels. Addressing these minor points will refine the manuscript and enhance its overall impact.

Response:

We sincerely thank the reviewer for the positive and constructive evaluation of our manuscript. We appreciate the thoughtful suggestions provided, which have helped us further improve the clarity, rigor, and reproducibility of the study.

Comment 1:

(1) provide a brief justification for choosing the three specific datasets and describe their clinical characteristics more clearly;

Response 1:

We thank the reviewer for this important comment. The three GEO datasets were selected to represent independent and clinically relevant cohorts of hepatocellular carcinoma (HCC) patients treated with immune checkpoint inhibitors, encompassing distinct therapeutic strategies and clinical contexts. As clarified in the revised Materials and Methods section, GSE279750 includes patients receiving first-line anti-PD-L1-based combination immunotherapy, with surgical tumor specimens collected after more than three months of treatment and classified according to modified RECIST criteria. GSE215011 comprises patients treated with nivolumab monotherapy (anti–PD-1), enabling direct comparison of tumor transcriptional profiles between responders and non-responders. GSE235863 represents HBV-positive HCC patients receiving anti–PD-1 plus lenvatinib combination therapy. In addition, the Results section has been revised to explicitly state that these cohorts reflect distinct but clinically meaningful immunotherapy settings, including monotherapy and combination regimens, thereby clarifying the clinical characteristics and rationale for integrating these datasets.

The corresponding revisions have been incorporated into the Materials and Methods section (page 2, lines 74–89).

“The three Gene Expression Omnibus (GEO) datasets[15] were selected to represent independent and clinically relevant cohorts of HCC patients treated with immune checkpoint inhibitors. GSE279750 (n = 10) includes patients receiving first-line anti–PD-L1–based combination immunotherapy, with surgical tumor specimens collected after more than three months of treatment and classified as responders or non-responders according to modified RECIST criteria[5]. GSE215011 (n = 10) comprises tumor RNA-seq data from patients treated with nivolumab monotherapy (anti–PD-1), enabling comparison of transcriptional profiles between responders and non-responders[16]. GSE235863 (n = 15) represents HBV-positive HCC patients receiving anti–PD-1 plus lenvatinib combination therapy (pembrolizumab or sintilimab) and includes paired samples collected before or after treatment initiation; responders were defined as patients achieving complete or partial response (CR/PR), whereas non-responders were defined as those with stable or progressive disease (SD/PD)[17]. All datasets were generated using next-generation sequencing–based transcriptomic profiling and provided clinical response annotations. For consistency, only tumor-derived gene expression data with available response information were included in the integrative analysis.”

The corresponding revisions have been incorporated into the Result section (page 4, lines 169–175).

“To evaluate gene expression patterns and identify predictive biomarkers associated with immunotherapy response, we integrated gene expression data from three independent GEO cohorts comprising 35 patients with HCC treated with immune checkpoint inhibitors (GSE279750, n = 10; GSE215011, n = 10; GSE235863, n = 15). These cohorts encompassed distinct but clinically relevant immunotherapy settings, including anti-PD-L1-based and anti-PD-1-based regimens, administered as monotherapy or in combination with lenvatinib, with tumor samples collected at different clinical time points.”

Comment 2:

(2) include additional details on hyperparameter tuning and model evaluation procedures, particularly for the LASSO and random forest models;

Response 2:

We thank the reviewer for this helpful suggestion. We have revised the Materials and Methods section to provide additional details on hyperparameter tuning and model evaluation procedures for both the LASSO and random forest analyses.

For LASSO logistic regression, feature selection was performed using the glmnet package with an L1 penalty (α = 1). The regularization parameter λ was selected by 10-fold cross-validation using the cv.glmnet function, and the value minimizing the cross-validated deviance (λ_min) was used to define the final LASSO model. Genes with non-zero coefficients were retained as LASSO-selected features.

Random forest analysis was implemented using the randomForest package primarily as an independent feature-ranking approach rather than as a final predictive model. Default parameters were applied (including 500 trees and mtry determined internally based on the number of predictors), and variable importance was assessed using the mean decrease in Gini index.

Model evaluation was conducted using a combination of predefined train/test splitting, receiver operating characteristic (ROC) analysis, and repeated stratified cross-validation. To minimize overfitting and information leakage, feature selection steps were completed before resampling, and only fixed gene signatures were evaluated during the validation process. These clarifications have now been incorporated into the revised Materials and Methods section.

The corresponding revisions have been incorporated into the Materials and Methods section (page 3, lines 120–135).

“2.5 Machine Learning-Based Feature Selection

LASSO logistic regression[20] was performed using the glmnet package (version 4.1-8) with an L1 penalty (α = 1). The regularization parameter λ was selected by 10-fold cross-validation using cv.glmnet, and the value minimizing cross-validated deviance (λ_min) was applied. Genes with non-zero coefficients were retained as candidate features. Random forest analysis was conducted using the randomForest package (version 4.7-1.1)[21] as an independent feature-ranking approach, with default parameters including 500 trees (ntree = 500) and internally optimized mtry values. Feature importance was quantified using the mean decrease in Gini index. LASSO-selected genes were further ranked by random forest importance, and genes with the highest combined scores were selected for signature development. The combined score was defined as the product of the absolute LASSO coefficient and the random forest importance (mean decrease in Gini). Model performance was evaluated using ROC analysis with predefined train/test splits and repeated stratified cross-validation. To minimize overfitting and information leakage, feature selection and model coefficients were fixed before validation, and only predefined gene signatures were evaluated during resampling.”

Comment 3:

(3) clarify whether AUC = 1.0 may indicate potential overfitting and discuss this briefly in the limitations section;

Response 3:

We thank the reviewer for this important comment. We agree that an AUC of 1.0 may indicate potential overfitting, particularly in studies with limited sample sizes. Accordingly, we have explicitly addressed this issue in the Limitations section of the revised manuscript. In addition, to avoid overstatement, we have revised the Results text to emphasize internal validation and model stability rather than definitive performance or generalizability. As discussed in the Study Limitations and Future Directions section, although repeated stratified cross-validation was used to reduce bias, independent external validation will be required to confirm the robustness of the five-gene signature.

The corresponding revisions have been incorporated into the Study Limitations and Future Directions section (page 15, lines 441–450).

“Several limitations should be acknowledged. First, the integrated cohort size was relatively small (n = 35), which may increase the risk of optimistic performance estimates despite internal validation. Although repeated stratified cross-validation with a fixed Signature Score was applied to mitigate overfitting and information leakage, the high AUC values observed should be interpreted as evidence of internal stability rather than generalizability. Therefore, external validation in larger, independent, and prospectively collected cohorts is required to confirm clinical utility. In addition, while bulk transcriptomic analysis enabled identification of a metabolically relevant signature, future single-cell or spatial transcriptomic studies may provide deeper mechanistic insights and enhance translational relevance for precision immunotherapy.”

Comment 4:

(4) improve the description of the biological mechanisms linking each of the five genes with immunotherapy response;

Response 4:

We thank the reviewer for this valuable suggestion. In response, we have revised the Discussion to provide a clearer and more comprehensive biological interpretation linking each of the five genes in the signature to immunotherapy response.

Specifically, we expanded the discussion of HOXA10 as a negative predictor of immunotherapy response, emphasizing its role as a transcriptional regulator associated with immune evasion and treatment resistance. The interpretation of CNTN3 was also strengthened by highlighting its function as a cell adhesion molecule that may enhance immune surveillance through improved immune cell–tumor interactions and immune cell infiltration.

In addition, we substantially expanded the mechanistic descriptions of PLPPR1, HAGLR (HOXD-AS1), and ENPP3. PLPPR1 is now discussed in the context of lipid signaling and membrane dynamics relevant to immune receptor signaling, immune cell activation, and metabolic fitness within the tumor microenvironment. HAGLR is described as a regulator of transcriptional programs linked to tumor progression, epithelial–mesenchymal transition, and immune suppression. ENPP3 is linked to extracellular nucleotide metabolism and purinergic signaling, particularly ATP–adenosine balance, a pathway known to modulate antitumor immunity.

Together, these revisions establish a more integrated biological framework connecting transcriptional control, metabolic regulation, and immune modulation to immunotherapy responsiveness. The corresponding changes have been incorporated into the revised Discussion section.

The corresponding revisions have been incorporated into the Discussion section (page 14, lines 404–450).

Comment 5:

(5) enhance the readability of figures by standardizing color schemes and improving axis labels. Addressing these minor points will refine the manuscript and enhance its overall impact.

Response 5:

We thank the reviewer for this helpful suggestion. In response, we have revised Figure 3A–C to improve visual clarity and readability. Specifically, the color schemes have been standardized across panels to ensure consistency and facilitate comparison, and the axis labels have been refined to improve clarity and interpretability.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

Reviewer’s comments

This original article written by Chao, et al. focuses on the biomarkers to predict the effects of immunotherapy in patients with HCC. This original article has its own originality and seems to be very interesting. However, the experimental protocols are not sufficient in several parts. The definitions of “responder” in the treatment has not been described at all. The several revisions will be required for the better quality of this original article. Please refer to the comments shown below.

Major

#1. Keywords were not appropriately selected. The authors should consider appropriate keywords to reflect the contents in this study.

#2. Which protocol was selected as the immunotherapy in patients with HCC? Was the immunotherapy naïve in the enrolled patients?

#3. The authors should describe the definition of “responder” in the immunotherapy? When was the judgement for the immunotherapy was done?

#4. In Figure 4C, only one core gene (LINC01554) was identified as the most common gene among three data base. However, this gene was not selected as the top five genes. The authors should describe the reason.

#5. In Figure 5B, how were the values of importance calculated?

#6. The statement on the combined score was found in page 10, lines 253-254. However, it was not easy to understand how to calculate the score.

Minor

#1. In Figure 3A, 3B and 3C, the text was too small to decipher.

#2. LASSO (page 2, line 58) should be spelled out.

#3. Hepatocellular carcinoma (page 1, line 36) and false discovery rate (page 2, line 136) can be abbreviated, respectively.

Author Response

Response to Reviewer 3 Comments

Comments:

Response:

We thank the reviewer for the careful evaluation of our manuscript and for the constructive comments provided. We appreciate the reviewer’s insightful suggestions, which have helped us substantially improve the clarity, methodological rigor, and overall quality of the manuscript.

Comments 1:

Major

#1. Keywords were not appropriately selected. The authors should consider appropriate keywords to reflect the contents in this study.

Response 1:

We thank the reviewer for this helpful comment. We have revised the keywords to more accurately reflect the core content and analytical framework of this study, including hepatocellular carcinoma, immunotherapy response prediction, metabolic gene signatures, biomarker discovery, and machine learning–based modeling. The updated keywords now better capture the study’s focus on integrative transcriptomic analysis and predictive biomarker development for immunotherapy response in HCC.

Keywords: hepatocellular carcinoma; immunotherapy response prediction; metabolic gene signature; biomarker discovery; machine learning

Comments 2:

#2. Which protocol was selected as the immunotherapy in patients with HCC? Was the immunotherapy naïve in the enrolled patients?

Response 2:

We thank the reviewer for raising this important question. According to the original GEO annotations and corresponding publications, different immunotherapy protocols were applied across the three cohorts. Specifically, GSE279750 consists of patients receiving first-line anti-PD-L1-based combination immunotherapy, with tumor samples obtained after more than three months of treatment; these cases were therefore immunotherapy-naïve prior to treatment initiation. GSE215011 includes patients treated with nivolumab monotherapy (anti–PD-1), while GSE235863 represents patients receiving anti-PD-1 plus lenvatinib combination therapy (pembrolizumab or sintilimab). For GSE215011 and GSE235863, the original GEO annotations and corresponding publications did not uniformly specify prior immunotherapy exposure; accordingly, these cohorts were analyzed based on documented clinical response status rather than assuming a fully immunotherapy-naïve baseline. These protocol details, sampling contexts, and related limitations have been explicitly clarified in the revised Materials and Methods section, and the Results section emphasizes that the integrated analysis includes tumor samples collected under distinct immunotherapy regimens and at different clinical time points, thereby avoiding the assumption of uniform baseline conditions.

The corresponding revisions have been incorporated into the Materials and Methods section (page 2, lines 74–89).

The corresponding revisions have been incorporated into the Result section (page 4, lines 169–175).

“To evaluate gene expression patterns and identify predictive biomarkers associated with immunotherapy response, we integrated gene expression data from three independent GEO cohorts comprising 35 patients with HCC treated with immune checkpoint inhibitors (GSE279750, n = 10; GSE215011, n = 10; GSE235863, n = 15). These cohorts encompassed distinct but clinically relevant immunotherapy settings, including anti–PD-L1–based and anti–PD-1–based regimens, administered as monotherapy or in combination with lenvatinib, with tumor samples collected at different clinical time points.”

Comments 3:

#3. The authors should describe the definition of “responder” in the immunotherapy? When was the judgement for the immunotherapy was done?

Response 3:

We thank the reviewer for this important question. The definition of “responder” and the timing of response assessment were based on the clinical criteria and annotations provided in the original GEO datasets and corresponding publications. In GSE279750, treatment response was evaluated using modified RECIST criteria after more than three months of anti–PD-L1–based combination immunotherapy, with complete or partial response defined as a responder and stable or progressive disease defined as a non-responder. In GSE215011, patients receiving nivolumab monotherapy were categorized as responders or non-responders according to documented clinical response assessments reported in the original study. In GSE235863, responders were defined as patients achieving complete or partial response (CR/PR), whereas non-responders were defined as those with stable or progressive disease (SD/PD), following anti–PD-1 plus lenvatinib combination therapy. These response definitions and assessment time points have now been explicitly clarified in the revised Materials and Methods section and are reflected in the Results section to ensure transparency and consistency across cohorts.

Comments 4:

Response 4:

We thank the reviewer for this insightful comment. The identification of LINC01554 as the only gene shared across all three datasets in Figure 4C reflects its high cross-dataset reproducibility at the differential expression level. However, the aim of the meta-analysis (Section 3.3) was to identify reproducible expression changes, whereas the subsequent machine learning–based feature selection (Section 3.4) focused on optimizing predictive performance for immunotherapy response. LASSO and random forest analyses were applied to the integrated expression matrix, in which candidate genes were evaluated based on their contribution to classification accuracy in a multivariate context rather than overlap frequency alone. Although LINC01554 showed consistent differential expression, it did not rank among the top features based on combined LASSO coefficients and random forest importance, indicating limited incremental predictive value when considered alongside other genes. Therefore, the final five-gene signature was selected based on predictive relevance and model performance rather than cross-dataset overlap alone. This distinction has been clarified in the revised Results section.

The corresponding revisions have been incorporated into the Result section (page 8, lines 239–256; page 10, lines 273–308).

“ 3.3. Meta-Analysis Reveals Consistent High-Confidence Core Genes Across Datasets

To identify genes with robust cross-dataset reproducibility, we employed MetaVolcano meta-analysis integrating statistical evidence from three independent cohorts. The MetaVolcano plot (Figure 4A) displays average log₂ fold-change versus cross-dataset significance, identifying 15 genes achieving statistical significance in at least two of three datasets (red dots). Cross-dataset log₂FC heatmaps (Figure 4B) of these 15 significant genes validated directional consistency across GSE215011, GSE235863, and GSE279750, with most genes maintaining uniform expression patterns. Venn diagram analysis (Figure 4C) revealed dataset-specific contributions across the three cohorts. GSE279750 identified the largest total number of DEGs (n=225), with 212 unique DEGs specific to this dataset. GSE235863 identified 162 DEGs, and GSE215011 identified 37 DEGs. Notably, only one core gene (LINC01554) achieved statistical significance across all three datasets, highlighting the importance of meta-analysis for identifying reproducible markers. The correlation heatmap (Figure 4D) revealed co-expression patterns among the 15 significant genes, with hierarchical clustering identifying distinct gene modules suggesting coordinated regulatory mechanisms. Collectively, these results establish a robust set of cross-validated candidate genes, providing a reliable foundation for subsequent machine learning-based feature selection and predictive model development.”

“3.4. Machine Learning-Based Feature Selection Identifies Five Core Predictive Genes

Importantly, genes identified by cross-dataset meta-analysis were used as candidates for feature selection; however, final model features were selected based on predictive contribution in multivariate machine learning models rather than overlap frequency alone. To identify the most informative features for predicting immunotherapy response, we employed two complementary machine learning approaches on the integrated expression matrix. LASSO regression with L1 penalty performs automatic feature selection by shrinking coefficients of less important variables toward zero (Figure 5A). Following cross-validation optimization, the coefficient plot displays the magnitude and direction of coefficients for 14 candidate genes selected based on predictive performance. Blue bars represent positive coefficients, indicating genes positively associated with immunotherapy response, while negative coefficients (PLPPR1, HAGLR, HOXA10) are associated with non-responders. Random forest analysis provides an independent assessment of feature importance through ensemble learning (Figure 5B). Green bars show the mean Gini decrease for each gene, with higher values indicating greater importance in classification decisions. ENPP3, PLPPR1, and CHI3L1 exhibited the highest random forest importance scores. To integrate insights from both algorithms, we calculated combined scores by multiplying the absolute LASSO coefficients by random forest importance (Figure 5C). Specifically, the combined score for each gene was defined as: Combined score = |βLASSO| × RF importance (mean decrease in Gini). Combined score ranking identified the top five genes: PLPPR1 (1.55), CNTN3 (1.51), HOXA10 (1.36), HAGLR (1.16), and ENPP3 (1.11), which were selected for further validation. Box plot analysis demonstrated significant expression differences between responders (blue) and non-responders (red) for all five genes (p < 0.05) (Figure 5D). To evaluate the predictive performance, we constructed a logistic regression model based on the five-gene signature and first assessed its discriminative ability using a predefined training and testing split. ROC analysis demonstrated strong discriminative performance in the initial training/testing evaluation (Figure S1A). Given the limited sample size, we further evaluated the stability of this result using repeated stratified cross-validation with a fixed Signature Score. Across resampling iterations, the model showed highly consistent discriminative performance, yielding consistently high AUC values (Figure S1B–C). Collectively, these results establish a robust five-gene predictive signature derived from dual-algorithm feature selection, supported by consistent performance across independent testing and repeated resampling, underscoring the internal robustness of the integrated machine learning framework for predicting immunotherapy response within the current cohort.”

Comments 5:

#5. In Figure 5B, how were the values of importance calculated?

Response 5:

We thank the reviewer for this important question. In Figure 5B, feature importance values were calculated based on the mean decrease in Gini index derived from random forest analysis. Specifically, random forest models were implemented using the randomForest package with 500 trees, and for each gene, importance was quantified as the average reduction in Gini impurity contributed by that feature across all trees in the forest, reflecting its overall contribution to classification decisions. As described in the revised Materials and Methods section (Section 2.5), the mean decrease in Gini index was used as an independent metric to rank feature importance in a multivariate, non-linear context. Genes with higher values indicate a greater role in improving node purity and discriminating between responders and non-responders. To further enhance clarity, we have also updated the Figure 5 legend to explicitly state that the importance values represent the mean decrease in Gini index across the forest. These revisions ensure transparency and consistency regarding the calculation and interpretation of feature importance in Figure 5B.

The corresponding revisions have been incorporated into the Materials and Methods section (page 3, lines 120–135).

“2.5 Machine Learning-Based Feature Selection

“Figure 5. Machine learning-based feature selection identifies five core predictive genes for immunotherapy response

(B) Random forest analysis provides an independent assessment of feature importance through ensemble learning. Green bars indicate the mean decrease in Gini index for each gene across the forest, with higher values reflecting greater importance in classification. ENPP3, PLPPR1, and CHI3L1 exhibited the highest importance scores.”

Comments 6:

#6. The statement on the combined score was found in page 10, lines 253-254. However, it was not easy to understand how to calculate the score.

Response 6:

We thank the reviewer for this comment. The combined score was used exclusively for feature ranking during the gene selection stage (Section 3.4) and is distinct from the multi-gene signature score described in Section 2.6. Specifically, the combined score was calculated by integrating two complementary measures of feature importance: the absolute value of the LASSO regression coefficient (|β|), reflecting linear predictive contribution, and the random forest importance quantified as the mean decrease in Gini index, reflecting non-linear classification relevance. For each candidate gene, the combined score was defined as: Combined score = |βLASSO| × RF importance. Genes were then ranked according to this combined score, and the top five genes were selected for downstream model development.

In contrast, the multi-gene signature score described in Section 2.6 was calculated using logistic regression coefficients applied to normalized gene expression values and was used solely for patient-level prediction and validation. We have clarified this distinction and the calculation of the combined score in the revised Materials and Methods and Results sections.

The corresponding revisions have been incorporated into the Materials and Methods section (page 3, lines 137–156).

“2.6. Signature Score Development and Validation

A logistic regression model was constructed using the glm() function with family=binomial(link="logit"). Maximum likelihood estimation optimized regression coefficients to maximize discriminative ability between responders and non-responders. The multi-gene signature score was calculated as the weighted sum of normalized expression values using regression coefficients (βᵢ) derived from the logistic regression model:

Signature Score = Σ βᵢ × Xᵢ

where βᵢ represents the regression coefficient for gene i, and Xᵢ denotes its normalized expression level. Model performance was evaluated through receiver operating characteristic (ROC) curve analysis using pROC (version 1.18.5)[22], calculating area under the curve (AUC) values for the signature score and individual genes. Bootstrap validation with 1,000 iterations assessed model stability and generated 95% confidence intervals[23]. To further mitigate potential bias from a single data split, additional internal validation was conducted using repeated stratified cross-validation (5-fold cross-validation repeated 100 times). Importantly, feature selection and coefficient estimation were fixed prior to cross-validation, and only the predefined Signature Score was evaluated during resampling, thereby minimizing the risk of information leakage. Differences in Signature Score distributions between responders and non-responders were assessed using the Wilcoxon rank-sum test.”

The corresponding revisions have been incorporated into the Result section (page 10, lines 289–295).

“To integrate insights from both algorithms, we calculated combined scores by multiplying the absolute LASSO coefficients by random forest importance (Figure 5C). Specifically, the combined score for each gene was defined as: Combined score = |βLASSO| × RF importance (mean decrease in Gini). Combined score ranking identified the top five genes: PLPPR1 (1.55), CNTN3 (1.51), HOXA10 (1.36), HAGLR (1.16), and ENPP3 (1.11), which were selected for further validation.”

Comments 7:

Minor

#1. In Figure 3A, 3B and 3C, the text was too small to decipher.

Response 7:

We thank the reviewer for pointing out this issue. We have revised Figures 3A-3C to improve readability by increasing the font size of all text elements, including axis labels, legends, and annotations. The updated figures have replaced the original versions in the revised manuscript.

Comments 8:

#2. LASSO (page 2, line 58) should be spelled out.

Response 8:

We thank the reviewer for this comment. As suggested, we have revised the text to spell out LASSO at its first occurrence. The sentence has been updated to read:

“Advanced computational approaches, particularly machine learning techniques combining Least Absolute Shrinkage and Selection Operator (LASSO) regression and random forest algorithms, offer powerful methodologies for identifying clinically relevant gene signatures from high-dimensional transcriptomic data [13,14].”

This revision has been made in the indicated section to improve clarity and readability.

Comments 9:

#3. Hepatocellular carcinoma (page 1, line 36) and false discovery rate (page 2, line 136) can be abbreviated, respectively.

Response 9:

We thank the reviewer for this helpful suggestion. As recommended, we have revised the manuscript to introduce appropriate abbreviations at the first occurrence of the terms. Specifically, hepatocellular carcinoma has been abbreviated as HCC after its first mention on page 1 (line 36), and false discovery rate has been abbreviated as FDR after its first mention on page 3 (line 107). The abbreviations are used consistently throughout the revised manuscript to improve clarity and readability.

Author Response File: Author Response.pdf

Round 2

Reviewer 3 Report

Comments and Suggestions for Authors

2^ndround

Major

#1. Well responded.

#2. Well responded. However, the authors should describe how many patients were regarded as responders.

#3. Well responded

#4. Well responded

#5. Well responded

#6. Well responded

Minor

#1. Well responded

#2. Well responded

#3. Well responded

Author Response

Response to Reviewer Comment

Comment:

#2. Well responded. However, the authors should describe how many patients were regarded as responders.

Response:

We thank the reviewer for this additional comment. As suggested, we have now explicitly clarified the number of responders and non-responders included in the integrated analysis. Across the three GEO cohorts, a total of 35 patients were analyzed, comprising 22 responders and 13 non-responders, as defined by the original clinical annotations and response criteria reported in the respective datasets. This information has been added to both the Materials and Methods and Results sections to improve transparency and ensure clarity regarding cohort composition.

The corresponding revisions have been incorporated into the Materials and Methods section (page 2, lines 89–91).

Materials and Methods

2.1. Data Acquisition and Preprocessing

“Across the three datasets, a total of 35 patients were included, comprising 22 responders and 13 non-responders based on the original clinical annotations.”

The corresponding revisions have been incorporated into the Results section (page 4, lines 175–177).

Results

3.1. Data Integration Reveals Consistent Gene Expression Patterns Across Cohorts

“Across the integrated cohort, 22 patients were classified as responders and 13 as non-responders based on the original clinical annotations.”