Skip to Content
MDPIMDPI

Advanced

by
  • Lu Qin 1,2,
  • Wen Zhu 3,* and
  • Yujing Zhang 1,2
  • et al.

Reviewer 1: Zhiping Zhang Reviewer 2: Anonymous Reviewer 3: Anonymous

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors
  • line 9: what is ‘GCN’?
  • line 11: what is ‘migration learning’?
  • The introduction should include related work in knowledge distillation.
  • line 126: based on figure 1 ROC curve, the statement ‘significantly surpasses all baseline models’ is far-fetched, given that SGCD, deepCDG, and M-GNN are very close.
  • line 175: what does ‘save’ mean here?
  • Figure 6 caption: why deepCDG?
  • line 436: it is unclear why needs knowledge distillation here, and also how it is incorporated in Figure 6.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

Dear Authors

Your manuscript presents an integrated framework for cancer driver gene prediction on a protein–protein interaction network by combining multi-omic signals (mutation, expression, methylation, and copy number variation) with a multi-branch GNN architecture. The design brings together modality-specific encoders, an attention-based fusion module, topology-driven feature enhancement, and a teacher-student knowledge transfer component. The overall direction is timely and the modular nature of the method is a clear strength.

That said, several parts of the paper would benefit from greater technical clarity and stronger experimental reporting. In particular, the concept you call “migration learning,” the exact construction and normalization of the omics-derived feature vectors, and the statistical framing of the comparative results need to be presented more explicitly. I would strongly encourage you to address the points below, as doing so will materially improve the transparency, reproducibility, and persuasiveness of the study.

Suggestions to improve the manuscript

  1. At present, “migration learning” is positioned as a central element, but it is not operationally defined in a way that the reader can follow and reproduce. Please state clearly whether this is transfer learning, domain adaptation, teacher–student distillation, multi-task learning, or another established learning paradigm. Then describe the training pipeline step by step. The description should include what is treated as source and target, what exactly is transferred, how the teacher is trained (or selected), and how the student is optimized. If the mechanism is essentially knowledge distillation, it may be better to use standard terminology or provide a strong justification for introducing a new label.
  2. You indicate that each gene is represented by a 64-dimensional vector derived from four omics modalities, with a 16-dimensional block per modality. However, the paper does not clearly explain what the 16 dimensions represent and how each modality-specific score is computed. Please provide concise definitions and, where possible, formulas for the mutation, differential expression, differential methylation, and CNV features. Also specify the tests, thresholds, and preprocessing steps used to derive “differential” signals. Keeping these details only in supplementary material makes it difficult to judge whether the reported gains stem from the architecture or from feature engineering choices.
  3. Z-score standardization is mentioned, but it is not clear whether it is performed per feature across all genes, per gene across features, or in another manner. More importantly, the standardization procedure must be described in a cross-validation-safe way. Please state explicitly that scaling parameters are fitted on the training folds only and then applied to the held-out fold. This is a small detail on paper, but it has a large impact on the credibility of reported performance.
  4. Using repeated 5-fold cross-validation is appropriate, but average scores alone are not enough to support strong superiority claims. Please report variability, such as standard deviation across repeats or confidence intervals. In addition, consider statistical tests comparing your method to the strongest baselines across the repeated folds. Even simple paired tests can help readers distinguish consistent improvements from fluctuations due to sampling.
  5. Constructing negatives by excluding known drivers is a common approach, but it can unintentionally label undiscovered true drivers as negatives. This risk should be acknowledged more directly because it can bias both learning and evaluation. If feasible, add a sensitivity analysis using an alternative negative sampling strategy or a stricter definition of negatives. Even a short experiment showing that conclusions are stable under reasonable choices would strengthen the work.
  6. The selected GNN-based baselines are relevant, but the current comparison may still leave readers wondering how the method stacks up against strong non-GNN approaches or other multi-omics plus network methods used in driver gene prediction. If adding additional baselines is not feasible, please justify the scope clearly and explain why the included baselines are the most appropriate references. Also describe how baseline hyperparameters were tuned to ensure a fair comparison.
  7. It appears that teacher information is used both as an auxiliary feature (teacher logits) and via a distillation loss. These are conceptually different contributions. I recommend ablations that separate them: one setting with teacher logits only, one with distillation loss only, and one with both. This will make your claims about the benefit of knowledge transfer more precise, and it will also help readers interpret what you mean by “migration learning.”
  8. AUC and AUPRC are standard and useful, especially under class imbalance. However, driver gene prediction is often used to prioritize a short list of candidates. Metrics such as precision@k, recall@k, hit@k, or NDCG would better reflect practical utility. Including these would align evaluation with how the outputs are likely to be consumed in downstream biological validation.
  9. The enrichment analysis would be more convincing if you clearly report the multiple testing correction method (for example, FDR control), the threshold used for significance, and the exact gene set definition (how “top genes” are chosen and how many are included). It would also help to show whether results are stable when k varies. For the drug sensitivity analysis, selecting genes “at random” weakens the interpretability. A more principled selection rule, such as top-ranked candidates or candidates absent from known databases, would make the narrative stronger.
  10. Some wording choices can confuse readers, especially if terms from other contexts are used to refer to omics modalities. Please keep the terminology consistent throughout. Also ensure that adjacency normalization is described consistently, and make clear which variant is used in implementation. These edits are minor in effort but important for clarity and reproducibility.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

This paper presents M-GNN, a Multi-modal Graph Neural Network framework designed to predict cancer driver genes by integrating multi-omics data (gene mutations, expression, DNA methylation, and copy number variations) with protein–protein interaction (PPI) network topology. The framework employs four independent GCN encoders per omics type, an attention-based cross-omics fusion layer, Node2Vec-derived structural embeddings, and a knowledge distillation strategy using a pre-trained teacher model to enrich feature representations. A dual-path classifier combining GNN and MLP heads further preserves both topological and intrinsic gene features. Evaluated across three PPI networks (STRINGdb, CPDB, IRefIndex), M-GNN consistently outperforms five baseline methods (EMOGI, HGDC, SGCD, deepCDG, IMVRL-GCN) in AUPRC and AUC metrics, with ablation studies confirming the contribution of each module. Biological interpretability is supported by GO/KEGG enrichment analysis, drug sensitivity correlation, and high overlap (91%) between predicted genes and curated cancer gene databases. Overall, the paper presents a methodologically solid contribution to computational oncology; however, it requires revision to address the following major and minor comments before it goes a further step.

Majors:

  1. In the Tables 1 and 2, they present point estimates (AUC, AUPRC) without confidence intervals or statistical significance tests (e.g., Wilcoxon signed-rank, McNemar). The performance margins between M-GNN and competing methods are often small (e.g., 0.8211 vs. 0.8014 on CPDB), making it impossible to determine if differences are statistically meaningful.
  2. The five baselines selected are not fully representative of the current state of the art. Notably, the TMO-Net [17] and Geneformer [18] are discussed in the introduction as relevant pre-training models but are excluded from comparative benchmarking without justification.
  3. The ‘deepCDG' in the Figure 6 caption was different from the ‘M-GNN’ mentioned in the method and the whole manuscript. This is a critical error that will cause significant confusion for readers and reviewers and suggests the figure may have been copied from a related source without full adaptation.
  4. In the Section 3.2.5, it describes using a 'pre-trained teacher model' for knowledge distillation, but critical details are missing: What architecture does the teacher model use? What training data and hyperparameters are used? How were the 'optimized configurations' determined?
  5. The paper acknowledges class imbalance (positive cancer driver genes are rare) and employs class-balanced cross-entropy loss, but does not report the actual class ratio in the dataset, nor does it evaluate sensitivity to the choice of negative sampling strategy. The author could report the number of positive and negative samples per PPI network. Include an experiment varying the negative-to-positive ratio or the sampling strategy to demonstrate model robustness to this design choice.
  6. In the Section 2.4, the GO/KEGG enrichment analysis is performed on the top-ranked predicted genes, not experimentally validated ones. The biological conclusions drawn (PI3K-Akt, focal adhesion) are well-established cancer pathways and may simply reflect the bias of the training data. Benchmark enrichment against background gene sets and provide enrichment p-values with multiple testing correction (e.g., Benjamini-Hochberg FDR). Discuss whether the identified pathways are enriched beyond what is expected from the training gene list alone.
  7. In the line 249, the 'Figures 4' for drug sensitivity results might match the following ‘Figure 5’. Additionally, the Figure 3 and Figure 4 could be integrated or put more detailed results into GO and KEGG figures separately.

Minors:

  1. The author could consider to move the figure 6 of overview of the project to the first figure position which helped the readers to understand the big picture. Then, more concepts and conclusions could be introduced to them by the overview logics.
  2. "multi-histology features" is non-standard terminology. Replace with 'multi-omics features' to align with established biomedical literature.
  3. In the lines 25 to 27, the sentence structure is complicated: 'Furthermore, cancer driver genes play a crucial role in the transformation of normal cells into cancerous cells. Therefore, accurately predicting…' Consider merging these two sentences for better flow.
  4. In the lines 64 to 65, the '…without accounting for variability across different omics' could add a citation to support this limitation claim, as it is a key motivation for the proposed method.
  5. In the lines 73 to 74, the sentence fragment: 'however, this approach does not consider the structural relationships of genes within the PPI network,' lacks a proper preceding clause and should be integrated into the previous sentence.
  6. Table caption refers to 'other baseline models' without listing them. Either list the models in the caption or add a footnote for self-contained readability.
  7. Table 2 is labeled 'Ablation study' without specifying which dataset or metric variant is used as the primary comparison condition. Add the full context, such as “STRINGdb, 10 × 5-fold CV”, to the caption.
  8. In the line 117, typo error: 'M-GNN and other baseline models inacross all networks' — remove 'in' before 'across.'
  9. Figure 1 caption is 'Comparison of ROC curves and PR curves of M-GNN with other baseline models.' Please specify the network (STRINGdb) since the figure results are network-specific.
  10. In the line 178, it’s better to distinguish the 'Table 2 in the Supplementary Material' from the Table 2 in the main text. Rename supplementary tables as 'Supplementary Table S1, S2' etc.
  11. The UpSet diagram in Figure 2 lacks axis labels and legend explanation adequate for standalone interpretation. Specify what 'Set Size' and 'Intersection Size' mean in this cancer gene context.
  12. In the line 341, the section heading typo: '3.2.4. Cross-group Interaction Fusion Layerr', extra 'r' in 'Layer.'
  13. In the lines 468 to 469, the conclusion states that 'topological features contributed to a performance improvement of more than 4%' but Table 2 shows 'Without structural features' yields AUPRC of 0.8135 vs. 0.8211 which the difference was less than 1%. Clarify or correct this claim.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

line 26-28 need a citation 

Author Response

Dear Reviewer,

Thank you very much for your insightful and constructive comments on our manuscript (metabolites-4193302). We have carefully considered all your suggestions and have revised the manuscript accordingly. All changes in the revised version are highlighted in blue to distinguish them from first‑round revisions (which were in red). Below we provide our point‑by‑point responses to your comments.

 

Comments and Suggestions for Authors

Comments1. line 26-28 need a citation 

Response 1: We thank the reviewer for pointing out that the statement in lines 26–28 lacks a citation. In the revised manuscript, we have added two supporting references [2,27] to this sentence to properly credit the established knowledge that cancer driver genes are critical for tumorigenesis and represent key targets for therapy. The added citation has been highlighted in blue to distinguish it from the first-round revisions (which were marked in red). We believe this addition strengthens the introduction and properly acknowledges prior work.

--- Page 1, Introduction, lines 27–30

 

We sincerely appreciate your time and expertise in reviewing our work. We hope that the revised manuscript now meets your expectations. Please do not hesitate to contact us if any further clarification is needed.

Sincerely,
Wen Zhu (on behalf of all authors)
School of Mathematics and Statistics, Hainan Normal University
Email: syzhuwen@163.com

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

I appreciate the work that went into this revision. The manuscript reads more coherently now, and several of the points that were previously difficult to evaluate have become substantially clearer. The explanation of the model components is more transparent, the terminology is more stable throughout the paper, and the added discussion around knowledge distillation makes the methodological contribution easier to understand. I also think the enrichment and drug-sensitivity sections are in better shape than before; they now feel less ad hoc and more analytically grounded.

Still, a few aspects of the revision feel incomplete to me.

  1. The first issue concerns how performance superiority is framed. I noticed that the revised manuscript now reports repeated cross-validation results with mean values and standard deviations, which is certainly an improvement. Even so, some of the wording still suggests a stronger and more universal advantage than the tables really show. In places where the margins are small, or where the model is not uniformly best across all settings, I would encourage the authors to be more measured. A brief statistical comparison against the strongest competing methods would help. If the authors prefer not to add formal tests, then the language of the claims should be adjusted accordingly. At this stage, what I would like to see is not necessarily more experimentation, but a tighter match between the evidence and the wording.
  2. I also remain unconvinced that the current discussion of ranking-oriented evaluation goes far enough. I understand the authors’ point that AUC and AUPRC are standard and informative, and I do not regard the absence of precision@k or NDCG as fatal. Still, for a task that is ultimately about prioritizing candidate driver genes, those measures are not peripheral. They speak directly to how useful the model would be in practice. Since these metrics have not been added in the revision, I think the manuscript should say more explicitly why they are absent and what that means for the interpretation of the reported results. Right now, this point is acknowledged, but it still feels slightly deferred rather than fully confronted.
  3. The revision on negative sampling is helpful, but I would encourage the authors to present it with a bit more caution. I appreciate that the manuscript now openly recognizes the risk that some unlabeled genes may in fact be true drivers, and the sensitivity analyses across random seeds and class ratios are worthwhile additions. Even so, these checks do not really eliminate the underlying labeling problem; they only suggest that the reported performance is not overly fragile under the chosen setup. I would therefore recommend softening any implication that the issue has been settled. A more explicit statement in the limitations or discussion section would make the paper more balanced and, in my view, more credible.
  4. I had a similar reaction to the comparison framework. The revised manuscript does a better job of explaining why the selected baselines were chosen, and I appreciate that the authors focused on conceptually relevant graph-based competitors rather than assembling an arbitrary benchmark. Even so, the comparison space remains relatively narrow. I do not think the paper needs a wholesale expansion of experiments at this stage, but I do think readers should be told more plainly that the study is benchmarking the proposed method primarily against GNN-based alternatives, not against the full landscape of possible driver-gene prediction strategies. That is a reasonable design choice, but it should be framed as such.

To be clear, I do think the paper is in a better state now. My remaining concerns are mostly about calibration, emphasis, and interpretive honesty rather than about the core idea of the work. The manuscript would benefit from one more careful revision aimed at making the claims slightly more restrained and the unresolved limitations slightly more explicit. In my experience, that kind of refinement often makes a paper more persuasive, not less.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

The authors have made substantial and thoughtful revisions that address the majority of reviewer concerns. The most significant improvements include: addition of rigorous statistical validation, comprehensive documentation of the knowledge distillation framework, explicit reporting of class imbalance ratios and sensitivity analysis for negative sampling, and proper application of Benjamini-Hochberg FDR correction. Overall, the revision demonstrates careful attention to methodological rigor and reproducibility. However, some major and minors were remaining, they were suggested to further revisions for better manuscripts’ publishing level standard.

 

Previous Major 6:

The authors were suggested to benchmark enrichment against background gene sets and discuss whether identified pathways are enriched beyond what is expected from the training gene list alone. But the authors did not explicitly compare their enrichment results against the training gene background. The concern about pathways like PI3K-Akt and focal adhesion simply reflecting training data bias remains partially unaddressed. The authors were suggested to include a control analysis showing enrichment p-values when using the training positive genes as the test set, demonstrating that predicted genes show additional pathway enrichment beyond known drivers.

 

Minors

  1. In the Line 7: 'Graph Convolutional Network (GCN)' - The acronym GCN is introduced here but not used consistently throughout. Consider using 'GCN' after first definition.
  2. In the Line 9: 'Knowledge distillation is incorporated' - This passive construction could be more direct: 'We incorporate knowledge distillation...'
  3. In the Line 96-100: The sentence introducing Figure 1's four modules is grammatically complex. Consider breaking into two sentences for clarity.
  4. In the Lines 136-140: The description of Z-score normalization could benefit from a brief explanation of why this prevents data leakage, as some readers may not immediately understand the concern.
  5. In the Line 226-229, the MLP subscripts (MLP1, MLP2, MLP3) suggest separate networks, but it's unclear if these share any parameters. Please clarify them.
  6. In the Line 270-271: 'hidden dimensions of 128 in the first classifier layer and 512 in the second'. This sentence could be clearer: specify whether these are sequential layers or parallel branches.
  7. In the Line 284: 'λ was set to 0.5 based on validation performance' - Consider mentioning what range of values was explored during tuning.
  8. In the Line 313 (Equation 15): The output layer bias is denoted b' (with prime), while Equation 19 uses b'' (double prime). This notation is inconsistent with standard practice; consider using different subscripts instead.
  9. In the Line 381-382: 'M-GNN consistently outperforms existing baseline methods' - This claim should be softened to 'consistently achieves the highest or second-highest performance' to match the actual Table 1 results.
  10. In the Figure 2: The legend font size appears small in the reproduced figure. Consider increasing for better readability in print.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf