Review Reports - Assessing the Value of Data-Driven Frameworks for Personalized Medicine in Pituitary Tumours: A Critical Overview

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The submission paper summarizes the application of machine learning (ML) and knowledge extraction in pituitary neuroendocrine tumors (PitNETs). The authors reviewed 471 omics studies, and describe several clinically annotated and publicly accessible datasets that are suited for robust machine learning model validation. The following comments could be considered to improve the manuscript:

1) In Section 2.1, the review paper declares four categories were defined and marked in different color. It is suggested marking these colors in Tables 1-5 for readers to follow.

2) In Tables 1-5, is there possible to provide the web page URL of the data repository? The authors could provide the procedures to access these data resources.

3) In Section 4, besides introducing the major machine learning paradigms for PitNET studies, it is better to discuss the advantages and weaknesses of the popular ML methods by comparing the state-of-the-art literatures.

4) In Section 6, the machine learning method is not clear. Pleases describe the details of the machine learners.

5) Regarding Figure 1 results, it is believed that the testing sample number was too small, so that the confusion matrices and the ROC curves were too extreme. Is that possible to include more samples for machine learning?

Author Response

We sincerely thank the reviewers for their thorough evaluation and constructive comments. We have revised the manuscript accordingly. All modifications made in the revised version are highlighted in red, and each query has been addressed point by point in this response document.

Reviewer 1

We thank the reviewer for the careful evaluation of our manuscript and for the constructive suggestions. We have revised the manuscript accordingly, as detailed below.

1) In Section 2.1, the review paper declares four categories were defined and marked in different color. It is suggested marking these colors in Tables 1-5 for readers to follow.

We appreciate this comment. The original submission included color-coded squares indicating each dataset category (RED, ORANGE, GREEN, BLUE), but these markings were inadvertently lost during the manuscript upload process. We have now restored the visual coding system directly within Tables 1–5. In addition, we have reinforced the color meaning within the table captions and the methods to ensure readers can follow the categorization throughout the manuscript.

2) In Tables 1-5, is there possible to provide the web page URL of the data repository? The authors could provide the procedures to access these data resources.

We thank the reviewer for this suggestion. We have now added the URLs of all publicly accessible repositories in Tables 1–5. To improve usability, we have also included a brief description in Section 2.1 summarizing typical access procedures. Most datasets deposited in GEO, ArrayExpress, and EGA (USA and European repositories) are directly downloadable without restrictions. For datasets stored in Chinese genomic repositories (e.g., NGDC), user registration is required, and in some cases access permissions are granted after a formal request. These details are now clearly indicated in the revised manuscript to assist readers in navigating the available resources.

We agree with the reviewer that a critical comparison is important. We have expanded Section 4 by adding a new subsection (4.6) that discusses the strengths and disadvantages of the main ML paradigms used in PitNET studies, including all the methods explained on the previous sections. This section clarifies which methods are most appropriate under common PitNET data constraints. We believe this addition substantially enhances the methodological depth of the manuscript.

4) In Section 6, the machine learning method is not clear. Pleases describe the details of the machine learners.

We thank the reviewer for this insightful comment. We have substantially revised both the Methods section (Section 2.3) and Section 6 to include a detailed description of the machine learning workflow. The revised text now specifies:
(i) the preprocessing steps applied to the methylation data,
(ii) the differential methylation analysis used for feature selection,
(iii) the rationale for choosing the elastic net classifier,
(iv) the hyperparameter optimization strategy based on repeated stratified 10-fold cross-validation,
(v) the iterative training process performed until model convergence, and
These additions clarify the ML pipeline and justify the methodological choices made, ensuring that readers can fully understand and reproduce the modelling procedure. We believe the revised sections now provide the level of methodological transparency requested by the reviewer.

We fully acknowledge the reviewer’s concern regarding sample size. Unfortunately, increasing the number of cases for this analysis is not currently feasible. Although additional PitNET omics datasets exist, the vast majority do not include the clinically relevant information required for prognostic modelling (follow-up data is the most difficult obtain). Furthermore, the lack of standardized definitions of clinical behaviour (variables that are essential for constructing and validating our predictive models). Moreover, there is no consensus across published studies on how to annotate outcomes such as aggressiveness, recurrence, or resistance to therapy, resulting in clinical labels that are not comparable between datasets. This limitation is one of the key conclusions of our review: despite a growing number of omics studies in PitNETs, the lack of standardized, clinically annotated datasets significantly constrains the development and validation of robust machine learning models.

Reviewer 2 Report

Comments and Suggestions for Authors

Authors presented a review paper titled "Assessing the Value of Data-Driven Frameworks for Personalized Medicine in Pituitary Tumours: a Systematic Review". The manuscript is interesting and can be improved by considering the following suggestions.

1.Motivation of the study must be needed.

2. Flowchart or pictorial representation of existing data-driven frameworks for personalized medicine in Pituitary Tumours can give more insights to the readers about this systematic review.

3. Tabular representation of existing research work is fine, but it will be beneficial to the readers if the authors discuss about the methodology, key findings, and limitations of the existing research work.

4. Authors have discussed about the elastic net classification and mentioned that detailed description is found in the methodology section. Please mention the page number of the methodology section, as it is not included in the manuscript.

5. Provide more details about the elastic net classification and why elastic net is chosen for this analysis.

6. The manuscript can be thoroughly checked for any grammatical or typographical errors.

Author Response

1.Motivation of the study must be needed.

We thank the reviewer for this observation. To address it, we expanded the Introduction to clearly articulate the motivation of our study. We now emphasize the lack of standardized, clinically annotated omics datasets in Pituitary Neuroendocrine Tumours (PitNETs) and the absence of consolidated methodological guidance for applying data-driven frameworks in this field. This addition clarifies the clinical and research needs that justify our systematic review and highlights the relevance of evaluating existing data resources, machine learning strategies, and knowledge extraction approaches for advancing personalized medicine in PitNETs. The revised text can be found at the end of the first paragraph of the Introduction.

Flowchart or pictorial representation of existing data-driven frameworks for personalized medicine in Pituitary Tumours can give more insights to the readers about this systematic review.

We thank the reviewer for this helpful suggestion. In response to Comment 2, we have added a new flowchart (Figure 2) at the new section 2.2 to provide a visual summary of the data-driven framework described in the manuscript.

Tabular representation of existing research work is fine, but it will be beneficial to the readers if the authors discuss about the methodology, key findings, and limitations of the existing research work.

In the revised manuscript, we have added a concise synthesis following the tabulated datasets (section 5.6) summarizing the predominant methodological approaches, the key findings reported in the literature, and the major limitations identified across existing studies. This addition highlights the current strengths and weaknesses of PitNET omics research and contextualizes the need for standardized clinical annotation and robust validation frameworks.

Authors have discussed about the elastic net classification and mentioned that detailed description is found in the methodology section. Please mention the page number of the methodology section, as it is not included in the manuscript.

We thank the reviewer for pointing this out. We have now specified the exact location of the methodological details by referencing Section 2.3 directly in the text. As page numbers may vary during typesetting, we refer to the section number rather than a page number, which ensures readers can locate the relevant information regardless of the final layout. The manuscript has been updated accordingly.

Provide more details about the elastic net classification and why elastic net is chosen for this analysis.

We thank the reviewer for this suggestion. Additional details regarding the elastic net classifier have now been incorporated into both the Methodology (Section 2.2) and Section 6 of the revised manuscript. Specifically, we clarify the rationale for selecting elastic net, its implementation, and its advantages for high-dimensional omics data. Elastic net was chosen because it combines LASSO and ridge regularization, making it particularly suitable for datasets such as DNA methylation arrays, where the number of features vastly exceeds the number of samples and multicollinearity is common. The updated text also explains the hyperparameter optimization process, the use of repeated stratified 10-fold cross-validation, and the iterative model training performed until stability was achieved. These modifications provide the methodological transparency requested and justify our choice of elastic net for the illustrative example.

The manuscript can be thoroughly checked for any grammatical or typographical errors.

The manuscript has now been carefully revised for grammar, clarity, and typographical accuracy. This revision was performed collaboratively, and the author group includes native English speakers who reviewed the text to ensure linguistic precision.

Reviewer 3 Report

Comments and Suggestions for Authors

The authors presented a review devoted to data-driven approaches in studies of pituitary neuroendocrine tumors. The manuscript follows the PRISMA 2000 guidelines for systematic reviews, covering a wide spectrum of existing literature on the subject. As the main advantage, the study provides a comprehensive list of various existing omic datasets and information about their usage in previous studies. The text is well written, and the conclusions are mostly supported by the presented results. Overall, I believe the review will be especially useful for the machine-learning researchers aiming for applications in the specified research domain.

I have the following comments:

Abstract contains rather extensive details on the number of publications processed (an initial and final numbers seem to be enough). Instead, it would be nice to find more specific results about the reviewed ML applications. For example, which approaches are more popular and why? Or something like this.
Supplementary Figure S1 was absent in the manuscript and was only provided later. I recommend including it as a figure in Methods.
There are problems with color coding throughout the text. The color codes introduced in lines 160–173 are not visible anywhere in the text. No color codes can be found in all tables; the meaning of the black box in the ‘Data sharing’ columns is not clear.
The authors mention the problem of small samples in ML analyses. As this problem seems to be common in clinical studies, it would be nice to know how researchers overcome this difficulty? Which specific methods are usually used in such cases?
Figure 1 and the whole idea of presenting this example looks questionable. Applying an ML model to 16 data cases is an a priori bad idea. The 64 cases are also not many, but even in this case the results (Fig. 1a) show low specificity. What message were the authors tried to transfer with this figure? In the current form, the figure may rather discourage researchers from using ML. I believe whether the dataset should be changed to a more numerous for the example, or the example itself should be discarded. Probably, presenting more specific details on previously published ML studies could be an alternative.

Author Response

I have the following comments:

Abstract contains rather extensive details on the number of publications processed (an initial and final numbers seem to be enough). Instead, it would be nice to find more specific results about the reviewed ML applications. For example, which approaches are more popular and why? Or something like this.

We thank the reviewer for this valuable suggestion. We have revised the abstract to reduce excessive numerical detail and added a sentence summarizing the predominant machine learning strategies identified in the reviewed studies.

Supplementary Figure S1 was absent in the manuscript and was only provided later. I recommend including it as a figure in Methods.

We thank the reviewer for noticing this. The absence of Supplementary Figure S1 in the initial submission resulted from an editorial management oversight during the file upload process. The figure has now been properly included in the revised manuscript within the Methods section, as recommended (Figure 1).

There are problems with color coding throughout the text. The color codes introduced in lines 160–173 are not visible anywhere in the text. No color codes can be found in all tables; the meaning of the black box in the ‘Data sharing’ columns is not clear.

We are sorry about this. The color coding was correctly implemented in the original version but was lost during the editorial conversion of the manuscript files. We have now restored the intended color scheme in all tables, ensuring that each dataset category is consistently represented according to the legend provided in the Methods section. In addition, we clarified the meaning of the black box in the ‘Data sharing’ column, which now explicitly indicates whether the dataset is publicly accessible.

The authors mention the problem of small samples in ML analyses. As this problem seems to be common in clinical studies, it would be nice to know how researchers overcome this difficulty? Which specific methods are usually used in such cases?

This challenge is inherent to research on rare diseases such as PitNETs, where cohort sizes are inevitably limited. We have now made explicit in Section 4.6 that the few existing studies addressing this issue rely on strategies such as regularization-based models, dimensionality reduction, and resampling procedures to mitigate overfitting and enable more robust prediction in high-dimensional settings. These clarifications were added to help readers identify where this methodological discussion occurs. However, given the very small number of available studies, it would be premature to define any of these approaches as “usually used” in the field.

Figure 1 and the whole idea of presenting this example looks questionable. Applying an ML model to 16 data cases is an a priori bad idea. The 64 cases are also not many, but even in this case the results (Fig. 1a) show low specificity. What message were the authors tried to transfer with this figure? In the current form, the figure may rather discourage researchers from using ML. I believe whether the dataset should be changed to a more numerous for the example, or the example itself should be discarded. Probably, presenting more specific details on previously published ML studies could be an alternative.

We thank the reviewer for this thoughtful comment. We agree that using machine learning on very small cohorts poses clear methodological limitations. However, this example was intentionally included to illustrate a central message of our review: although PitNET omics datasets are increasingly available, the lack of clinically annotated and numerically sufficient cohorts critically constrains the development of robust predictive models. At present, the dataset used in Figure 1 is the largest by far publicly accessible PitNET methylation cohort with outcome-related clinical annotation, and no alternative dataset exists that would allow a more adequately powered demonstration.

To address the reviewer’s concern, we have revised the text accompanying Figure 1 (now Figure 3) to explicitly state that the purpose of the example is not to promote the specific model or its performance metrics, but rather to demonstrate the feasibility of the approach and highlight the current barriers that prevent meaningful ML deployment in this domain. We now emphasize that the results underscore the need for multicentric efforts, dataset harmonization, and standardized clinical follow-up—without which ML models cannot be expected to generalize. With this clarification, we believe the figure serves as a constructive illustration of the field’s unmet requirements rather than a discouraging result for people working on PitNETs field.

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

I have read the revised manuscript and authors' response to review comments. The revised manuscript now is satisfied.