PREreview of A multi-scale segmentation-free self-supervised AI model to characterize the heterogeneity of the brain tumor microenvironment

This review resulted from the graduate-level course "How to Read and Evaluate Scientific Papers and Preprints" from the University of São Paulo, which aimed to provide students with the opportunity to review scientific articles, develop critical and constructive discussions on the endless frontiers of knowledge, and understand the peer review process.

The manuscript proposes a new approach to investigate clinical images based on an AI segmentation-free model. Using the trained model, authors found patterns of cellular organization in brain tumors related to patient survival. The research question is clear and relevant considering the error-prone cell-segmentation model that is currently in use. The analysis was made in a considerable sample space (389 images from 185 patients), showing potential to validate cellular composition with prognosis. Data presented is clear in general. However, below we make some considerations regarding the experimental approach and presentation that could enhance manuscript contribution to the field and its impact. 

The work is promising, but may be perceived as speculative regarding the proposed biological context of the findings. A suggestion is to carry out in vitro assays to validate the hypotheses raised by the segmentation-free model, as suggested by the authors. Also, the work may benefit from experiments comparing the segmentation-free model and the cell-segmentation model (maybe even in the same image dataset), further supporting the argument raised by the authors to develop this tool.

MAJOR REVISIONS

1. Authors might benefit from a text review regarding formal language, text clarity, referencing relevant existing literature throughout the text and acronyms definition upon their first use.

2. To enhance clarity and comprehension, we suggest that the figure legends are revised in order to be self-explanatory, including number of biological replicates, statistical approach used, and the general significance of the data presented. Additionally, it would be helpful to number the figures and supplementary figures according to their citation order in the manuscript text.

3. There is some overlap between the Methods and Results sections. Providing additional details in the Methods section would further enhance clarity and ensure the reproducibility of the presented data, below we outline some key points for revision in this section:

   • It would be helpful to clarify how the image dataset was acquired and to cite the relevant database or paper that provided it.

   • The protein markers list could be referenced by papers that validated them.

   • The mention of the 'custom web app' at the end of the manuscript could benefit from further explanation. It would be useful to indicate whether this is a new tool developed by the lab or a pre-existing app customized for this specific work. Additionally, details about the parameters used for analysis within the app, its availability, and any modifications made (if applicable) should be provided.

   • It would be beneficial to cite the original works from which the sc and snRNAseq data were extracted, and to provide a clear description of the reanalysis protocol in the Methods section. If new code was generated, it is advisable that it is made available on platforms like Zenodo, and a statement in the data availability section should be included. Furthermore, more clarity regarding the correlation between the samples used for RNAseq and the conditions assessed in this work would be appreciated. It should also be explicitly stated that this is a reanalysis of the RNAseq data, rather than an original analysis.

   • To improve clarity, all parameters used in the AI training protocol could be detailed.

   • While the use of statistical approaches such as the Benjamini-Hochberg method can reduce false discovery rates, it would be beneficial to reference existing literature that demonstrates the appropriateness of the chosen statistical methods. Additionally, details about the data visualization and statistical analysis software used would help to clarify the methodology.

   • Some supplementary figures, such as Figures 1 and 6, which relate to model training and other aspects of the methods section, could be more appropriately cited within that section.

4. Even if this is a methods paper, the Discussion section could benefit from a broader scope, including not only a recap of the key results but also a comparison with findings from other methods available in the literature. Additionally, it would be helpful to highlight the significance of the data, address potential limitations, and consider future perspectives. For example, the visual comparison between LTME reconstruction+visibility and SCME suggests that the former may be less efficient. Could this be considered a limitation, or is there another explanation?

5. "We experimented with masking the area outside the tile boundary but ultimately decided to keep it as it improved  model performance, likely because our model was able to learn from the additional context.” It might be useful to present this data in supplementary figures to provide further clarity.

MINOR REVISIONS

1. Tables 1 and 2, as well as the first topic in the results section, could be moved to the methods and supplementary materials, since it does not contain any new data. Also, the relevance of information present in these tables, as treatments and stage, might be further discussed, as well as sample inclusion/exclusion criteria considering these information.

2. Figure 4 has so many color variables that makes it difficult to understand, maybe consider to split the information into separate graphs.

3. In the context of text and data clarity, the authors might benefit from merging figures into figure panels since many of them are closely connected, e.g. figures 1 to 4 and figures 5 and 6.

4. Biological significance of figure 5 findings regarding association between CD20 B cells/CD8 T cells and survival is very interesting and should be further discussed in the manuscript.

5. Figure 6 and supplementary figure 6 are difficult to understand. If a structure is present/absent, it should be clearly indicated by arrows and boxes. Also, the meaning of yellow boxes in these images is unclear.

6. Legend key is blocking data points in figure 7B and should be repositioned.

7. p-value = 0 appears in figures containing kaplan-meier plots. Was this intended or it was supposed to depict a threshold of sensitivity, such in p < 0.0001? A clarification to this matter would be opportune, such as in a statistical analysis subsection within the Methods section.

8. A suggestion is to deposit codes on a permanent record database, as Zenodo, instead of Colab notebooks and github, to improve availability and reproducibility by future studies.

Competing interests

The authors declare that they have no competing interests.