A collection of annotated and harmonized human breast cancer transcriptome datasets, including immunologic classification [version 2; peer review: 2 approved]

The increased application of high-throughput approaches in translational research has expanded the number of publicly available data repositories. Gathering additional valuable information contained in the datasets represents a crucial opportunity in the biomedical field. To facilitate and stimulate utilization of these datasets, we have recently developed an interactive data browsing and visualization web application, the Gene Expression Browser (GXB)


Introduction
Technological progress in the field of biomedical research has resulted in an increased utilization of platforms generating information on a system-scale, e.g. genome, transcriptome and proteome. As researchers are typically willing and often required to share their data collections, the availability of 'big data' is expanding rapidly. At this moment, the NCBI Gene Expression Omnibus (GEO), a public repository of transcriptome profiles, holds over 2 million individual transcriptome profiles from more than 76,000 studies ("Home -GEO -NCBI", 2016). This large amount of available transcriptomic data provides major opportunities as well as challenges to researchers. Identification of differential gene expression in healthy versus diseased individuals, for example, has the potential to increase our understanding of the disease process, can lead to the identification of novel disease biomarkers or to the recognition of potential therapeutic targets. However, utilization of the available system-scale information can be challenging, since data repositories often lack the analytical and visualization tools needed for data assessment and interpretation. For this reason, proper analysis relies on elevated bioinformatics skills.
To overcome the challenges faced when analyzing transcriptomic data, we previously developed a web application called gene expression browser (GXB), which makes datasets more accessible and interactive (Speake et al., 2015). The application graphically visualizes gene expression data in bar chart or box plot representation and is capable of dynamically changing its interface views upon user input. GXB allows users to upload microarray data, add data annotations, which enables overlay of clinical data, explore gene rank lists based on their differential expression patterns between groups, view the data on a gene-by-gene basis and compare different datasets and diseases. These capabilities stimulate the acquisition of new knowledge from public datasets, as demonstrated by the first paper that employed GXB to identify a previously unknown role of a specific transcript during immune-mediated processes (Rinchai et al., 2015).
In recent years, a large number of transcriptional studies have been conducted with the aim to characterize breast cancer on a genetic basis. GEO holds about 1297 datasets relating to breast cancer. One of the main impacts gene expression profiling has had on our understanding of breast cancer has been through the classification of breast cancer into intrinsic molecular subtypes (IMS). Three main methods have been described to achieve this, which have the same subtypes, but actually use different gene sets to stratify the patients (Hu et al., 2006;Parker et al., 2009;Sorlie et al., 2003). Four major IMS of breast cancer have been identified: Luminal A, Luminal B, HER2-enriched and Basal-like. A less common molecular subtype called Claudin-low has been characterized at a later time point (Prat et al., 2010). Stratified IMS groups present critical differences in incidence, survival and response to treatment, and most importantly add prognostic information that is not provided by classical stratifications, like estrogen receptor status, histologic grade, tumor size, and node status (Parker et al., 2009).
Recent breakthroughs in the field of cancer immunotherapy and especially the application of checkpoint blockade inhibitors has ignited a fierce drive to understand the genetic basis for the huge differences observed between patients with different immune phenotypes. Several papers have shown that expression profiles are able to distinguish between those patients that have an active immune environment and those that do not , a novel survival-based immune classification system was devised for breast cancer based on the relative expression of immune gene signatures that reflect different effector immune cell subpopulations, namely antibody-producing plasma B cells (the B/P metagene), cytotoxic T and/or NK cells (the T/NK metagene), and antigen-presenting myeloid/dendritic cells (the M/D metagene). The system defines a tumor's immune subclass based on its survival-associated immunogenic disposition status (IDS), which discriminates between poor immunogenic disposition (PID), weak immunogenic disposition (WID) and favorable immunogenic disposition (FID). The ability of IDS to distinguish patients with differential prognosis is dependent on the tumor's immune benefit status (IBS), which is defined by IMS and the expression of cell proliferation markers. The IBS classification segregates immune benefit-enabled (IBE) and immune benefit-disabled (IBD) tumors. In IBE tumors, but not IBD tumors, FID status confers a protective survival benefit compared to WID and PID status ( Figure 1B) (Miller et al., 2016;Nagalla et al., 2013). In this data note, we demonstrate the use of GXB to evaluate cancer gene expression across immunologic classifications of breast cancer.
Since the amount of possible datasets to be included in GXB is enormous, we chose to start with the GEO datasets underlying the

Amendments from Version 1
The manuscript was revised according to referees' suggestions. For clarity, names of datasets have been renamed to reflect the institution that has generated the data, both in the manuscript and the online GXB browser. The first example case of utilization of GXB has been adapted so the different strategies to explore the data are better represented. In addition, the existence of different probes for single transcripts is pointed out and explained in this example. As we recognize that we use quite a lot of abbreviations corresponding to the different classifications applied to the breast cancer datasets, we created an extra table (Table 2)  immunologically classified breast cancer datasets by (Miller et al., 2016). In Hendrickx et al. (2017), these same datasets were classified according to ICR. This will allow us to share our immune related classifications in a comprehensible way and allow others to reuse them. A harmonization effort of the other available clinical data had been undertaken and should help the downstream analysis of the expression data. Therefore, gathering these datasets with their detailed study and sample information will facilitate the identification of clinically-relevant genetic signatures for biomarker and/or therapeutic purposes.
In this data note, using GXB, we have made available a curated compendium of 13 public datasets relevant to human breast cancer, representing a total of 2142 cases.

Selection of breast cancer datasets
The starting point of our selection of breast cancer datasets are the patient cohorts included in the multi-study breast cancer database described by Nagalla et al. (2013). These 13 NCBI GEO datasets (GEO accession numbers: GSE45255, GSE2034, GSE5327, GSE12093, GSE9195, GSE11121, GSE1456, GSE2603, GSE6532, GSE7390, GSE7378 and GSE4922) resulted in 2142 cases initially uploaded in GXB. 22 of these cases reflect data from breast cancer cell lines and were therefore excluded from our data collection. A total of 1839 cases represent primary invasive breast tumors sampled at the time of surgical resection without prior neoadjuvant treatment and were therefore annotated with survival data, IMS, IBS, IDS and ICR status (Hendrickx et al., 2017;Miller et al., 2016). 281 of the cases did not fulfill these criteria and were therefore not annotated. Of note, 115 cases of original meta-cohort used Nagalla's study (n=1954) were not shared within GEO, but shared within other platforms (caArray and ArrayExpress). For this reason, these samples were not included in our GXB collection ( Figure 2).
The datasets that comprise our collection are listed in Table 1 and can be searched interactively in GXB. All GEO datasets consist of unique cases with the exception for 36 cases from NUH Singapore, which are both present in the Bordet Radcliff NUH (GSE45255) dataset and the Uppsala and Singapore (GSE4922) dataset.
Data of the 1839 GEO-cases annotated with survival data that were previously combined and used in the Nagalla study, have been uploaded to GXB in the dataset "Nagalla 2013 reconstituted public dataset".
Dataset upload into GXB All datasets were downloaded from NCBI GEO in SOFT file format and were uploaded into GXB with the exception of the Guy's hospital dataset (GUYT2; GSE9195). Expression data in the SOFT file of this dataset was expressed as fold change. Therefore, we had to revert to reprocessing of the CEL files found attached to the GSE on GEO. In this case, the cell files were read into R (v3.2.2) using the 'affy' package (v1.50.0). Data was normalized using the RMA (Robust multichip averaging) and gene annotation data was added using the hgu133plus2.db package (v3.2.3).
GSE records containing data generated with different or multiple platforms have been split by platform using the import process of GXB. GSEs containing data from both clinical as in vitro origin (GSE2603) have been split manually using the GXB Graphical interface.
Metadata of the different studies was added to GXB both from the descriptive information found on GEO or from the method Figure 2. Schematic representation of dataset selection and annotation. Breast cancer cases included in 13 NCBI GEO datasets were uploaded in GXB (n=2142). 22 cases described data from breast cell lines and were excluded from our data collection. We annotated 1839 cases with survival data, IMS, IBS, IDS and ICR status. 281 cases were either neoadjuvant treated, did not represent a primary invasive tumor, were not sampled at the time of surgery or without available survival data and were therefore not annotated. The total collection includes 1839 cases from the original cohort described in Nagalla et al. (2013) (n=1954). Of note, 115 cases of this cohort are not included in our collection as these were not shared via GEO. *251/1839 cases have been classified for IMS "Normal-like". IDS is not applicable for normal-like breast cancer tissue; therefore, IDS is non-classified for these samples. DMFS, Distant Metastasis Free Survival; GXB, Gene Expression Browser; IMS, intrinsic molecular subtype; IBS, immune benefit status; IDS, immune disposition status; ICR, immunologic constant of rejection. sections of the publications linked to these datasets. Short links to PMID (Pubmed) and GEO records were added.

Construction of the Nagalla's dataset
The constitution of the complete cohort has previously been described by (Nagalla et al., 2013). The dataset "Nagalla 2013 reconstituted public dataset" available in GXB contains only the samples that were publicly available via GEO. Briefly, raw data (CEL files) were extracted from GEO. The array platforms employed for these 13 datasets were Affymetrix U133A, U133A2, and U133 PLUS 2.  Table 2). IMS (i.e., Basal-like, HER2-enriched,

Classifications Categories Reference
Intrinsic To enable comparisons between datasets and to facilitate efficient data analysis, the clinical data was harmonized to reflect a nomenclature similar to that of The Cancer Genome Atlas (TCGA). Clinical variable names and availability in datasets are listed in Table 3. In general, variable values have been replaced by descriptive values (e.g. "1" and "0" are replaced by "ER+" and "ER-", respectively). For disease free survival, variable values have been adapted to "DiseaseFree" or "Recurred/Progressed", and for distant metastasis survival to "DistantMetastasisFree" and "Distant-Metastasis". Numeric values of variable "tumor size" have been converted to units in cm for all datasets. This variable was used to generate the additional variable pathology T stage according to the 7th edition of the AJCC staging system for breast cancer (Edge & Compton, 2010). For tumors with a diameter larger than 5 cm, pathology T stage could be either T3 or T4, therefore value "T3/T4" has been assigned to these cases.
Standardized clinical datasets can be found in the 'downloads' tab in GXB under the heading "additional files". All datasets start with the following 21 clinical variables in fixed order: "sample.ID", "array sample id", "sample title", "series", "IMS", "IBS", "IDS", "ICR", "DMFS_10Y_EVENT", "DMFS_10Y_TIME", "disease free survival event", "disease free survival years", "distant met free survival event", "distant met free survival", "age at initial pathologic diagnosis", "lymph node status", "ER status", "PR status", "histology differentiation grade", "tumor size cm", and "pathology T stage". In case one of these variables is not available in a specific dataset, values in this column are all NA.
Group sets for IBS/IDS, ICR cluster, Lymph Node (LN) Status, IMS, Histological grade, stage and Estrogen Receptor (ER) status were defined with matching differential gene expression rank lists. Rank lists are based on differential gene expression between two relevant groups for each group set: IBD-FID vs IBE-FID (IBS/IDS); ICR1 vs ICR4 (ICR1/ICR4); LN+ vs LN-(LN status); G1 vs G3 (histological grade); ER+ vs ER-(ER status). For IMS, no rank list was generated, as this variable is not ordered. For tumor stage, no rank list was generated because the spread of samples between categories was small (R scripts for this harmonistaion have been made available on Github.). For some of the transcripts, as is the case for the HLA-G transcript, multiple probes are available. Probe ID can be displayed by clicking "Tools" in the central panel and clicking "Show Probe ID" in the dropdown list. In this example, which is for illustrative purposes only, we selected probe ID 211528_X_AT at random.

Dataset demonstration
2) Sample grouping is default as "All sample", it is changed by selecting "Immunologic Constant of Rejection" and 3) plot type is set to "Box Plot" in drop down menus in the central panel. The central panel now presents a graphical display of the observed abundance of HLA-G transcripts in breast cancer samples across the different ICR groups, each sample is represented by a single point in a boxplot ( Figure 3A). A tendency of increased HLA-G expression in groups with the highest expression of ICR genes can be observed. 4) To verify whether this trend can also be observed in other breast cancer datasets, GXB's "Cross Project View" is used. By selecting "Cross Project View" in the "Tools" drop-down menu located in the top right corner of the user interface, a list of available datasets/projects appears in the left pane. By consecutive selection of single datasets, box plots with HLA-G transcripts across ICR groups are displayed for each individual dataset.
Each of the boxplots corresponding to the 13 datasets show a similar pattern, indicating an increased HLA-G expression in breast tumors with a high expression of ICR genes (representative plots are shown in Figure 3B). In the combined dataset containing the total of 1839 annotated cases from these datasets, this trend is also observed ( Figure 3B). This observation made by exploring transcriptome data in GXB provides an interesting starting point for further analysis. Statistical analysis of this potential association is required and, of course, the clinical relevance of the observed difference in abundance of transcripts should be determined. Most importantly, the functional relevance of HLA-G expression depends on its interaction with inhibitory receptors including ILT2, ILT4 and KIR2DL4 (LeMaoult et al., 2005). Therefore, combined analysis of both HLA-G and these inhibitory receptors is suggested in future analyses.
This example illustrates the convenience of exploring gene expression data in GXB. The browser facilitates intuitive navigation and visualization of gene expression across different group sets.

Differential gene expression between IBS/IDS subgroups
The breast cancer datasets uploaded in GXB are provided with a rich context of immune classifications and clinical parameters. As opposed to start a search with a specific gene of interest, as presented in the HLA-G example case, differential gene expression between groups of interest can be explored in GXB by evaluation of gene rank lists. Here, we demonstrate the use of GXB to explore differential gene expression across IBS/IDS groups.
The   This demonstration indicates that GXB allows for easy and efficient visualization of differential gene expression between subgroups. Subsequently, elaborate statistical analysis is required to confirm the differences in gene expression observed in GXB.

Overview of breast cancer immune classifications in GXB
Since this GXB data collection is provided with multiple immune classifications of breast cancer, it is interesting to visualize the relationship between these classifications in GXB. The overlay feature in GXB can be used to visualize the assignment of different classifications to individual samples simultaneously.
To illustrate this overlay option, we choose to select the Erasmus Medical Center dataset 2 (EMC2) with CXCL9 expression, as this is one of the chemokines included in the ICR gene signature. Graphical representation in GXB is set to bar plot and group set ICR is selected. As anticipated, the CXCL9 expression gradually increases from ICR1-ICR4. The drop down menu "Overlays" is used to add multiple layers of additional variables, "IBS", "IDS" and "IMS". Boxes underneath the individual bars (each bar represents a single case) display the assigned classifications ( Figure 5A). When comparing IBS classifications across ICR groups, it is evident that IBE tumors are frequently assigned  to the higher ICR clusters, ICR3 and ICR4, while IBD tumors tend to concentrate to the clusters with a low expression of the ICR signature (ICR1, ICR2) ( Figure 5A). This result is consistent with our previous observations: pathways that distinguish IBE and IBD are associated with the immune functional orientation of the tumor, and genes in these same pathways are crucial and PID tumors prevail in the ICR1 cluster ( Figure 5A). This observation is also in line with our expectations, the IDS classification is based on an immune metagene model that relies on immune gene subclusters that reflect the relative abundance of tumor-infiltrating immune cells (Nagalla et al., 2013). As markers of immune cell infiltration are also included in the ICR signature, IDS is closely associated with ICR.
For a more comprehensive overview of the relationship between different immune classifications in breast cancer, the overlay of immune classifications was evaluated in the Nagalla 2013 reconstituted public dataset (n=1839). The observations made in the EMC2 dataset (n=58; Figure 5A) are also apparent in the dataset containing all annotated cases of this GXB breast cancer instance ( Figure 5B). Moreover, in this dataset it is clearly visible that IBS/IDS subgroups with an improved prognosis are more prevalent in the ICR4 cluster. For example, IBE-FID tumors are relatively more frequently assigned to ICR4 compared with IBD-FID. Vice versa, IBD-PID tumors are proportionally more frequently observed in the ICR1 cluster compared with IBE-PID tumors, which are in comparison more frequently assigned to ICR2 ICR3.
The overlay of the different immune classifications demonstrates a coherency between the IBS/IDS classification and the ICR clusters. Bearing in mind that the ICR signature is associated with a broader phenomenon of immune-mediated, tissue-specific destruction, this coherency strengthens the hypothesis of a common final pathway of tissue destruction.

Conclusions
In this data note, we highlighted the opportunities provided by the availability of public datasets. We uploaded 13 public datasets on human breast cancer, including a combined dataset, with harmonized clinical data annotation and immune classification to GXB to facilitate the reuse of gene expression data. The use of GXB to explore gene expression and the different possible approaches were illustrated by the following: (1) an example case of a specific gene of interest, HLA-G; (2) comparison of gene expression between specific subgroups, IBD-FID vs IBE-FID; and (3) the evaluation of the relationship between different categorical variables, IBS/IDS and ICR immune classifications. To conclude, GXB provides a convenient environment to explore gene expression profiles in the context of breast cancer.

Data availability
All datasets included in our curated collection are available publicly via the NCBI GEO website: http://www.ncbi.nlm.nih. gov/geo/, and are referenced throughout the manuscript by their GEO accession numbers (e.g. GSE7390). Signal files and sample description files can also be downloaded from the GXB tool under the "downloads" tab.

Competing interests
No competing interests were disclosed.

Christos Hatzis
Breast Medical Oncology, Yale School of Medicine, New Haven, CT, USA Roelands et al present a web application that provides real-time analysis and visualization of a large compendium of microarray datasets generated from primary breast cancer biopsies. The uniqueness and distinct utility of this dataset is the focus on immune-related signatures. Before discussing potential shortcomings according to this reviewer's opinion, the authors should be commended for investing the energy and resources to develop and implement a focused public repository that would facilitate further analyses.
Although the overall objective behind this website and tools is laudable, I believe that the approach followed, as presented in the manuscript, has several potential shortcomings that would be worth addressing.
Regarding the protocols followed, I would echo the comments by the first reviewer and question the use of MAS5 and COMBAT, given that more robust algorithms are now available. This may be a key issue given the quantitative nature of many of the indices used. Also, it is not clear that this approach removes bias between the older Affymetrix platforms and the newer ones (PLUS2.0). On page 5 it is mentioned that "the 22,268 probe sets present in each of these platforms were included", but the PLUS 2.0 chips include twice as many. They probably meant to say "present in all platforms".

1.
In terms of whether sufficient details were provided to allow replication, it would be great if the authors provided the complete set of scripts used to pre-process and normalize the data included in the analysis. This will allow transparent assessment of the methods and replication of the dataset if needed.

2.
Related to the type of data included, I believe that the resource would be more valuable if additional clinically-relevant information was provided with each sample. For example, disease progression clearly depends on treatment and treatment information is missing entirely -we do not know whether patients even received any treatment at all. It would be useful to curate treatment information available in the original datasets. Also, it may be worth providing annotations for additional immune signatures (e.g. Rody et al, 2011 1 ) and also of additional breast cancer subtypes (e.g. Lehmann et al, 2016 2 ). This will allow a more thorough assessment of the overlap between the various signature-defined subtypes.
Finally, although there are not as many profiles from breast cancer metastasis available in GEO, it may be worth considering curating those and including them in the platform. It would be extremely interesting to know how the immune signatures differ between primary 3. and metastatic samples, and whether the trends suggest immune evasion.
Finally, from a usability standpoint, I have a few suggestions outlined below: -Although the functionality may be available, I found it very difficult to filter samples based on a particular feature. For example, I was trying to do the HLA-G by ICR group analysis only for the ER-negative or basal cases, but could not find an easy way to do that.
-It would be useful to provide tools for correlational analysis of two or more genes (e.g. coexpression patterns) and heat maps. The one gene at a time visualization is very restrictive.

4.
Minor points: Worth considering having another table after Table 1 that summarizes the immune subtype signatures and annotations with references. Intrinsic subtype signatures used can also be summarized in that table. That will also eliminate the need to explain the acronyms (e.g. Table 2 is marred by too many acronyms IMS, IBC, IDS, ICR that are not explained).

1.
When running the HLA-G example as outlined in the manuscript, I was presented with 4 different HLA-G transcripts with no explanation as to what was the difference between them. Are these different probe sets? If so, that should probably be explained. Also, in that case, is there a recommendation as to which of these probe sets is the most representative (e.g. based on specificity, range, variance etc).

2.
expertise to confirm that it is of an acceptable scientific standard, however I have significant reservations, as outlined above.
Author Response 21 Jan 2018 Wouter Hendrickx, Sidra Medical and Research Center, Qatar Major points: Regarding the protocols followed, I would echo the comments by the first reviewer and question the use of MAS5 and COMBAT, given that more robust algorithms are now available. This may be a key issue given the quantitative nature of many of the indices used. Also, it is not clear that this approach removes bias between the older Affymetrix platforms and the newer ones (PLUS2.0). On page 5 it is mentioned that "the 22,268 probe sets present in each of these platforms were included", but the PLUS 2.0 chips include twice as many. They probably meant to say "present in all platforms".

The primary goal of this data note was to share the transcriptomic and annotation data associated with this previous publication (in GXB defined as "the Nagalla 2013 reconstituted public dataset") and the breast cancer transcriptomic datasets from GEO that it constitutes of in a more interactive, comprehensible format to facilitate usage of the data and did not involve data processing. The selection of 22,268 probe sets, combination and normalization of the 13 datasets was also previously performed (Nagalla et al). For your second point, we indeed meant the probe sets that are present in all platforms, for this reasons probe sets that are exclusively present in Affymetrix U133 PLUS2.0 are not included in these 22,268 probe sets. We have revised this sentence in the manuscript accordingly.
In terms of whether sufficient details were provided to allow replication, it would be great if the authors provided the complete set of scripts used to pre-process and normalize the data included in the analysis. This will allow transparent assessment of the methods and replication of the dataset if needed.

The uploads from individual GEO datasets into GXB did not involve any data processing. For the Nagalla et al this has been performed as part of the work described in Nagalla et al. This data note did involve clinical data harmonization performed in R, scripts have been made available on Github.
Related to the type of data included, I believe that the resource would be more valuable if additional clinically-relevant information was provided with each sample. For example, disease progression clearly depends on treatment and treatment information is missing entirely -we do not know whether patients even received any treatment at all. It would be useful to curate treatment information available in the original datasets. Also, it may be worth providing annotations for additional immune signatures (e.g. Rody et al, 2011 1 ) and also of additional breast cancer subtypes (e.g. Lehmann et al, 2016 2 ). This will allow a more thorough assessment of the overlap between the various signature-defined subtypes. Finally, although there are not as many profiles from breast cancer metastasis available in GEO, it may be worth considering curating those and including them in the platform. It would be extremely interesting to know how the immune signatures differ between primary and metastatic samples, and whether the trends suggest immune evasion.

When uploading these GEO datasets into GXB, all available clinical data on GEO was also transferred to GXB. For some of the datasets (e.g. GSE4922), a treatment parameter is available which can be selected as a parameter using the overlay function in the bar plot generated in GXB. We completely agree that information on adjuvant treatment is essential for interpretation of the supplied survival data (Distant Metastasis Free Survival and Disease Free Survival). For proper survival analysis, we recommend users to download the csv file supplied which can be found under the "Downloads"-tab of each dataset and use available treatment information for stratification. Addition of more gene signatures can indeed be valuable to compare immune based classifications as demonstrated in this data note for Immunological Constant of Rejection (ICR) classification and Immune Benefit Status (IBS)/ Immune Disposition Status (IDS). Similarly, uploading additional GEO datasets that contain matched primary and metastatic samples would be very interesting and allow for comparison of both immune microenvironments. Although not in the scope of the current data note, we will definitely take these suggestions into consideration for further development of the GXB breast cancer instance.
Finally, from a usability standpoint, I have a few suggestions outlined below: ○ -Although the functionality may be available, I found it very difficult to filter samples based on a particular feature. For example, I was trying to do the HLA-G by ICR group analysis only for the ER-negative or basal cases, but could not find an easy way to do that.

It is indeed possible to visualize the expression of a gene of interest, for example HLA-G, in a specific subgroup of samples like all ER-negative or basal cases. The most convenient way to do this is to (1) select barplot, (2) set group by Molecular Subtype and (3) overlay the barplots with ICR group annotation. To make the figure easier to analyze, you can subsequently sort by ICR.
-It would be useful to provide tools for correlational analysis of two or more genes (e.g. coexpression patterns) and heat maps. The one gene at a time visualization is very restrictive.

We realize that gene per gene visualization has its limitations and analysis of groups coordinately expressed genes provides more biological significance. For this reason, the GXB development team has already created a module analysis tool (MAT) that analyses expression data to find pre-defined groupings of co-expressed genes (modules) to obtain a molecular fingerprint of gene expression for each individual sample in your dataset. MAT has already been applied to many datasets with samples of various immune related diseases and we are considering to also use this platform for cancer transcriptomic datasets. As the desired number of group comparisons in breast cancer (i.e., Molecular Subtypes, Immunologic classifications, Pathology T Stage) is higher compared to the existing group comparisons, some work is required before we can implement these dataset in this tool.
Minor points: Worth considering having another table after Table 1 that summarizes the immune subtype signatures and annotations with references. Intrinsic subtype signatures used can also be summarized in that table. That will also eliminate the need to explain the acronyms (e.g. Table 2 is marred by too many acronyms IMS, IBC, IDS, ICR ○ that are not explained).

We have added the suggested table to the manuscript. This is both useful when reading the manuscript as well as for the use of this breast cancer GXB application, thank you for this suggestion.
When running the HLA-G example as outlined in the manuscript, I was presented with 4 different HLA-G transcripts with no explanation as to what was the difference between them. Are these different probe sets? If so, that should probably be explained. Also, in that case, is there a recommendation as to which of these probe sets is the most representative (e.g. based on specificity, range, variance etc). I think GXB is a welcome contribution to the breast cancer research community and its public availability will help other scientists to leverage this valuable collection of datasets. However, I have several concerns that need to be addressed before indexing, as listed below.
The manuscript is well written and easy to understand.

Major comments
In Table 1 and on the front page of the web-application, Princess Margaret Cancer Centre is listed in the sample set name even though these data have not been generated in this institution, which is misleading.
When selecting "Breast Cancer", which I assumed contained the whole database, many datasets got filtered out, which is confusing.
I tried AURKA and selected GSE9195 as dataset. Then I played a bit with the barplot to overlay different kinds of information. I added DMFS 10Y (categorical) and sorted the patients based on this but there were some patients with very low DMFS at the beginning and the end of the plot, which is confusing.
Since the authors are dealing with survival data, I was expecting survival statistics and/or survival curves but I did not see any in GXB or the manuscript. Figure 1 is somehow misleading as these plots do not seem to be part of the web-application.
The authors decided to limit their database to Affymetrix data. In that case, I suggest the authors reprocess all the CEL files consistently using (f)RMA and up-to-date chip description files (CDF), such as BrainArray CDFs. This would increase the consistency across all the datasets.
Given the authors' large collection of datasets, mining each dataset separately is cumbersome. Implementing meta-analysis pipelines would help draw an overall conclusion before digging into dataset-specific results (e.g., Figure 3). This is what the authors seem to have done with the "Nagalla" dataset that is a compendium of all the datasets reprocessed using MAS5 and further corrected with ComBat. It is not clear why the authors used MAS% to normalize the data in this case vs (f)RMA for some other datasets. Please clarify. Why not recommending users to first explore Nagallia before going after each dataset separately to detect trends that may be (in)consistent with the majority of the datasets?
There have been many molecular subtyping schemes published, including the Integrative Subtypes (IntClust; Curtis Table 2 would be better represented as a barplot or a similar figure.

Minor comments
GXB would benefit from the inclusion of normal samples (from healthy patients or adjacent normal samples). METABRIC, TCGA and GTEx are relevant data sources for such gene expression profiles.

Is the rationale for creating the dataset(s) clearly described? Yes
Are the protocols appropriate and is the work technically sound?

Are sufficient details of methods and materials provided to allow replication by others? Partly
Are the datasets clearly presented in a useable and accessible format? Yes

Major comments
In Table 1 and on the front page of the web-application, Princess Margaret Cancer Centre is listed in the sample set name even though these data have not been generated in this institution, which is misleading. When selecting "Breast Cancer", which I assumed contained the whole database, many datasets got filtered out, which is confusing.

Thank you for noticing this. Only a single disease type can be assigned per dataset. We can change ER+ and LN-datasets to disease type: Breast Cancer , effectively disabling the user to filter datasets based on ER+ and LN status. We preferred changing the name of Breast cancer (general) to "mixed breast cancer types" for clarity.
I tried AURKA and selected GSE9195 as dataset. Then I played a bit with the barplot to overlay different kinds of information. I added DMFS 10Y (categorical) and sorted the patients based on this but there were some patients with very low DMFS at the beginning and the end of the plot, which is confusing.

We were able to reproduce the barplot you described. By sorting by the categorical variable DMFS 10Y EVENT, you only sort based on this variable (DistantMetastasisFree or DistantMetastasis). The continuous counterpart of this variable (DMFS 10Y TIME) is not taken into account when you perform the sorting, explaining your observation. An alternative is to sort based on the continuous variable DMFS 10Y TIME, though this sorting results in a plot that mixes DistantMetastasisFree and DistantMetastasis categories. Unfortunately, subsequent sorting based on 2 variables is not possible using GXB. We have suggested this improvement to the GXB developing team.
Since the authors are dealing with survival data, I was expecting survival statistics and/or survival curves but I did not see any in GXB or the manuscript. Figure 1 is somehow misleading as these plots do not seem to be part of the web-application.

Survival statistics are not yet part of the GXB web-application. To prevent any misconception, we have added an extra sentence in the legend of figure 1: "This figure is for explanatory purposes only and does not serve as a demonstration of the GXB web application."
The authors decided to limit their database to Affymetrix data. In that case, I suggest the authors reprocess all the CEL files consistently using (f)RMA and up-to-date chip description files (CDF), such as BrainArray CDFs. This would increase the consistency across all the datasets.

The purpose of this data note is to share existing data from GEO in a more comprehensible format and does not involve any data processing. We had to make an exception for dataset GSE9195.GPL570, since only raw format was available on GEO.
Given the authors' large collection of datasets, mining each dataset separately is cumbersome. Implementing meta-analysis pipelines would help draw an overall conclusion before digging into dataset-specific results (e.g., Figure 3). This is what the authors seem to have done with the "Nagalla" dataset that is a compendium of all the datasets reprocessed using MAS5 and further corrected with ComBat. It is not clear why the authors used MAS5 to normalize the data in this case vs (f)RMA for some other datasets. Please clarify.

Normalization of the complete Nagalla cohort was previously performed as described in Nagalla et al. 2013 and Miller et al. 2016 To make this data available as published in this article, we used the same data file to upload to GXB. Normalization is different between datasets because the data was transferred straight from GEO and different contributors use different methods depending on the times it was performed and the need of their specific work.
Why not recommending users to first explore Nagallia before going after each dataset separately to detect trends that may be (in)consistent with the majority of the datasets?

In this dataset collection, it is indeed possible to first explore the Nagalla dataset, which is a dataset that includes all samples from the collection. This advantage is specific for this breast cancer compendium of datasets, so your suggestion is a very good strategy to explore these datasets. We have added this recommendation in the text of the Dataset Demonstration section of the data note.
There have been many molecular subtyping schemes published, including the Integrative Subtypes  Table 2 would be better represented as a barplot or a similar figure. For the purpose of listing which clinical variables are available in how many datasets from the collection, we found a table the more appropriate option.

Minor comments
GXB would benefit from the inclusion of normal samples (from healthy patients or adjacent normal samples). METABRIC, TCGA and GTEx are relevant data sources for such gene expression profiles.

Comparing healthy control tissue with cancerous tissue on gene expression level indeed represents an interesting aspect that can be explored using the GXB browser. At this moment, different types of dataset collections have already been uploaded to GXB, including case versus control comparisons. For example, a dataset with gene expression of whole blood samples from both lung cancer patients and healthy controls and a dataset with head and neck cancer and cervical cancer tissue samples, matched with site-matched normal epithelial samples are currently available. The TCGA, METABRIC dataset and datasets in the GTEx portal are valuable resources that can be used to upload new datasets into GXB. We will definitely take this suggestion into consideration for our new dataset collection to upload into GXB.
Discussed changes will be applied in version 2 of the manuscript.
Competing Interests: none