Address some of the reviewers' comments

bibliography.bib

@@ -67,6 +67,20 @@ @article{armstrong2021
   month         = sep,
   pages         = {2131–2137}
 }
+@article{aton2025,
+  title = {Scikit-bio: a fundamental Python library for biological omic data analysis},
+  volume = {23},
+  ISSN = {1548-7105},
+  url = {http://dx.doi.org/10.1038/s41592-025-02981-z},
+  DOI = {10.1038/s41592-025-02981-z},
+  number = {2},
+  journal = {Nature Methods},
+  publisher = {Springer Science and Business Media LLC},
+  author = {Aton,  Matthew and McDonald,  Daniel and Cañardo Alastuey,  Jorge and Azom,  Raeed and Batra,  Paarth and Bezshapkin,  Valentyn and Bolyen,  Evan and Cagle,  Alexander and Caporaso,  J. Gregory and Debelius,  Justine W. and Gorlick,  Kestrel and Hamsanipally,  Nirmitha and Hunger,  Lars and Keluskar,  Aryan and Liao,  Disen and Lu,  Yang Young and Navas-Molina,  Jose A. and Pitman,  Anders and Rideout,  Jai Ram and Sazonov,  Anton and Sathappan,  Bharath and Schwarzberg Lipson,  Karen and Sfiligoi,  Igor and Tapo,  Chris and Vázquez-Baeza,  Yoshiki and Wu,  Zijun and Xu,  Zhenjiang Zech and Ye,  Mingsong Sam and Zhao,  Jianshu and Knight,  Rob and Morton,  James T. and Zhu,  Qiyun},
+  year = {2025},
+  month = dec,
+  pages = {274–276}
+}
 @article{awais2025,
   title         = {Foundation models defining a new era in vision: a survey and outlook},
   author        = {Awais, Muhammad and Naseer, Muzammal and Khan, Salman and Anwer, Rao Muhammad and Cholakkal, Hisham and Shah, Mubarak and Yang, Ming-Hsuan and Khan, Fahad Shahbaz},

manuscript.qmd

@@ -332,13 +332,14 @@ extend the capacities far beyond the underlying technical framework [@katz2019].
 However, frameworks focused solely on microbiome data science fall short on 
 addressing the need to integrate data across diverse modalities, such as 
 metagenomics, metabolomics, and host genomics.
-
 The Bioconductor project embeds the methods of microbiome data science
 into a wider statistical programming framework
 alongside the concurrent development of interoperable methods and
 workflows for transcriptomics [@righelli2022], single-cell [@amezquita2020],
 metabolomics [@gatto2025], proteomics [@gatto2025; @anwar2025], and other
-fields. Here, we introduce a mature data science ecosystem supporting the 
+fields.
+
+Here, we introduce a mature data science ecosystem supporting the 
 integration and analysis of multi-modal datasets in microbiome research. 
 It provides harmonized data import and presentation, and fast, 
 robust methods for transforming abundance tables into biological insights. The
@@ -350,6 +351,33 @@ evidence-based best practices, we present the online book
 [https://microbiome.github.io/OMA](https://microbiome.github.io/OMA/docs/devel)
 that introduces many dedicated topics in microbiome analysis.
 
+# Data analytical frameworks {#sec-data_frameworks}
+
+Several data analytical frameworks are available for microbiome data science,
+each distinguished by its underlying paradigms, user communities, and scope.
+For example, tools such as QIIME 2 [@bolyen2019], Anvi'o
+[@eren2021], and Mothur [@schloss2009] emphasize standardized, workflow-oriented
+approaches tailored to microbiome analysis. These frameworks provide
+validated and user-friendly solutions, enabling development of efficient
+end-to-end workflows for microbiome research. However, they are typically
+designed around specific data types, which can limit flexibility
+when integrating heterogeneous data modalities within a unified analytical
+framework. In contrast, Bioconductor supports a more exploratory, statistical
+analysis oriented paradigm spanning multiple areas of bioinformatics beyond
+microbiome research.
+
+This data analytical paradigm of Bioconductor is similar to scikit-bio's
+[@aton2025]. While scikit-bio as a Python-based ecosystem benefits from rapidly
+evolving machine learning and artificial intelligence tools, Bioconductor’s
+strengths lie in robust statistical modeling and strong interoperability with
+other computational environments. Another key distinction is community
+structure: scikit-bio is relatively centralized, whereas Bioconductor packages
+are developed and maintained by a distributed community, resulting in a broad
+toolset. Although this can introduce challenges, such as dependency management
+and package compatibility, these are actively addressed through community
+standards, coordinated release cycles, and continuous integration practices,
+including shared data structures.
+
 # Data infrastructure {#sec-infra}
 
 Data scientists routinely deal with interlinked assays, hierarchical data
@@ -358,11 +386,13 @@ side information. This side information can include
 sample attributes like treatment group, collection site, or time point. The
 ability to integrate heterogeneous data elements is essential in biological
 research, where the individual components of a system can seldom be studied in
-isolation. However, input data formats vary across different methods, leading
-to technical complications and time lost on data wrangling.
-Bioconductor addresses this through standardized data structures supported
-by an ecosystem of interoperable methods, allowing the analyst to invest more
-time in core analytical tasks.
+isolation. However, this growing complexity introduces additional overhead, as
+researchers must track samples and features across multiple tables and manage an
+increasing number of data elements. Moreover, input data formats vary across
+different methods, leading to technical complications and time lost on data
+wrangling. Bioconductor addresses these challenges through standardized data
+structures supported by an ecosystem of interoperable methods, allowing the
+analyst to invest more time in core analytical tasks.
 
 Bioconductor data structures support statistical analysis of complex
 data combinations [@huber2015] and are widely adopted by various
@@ -517,7 +547,8 @@ sequences, and incorporates results from common analytical procedures such as
 dimensionality reduction, statistical transformations, and agglomeration by 
 taxonomic, functional or other information. The interoperability with other
 Bioconductor classes facilitates scalable analysis and the design of modular
-yet flexible workflows. In particular, the direct interoperability with the
+yet flexible workflows, as TreeSE inherits full compatibility with methods
+designed for SE. In particular, the direct interoperability with the
 widely adopted SE data science ecosystem distinguishes our TreeSE-based
 approach from phyloseq, which is another popular Bioconductor data container
 developed initially with a focus on 16S amplicon data analysis [@mcmurdie2012]. 
@@ -530,7 +561,9 @@ other profiles and alternative feature representations. It also offers
 substantial gains in memory efficiency and computing speed over phyloseq
 (@fig-benchmarking). While speedyseq is faster in certain operations, TreeSE
 tends to scale more efficiently in terms of execution time (t) and allocated
-memory (m) as sample size (n) increases. Despite these advantages, optimizing
+memory (m) as sample size (n) increases, especially as SE-based data containers
+also offer built-in support for sparse and delayed matrices, which can improve
+performance [@huber2015]. Despite these advantages, optimizing
 time and memory consumption presents an obvious area for further development
 in the TreeSE-based framework. Conversion functions and detailed cheat sheets
 between the two formats are available to support interoperability between the
@@ -811,18 +844,27 @@ taxonomic, resistome, transcriptome, proteome, and metabolome profiling data,
 dedicated data containers have been developed for certain modalities by the
 Bioconductor community to address their specific characteristics; examples
 include single cell sequencing [@amezquita2020] and host genomics
-[@lawrence2013]. Subsequent linking of the data from distinct modalities often
-presents an additional challenge: multi-omics data is often sparse, so that
-samples can be only partially matched between experiments. In other cases,
-a single sample in one modality may link to two or more samples in another
-modality [see @husso2023]. Integrating data across multiple modalities could
-thus benefit from flexible sample mappings and the ability to accommodate
-potentially different data containers corresponding to the different modalities.
-The MultiAssayExperiment (MAE) data container [@ramos2017] provides a solution
-to these two tasks and facilitates the integration of different omics
-modalities (@fig-datacontainers b). Integration of multi-table data without such
-a structured framework could lead to ambiguous mappings and inconsistent
-results. For instance, diagonal integration methods
+[@lawrence2013]. Linking data across modalities presents an additional technical
+challenge, as multi-omics data are often sparse and only partially matched
+across experiments. For example, one study may include 100 samples with
+metagenomics sequencing data but only 40 with matched metabolomics
+[@sherwani2025]. In other
+cases, a single sample in one modality may link to two or more samples in
+another modality (*e.g.*, due to technical or spatial replication)
+[see @husso2023]. These scenarios require flexible sample mappings and the
+ability to accommodate heterogeneous data containers across modalities.
+The MultiAssayExperiment (MAE) data container [@ramos2017] addresses these
+challenges by introducing a formal mapping layer (sampleMap) that links
+biological samples to their corresponding measurements in each assay
+(@fig-datacontainers b). This design supports one-to-one, one-to-many, and
+partially overlapping sample relationships across experiments. Importantly, the
+mapping is used during downstream operations: when subsetting an MAE object
+(*e.g.*, selecting samples with both microbiome and metabolome data), the
+framework automatically resolves the intersection of available samples and
+ensures consistent alignment across all assays. Such structured handling of
+sample relationships is particularly important for integrative methods, as it
+reduces the risk of sample mismatches, a common source of irreproducibility in
+ad hoc multi-table analyses. For instance, diagonal integration methods
 [@xu2022; @argelaguet2021computational]—where shared features across
 data types may not be explicitly defined—can benefit from
 such a systematic framework that enforces anchor alignment. This mirrors lessons
@@ -1213,8 +1255,12 @@ modalities, such as resistomes [@salehi2025], metabolomes [@hintikka2021],
 host genomes [@qin2022], transcriptomes, or proteomes
 [@nyholm2020; @moreno-indias2021] (@fig-downstream_analyses). As we have
 demonstrated, the integrative data containers support complex combinations of
-omics data sets [@huber2015; @ramos2017; @amezquita2020].
-Such integrative approaches in microbiome
+omics data sets [@huber2015; @ramos2017; @amezquita2020]. By leveraging
+standardized multi-assay data structures, users can apply a growing
+number of integrative methods for multi-omics analysis. This eliminates the
+need for manual data wrangling, reduces the risk of errors (*e.g.*, sample
+mismatches), and enables the creation of modular, efficient, and reproducible
+workflows. Such integrative approaches in microbiome
 research can be broadly categorized into three groups: Machine Learning (ML)
 prediction [@mallick2024], association-based approaches, and
 latent variable modeling  [@jeganathan2021; @sherwani2025;
@@ -1225,7 +1271,7 @@ analysis include cross-correlation, canonical correspondence analysis
 cross-correlation analysis can be used to quantify direct associations of
 samples or features between two data sets, anansi can supervise such analyses 
 by incorporating known functional relations, and CCA and Joint-RPCA 
-provide further tools to discover and visualize shared latent variation. 
+provide further tools to discover and visualize shared latent variation.
 We hope that systematic integrative frameworks, such as the one proposed here, 
 will accelerate the development, benchmarking, and standardization of 
  multi-modal extensions for microbiome analysis.
@@ -1469,8 +1515,17 @@ demands, the development and quality control of academic software
 critically relies on active and critical communities that can
 provide continuous benchmarking and support the identification and
 dissemination of evidence-based best practices [@berg2020;
-@barker2022; @hocquet2024]. We actively support community formation
-by online training resources and communication channels. Together with the
+@barker2022; @hocquet2024]. For example, starting with simple methods, early
+computation of key measurements, and initial visual exploration improves
+transparency, interpretability, and understanding of the data, while explicit
+code and clear commenting further support traceability and verification.
+
+Harmonized data structures and access to validated
+methods help shape such best practices by enabling transparent benchmarking and
+fostering community-driven consensus, while allowing users to flexibly select
+and evaluate methods appropriate for their specific data and research questions.
+Community formation is actively supported through
+online training resources and communication channels. Together with the
 interactive applications, these resources provide accessible entry points for
 all developers and users.
 
@@ -1487,17 +1542,19 @@ Bioconductor data structures [@huber2015; @ramos2017; @huang2021] have been
 adopted as a solution for multi-table data integration tasks in various
 domains such as single-cell sequencing
 ([OSCA book](https://bioconductor.org/books/release/OSCA/)) [@amezquita2020]
-and transciptomics ([OSTA book](https://lmweber.org/OSTA/)). These resources
+and transciptomics ([OSTA book](https://bioconductor.org/books/release/OSTA/)).
+These resources
 demonstrate how standard data containers can support statistical
 programming on interlinked multi-table data sets.
 
 **Interactive applications.** Graphical user interfaces provide
 accessible entry points to the framework. The iSEEtree and miaDash
 packages provide interactive analysis and visualization tools for
-microbiome data [@benedetti2025]. Their
-interface supports automatic generation of source code and replicable
-workflows as well as analysis templates for users without a strong
-programming expertise. Moreover, the tidy R paradigm, available
+microbiome data [@benedetti2025]. The former extends iSEE to provide support for
+TreeSE objects, while the latter builds on iSEEtree to enable analyses without
+requiring any programming. Their interface supports automatic generation of
+source code and replicable workflows as well as analysis templates for users
+without a strong programming expertise. Moreover, the tidy R paradigm, available
 through tidyomics, simplifies the syntax and can streamline the
 workflows [@hutchison2024]. These tools allow users to analyze and 
 visualize complex data combinations with ease.
@@ -1571,7 +1628,13 @@ In order to enjoy the benefits of automation while
 maintaining flexibility, data scientists routinely remix open-source
 software components into custom workflows. However, the lack of commonly adopted
 data representation standards often hinders the development, sharing,
-and application of modular, interoperable statistical workflows. 
+and application of modular, interoperable statistical workflows. Here we present
+a R/Bioconductor ecosystem for multimodal microbiome data science. While
+TreeSE and MAE are prior contributions [@huang2021; @ramos2017], we have built
+an ecosystem around them through further
+method and community development. We also provide a comprehensive online book
+supporting learning and reproducible analysis, resulting in an ecosystem ready
+for microbiome research.
 
 **The Bioconductor community** is widely recognized for
 developing open research software for life sciences. Its
@@ -1600,7 +1663,7 @@ omics along varying spatial and temporal dimensions. In concert with parallel
 developments in other omics, our framework specifically addresses the need to
 integrate and jointly analyze interlinked abundance tables and supporting
 hierarchical and tabular side information across different omics. Whereas our
-framework supports the basic data wrangling and microbiomedata science
+framework supports the basic data wrangling and microbiome data science
 operations, it also provides the structured input necessary for probabilistic
 inference and scalable computation. This is essential as the recent 
 developments in microbiome and multi-omics 
@@ -1625,7 +1688,9 @@ documented the open data science ecosystem for multi-modal microbiome research
 built on widely adopted Bioconductor data containers, software libraries,  
 data resources, tutorials, and a growing community of developers. 
 Further extensions could include methods for temporal, spatial, and multi-omic
-analyses. Finally, supporting interoperability between R and other
+analyses along with an end-to-end case study demonstrating how an example
+analysis can be conducted in practice from raw data to results.
+Finally, supporting interoperability between R and other
 programming languages could facilitate the application of other machine learning
 frameworks. As the field evolves, the developer communities will continue
 to integrate new tools and approaches to improve integrative analyses of