Creating an executable paper is a journey through Open Science

Executable papers take transparency and openness in research communication one step further. In this comment, an early career researcher reports her experience of creating an executable paper as a journey through Open Science.

culminating in the recent retraction from the Lancet of the paper claiming adverse effects of hydroxychloroquine on Covid-19 patients 12 .While the present Covid-19 pandemics has magnified the issue of open and transparent research, the reproducibility crisis is a well known issue within the wider academic community 13,14 , and how this could be tackled by implementing Open Science and open research practices is a current matter of debate 14,15 .
As researchers, we aim at having a direct and positive impact on our society through our work.However, in order to inform policy, research has to be open, transparent, and reproducible.This does not stop at the data being available, but includes transparency of the methods employed to reach conclusions, as well as an efficient dissemination of the results 16 .
A way to support data availability and methods transparency is the "executable paper" format -i.e.formatting a publication as a dynamic piece of software that combines text, raw data, and the code used for the analysis, and that a reader can interact with.While the reader still has to trust that data have been collected properly, being able to directly reproduce the analysis is already more than many classical publications currently allow.As classical journals do not yet accept executable papers as submissions, the executable paper fulfills the role of an extended preprint that exists next to a peer reviewed publication in a journal.Recently, I have been exploring the executable paper format as part of my Wikimedia Open Science fellowship by creating one from scratch for my last work on pattern formation in salt deserts 17 .My executable publication's latest release can be found on GitHub 18 .In the following, I share my experience of creating this executable paper.I will introduce the tools I used (and why), touch on the challenges I encountered (and how I solved them), and outline how working on the executable paper benefited my research practice.
What is an executable paper?Before I start to dive into the details, I want to explain what an executable paper is (or at least what I think it should be).In principle the idea is to enable the reader to reproduce each and every step taken to arrive at a publication's conclusions, from the raw data to the polished plot.Therefore the paper should be linked to a repository containing the raw data and should support the necessary functionality to make the research process transparent and reproducible.Specifically, the executable paper has to display nicely formatted text, including references and links (just like a journal article), as well as figures, plots, and possibly also videos and interactive elements.Additionally, the analysis code used for creating plots from data should be displayed and interpreted by the executable paper, also allowing for user input to modify the analysis.Lastly, the executable paper should be composed of completely Open Source components and hosted under a free license to be easy to share and reuse.
As I started working on my executable paper, I quickly found out that the whole concept of executable papers was not as established as I thought.There was a short outburst of activity back in 2011, driven by Elsevier 19 , but other than that little turns up when searching for the term online.Other than a couple of templates 20 , very short guidelines 21 , or very detailed but not very broadly applicable how-to's 22 , there is currently not much to build on.
A variation of the executable paper is the "reproducible paper" 23 , a publication centered around a simulation code that can be run by the reader to generate data and perform their analysis, thus reproducing the whole research process down to the results.The main difference with the executable paper is that, while the executable paper relies on empirical data, the reproducible paper's code generates the data, thus putting a larger emphasis on the description of the code.
Along a similar line, the large crowd of Jupyter Notebook 24 users is driving the development and dissemination of executable documents containing scientific research.There is a growing gallery 25 of amazing examples of scientific publications and books written in Jupyter Notebooks, including a notebook that walks the reader through the process of detecting gravitational waves from LIGO data 26 .Therefore I decided to adopt the Jupyter Notebook as the basic tool for my executable paper.

The document
The Jupyter Notebook document is written in Markdown 27 a very easy-to-learn markup language that allows one to quickly format text in an easy-to-read and web-friendly way.Important for me as a physicist: Jupyter Notebooks are also able to render equations using LaTeX notation.Since I usually write my publications in LaTeX, this allows me to copy-paste most of the text and equations over to the executable format.Figure 1 shows a small excerpt of the math rendering in the executable paper.
Next to the text, the document is composed of code cells that contain and interpret code from a variety of languages.In Fig. 2 is an example of such a code cell, in which a data set is loaded and displayed as a table.Personally, I believe that being able to see the first few lines of a data set can be very useful, as it immediately gives an intuition of what information the data set contains, and what are the typical values.
The code displayed in this specific cell is Python 29 .I chose Python for my executable document because it has a large user base and a large variety of well-documented and well-maintained libraries for scientific computing.Another advantage of Python is that with a little bit of effort and by writing readable instead of elegant code, it can be read almost as a description of what the code is actually doing.This allows readers who are unfamiliar with the syntax to follow the steps of the analysis and validate them by running the code themselves.Jupyter Notebooks also support other languages 30 , such as R 31 , haskell 32 , or javascript 33 .In general, the format of the Jupyter Notebook makes the analysis more open and accessible, by allowing readers that are potentially unfamiliar with the programming language to run any code without locally installing any additional software (more on that later under "hosting").

Hosting
A fundamental part of producing an executable paper is making all material, from the actual paper to the data, available to readers Fig. 1 Screenshot of a text cell containing symbolic math notations in a Jupyter Notebook.Equations are typeset using LaTeX notation and mathjax 28 and can be referenced using hyperlinks.
by hosting it on an accessible platform.In the future, journals adopting this kind of format will take over all the tasks associated with hosting.So far, however, creating an executable paper presents two challenges in terms of hosting: first all the resources (code, images, and data) need to be accessible to the reader.Second, since the executable paper is not a static artefact, but rather an interactive and dynamic piece of software, it needs to be hosted such that the user can interact with the document, preferably without having to install additional software.
The first challenge is easily solved by using Git and GitHub.Putting the text and code under version control in a Git repository presents additional advantages: it allows to save the history of changes, which can be used to revert back to earlier versions of the document if needed.It also allows for collaborative working on text and code by managing changes from different collaborators.Uploading the code on a remote repository at GitHub is a minor additional step, and has the benefit of creating a backup in a remote server.Git also allows for an easy setup of a project description through the README file, and the creation of a license.Furthermore, it is easy and fast to create a release of the materials and add a persistent DOI, for example by using Zenodo 34 .Ultimately, publishing an executable paper on GitHub is equivalent to publishing a preprint on your own homepage or on a preprint server.
As regards the second challenge, most people would agree that getting a programming language to work locally on one's machine is far from being straightforward.Operating systems would interfere with development environments or compilers, and even producing a simple "hello world" program with a new language could easily take up to an hour.Fortunately, for interpreted programming languages (such as Python) online hosting solutions such as binder 35 are available.GitHub allows for easy integration of binder with a simple badge (just like a license badge).This way, readers can execute the code without the need to download any data or install any software.Large projects (such as my executable paper 18 ) can take a couple of minutes to load on binder, but that is a minor tradeoff when compared to the need to locally install a new programming language.So now the technical aspects are covered, I will say a few words about the challenges I encountered when creating the content of the executable paper.

Clean and shareable data
To fulfill the aim of enabling a reader to follow my analysis from the first raw data point to the polished plot, I need to make my raw data accessible.This means raw data has to be well-ordered, well-described, and uploaded somewhere publicly accessible.
Formatting data so that they are accessible is actually extremely time consuming, even when one would consider themself as a well-organized person.In particular, my publication draws on data from experiments, field studies, and simulations, all of which features different formats, sizes, and analysis pipelines that needed to be made consistent.While going through the process, I realized I had made countless implicit assumptions that guided the data processing.Therefore, a part of making my data and analysis open, included giving data sets more telling names, documenting meta-information, polishing, and annotating my analysis scripts.
Particularly relevant points turned out to be (i) using telling and unique file names that describe the data content, and possibly the dates when the data was collected or processed; (ii) documenting all relevant meta-information in a dedicated readme file any time data are modified; (iii) storing a copy of the original raw data; and (iv) having multiple and synchronized backups of the processed data.
Beside being fundamental in terms of Open Science practice, formatting data can bring further benefits.For example, I could publish my reformatted data as six independent datasets on PANGAEA [36][37][38][39][40][41] , and write a dedicated descriptive publication 42 .I definitely learned a lot throughout the process and I am applying what I learned to my new research projects -in the hope that next time it will not take me so much time to get my data in order.

Choosing the right level of detail
Making every single analysis step transparent and reproducible is desirable for the sake of openness.Nevertheless, as for every scientific publication, the level of detail needs to be carefully balanced with the narrative flow in order to ensure readability.To this end, the executable paper format gives more options than a regular publication: Jupyter Notebooks offer collapsing code cells and the possibility to add links to more detailed descriptions or separate analysis scripts.Analysis scripts and additional methods descriptions can also be presented in the supplement of a regular publication.Nevertheless, the strength of the executable paper is to seamlessly collect and combine all relevant information in one master document.This way, following the details necessary to reproduce the research does not require multiple switches between different documents and software, thus becoming more transparent.
In my specific case, I tried to strike a balance between brevity and reproducibility by merging a few important parts of the original appendix and methods into the main part of the article.In addition, I added many links that lead to more detailed explanations.For the more complicated data analysis I created a Fig. 2 Screenshot of a code cell in which a data table is read and displayed in a Jupyter Notebook.The reader can see which data file is read into the notebook and which columns it has.Executing the code cell reads the data table into a variable and displays the first five rows of the table below the code cell.
separate script that can be downloaded with the main repository and can be re-run to analyze the data from scratch.Since alongside raw data I also provide the already analyzed data, rerunning that script is a reader's choice and not a must.

Interactive elements
In my opinion, one of the major benefits of executable papers in terms of supporting the openness of the research process, is the possibility to extend the analysis beyond what is typically reported in a traditional publication.Usually, only a single (often "the successful") way of analyzing the data is described in a paper, and things like sensitivity analysis, or the results obtained with different parameters (for example a cutoff value) are reported in the supplementary information, or omitted altogether.Conversely, an executable paper allows for user input in the code cells, that is, plots can be interactively modified and re-computed, taking the reader input into account.If a certain choice was made for a parameter, the reader can modify the chosen value and check to what extent the conclusions drawn from the analysis still hold.
In my executable paper, I incorporated interactive elements as a sensitivity analysis within the key figure of the manuscript by allowing the user to try out different approaches to calculate the permeability of desert soil.I chose this specific parameter because it has a large impact on the results and because there are many different and ambiguous ways to calculate it.In this way, the figure shows the main result, and at the same time that the choices for the analysis only have a minor impact on the outcome (except for the very extreme cases).I believe such a solidity check not only strengthens the conclusions in the eyes of the reader, but is also important for an author to be more confident about the choices made during the analysis process.
A journey through Open Science I think that executable papers would greatly improve research standards, especially if they were to be accepted and reviewed by journals just like regular publications.In fact, the executable paper format enforces publishing data and analysis code in a clean and well-documented way alongside the article.Preparing such polished material takes time, but also gives an additional chance of catching mistakes, and leaves it ready-to-use and reuse for future research.Moreover, the additional transparency makes it hard to "hide" sloppiness in the analysis process or even fraudulent practices.
Sincejust like regular readersreviewers of executable papers would not need to install additional software, the technological barrier to reviewing executable papers is minimal.Actually, weaving together into a seamless environment all components of the research process has the potential to reduce review time: reviewers would not need to attempt reproducing the calculation themselves, as they could check the code straight away, and well explained and documented methods would probably prevent many misunderstandings.Of course, reviewers should have a minimum of knowledge about coding and data standards to be able to judge the quality of all aspects of the publication.
Similar to publishing preprints, publishing an executable paper ahead of a classical journal publication bears the risk of other scientists trying to "steal" ideas.But just like preprints, a DOI can be assigned to executable papers which can be used to prove the origin of the idea in case of a dispute.Similar to publishing a data set in a public repository, publishing data in an executable paper enables other people to use such data for their own research.While this might seem scary to some, I think that it has many benefits: publishing data broadens the context in which the research is useful to other researchers, resulting in more citations or even new possibilities for collaborations.
Creating an executable paper takes a significant amount of time, but I can only recommend the experience.It gave me new confidence about my own research, and forced me to revisit choices I had done in the past, sometimes in a rush and with less understanding than I have today.Creating an executable publication also gave me the chance to improve my research practice by using many open tools and workflows I already knew about, and pulling them together into a more open and transparent way of academic working.In retrospective, creating an executable paper is a journey through almost all aspects of Open Science: from open and accessible data to open and transparent tools and software, ending with an open publication composed of open and transparent content.