It is often the case that scientific knowledge exists but is scattered across numerous experimental studies. Because of this dispersion in different formats, it remains difficult to access, extract, reproduce, confirm or generalise. This is the case in crop science, where Mahmoud et al [1] propose to collect and assemble data from numerous field experiments on intercropping.

It happens that the construction of the global dataset requires a lot of time, attention and a well thought-out method, inspired by the literature on data science [2] and adapted to the specificities of crop science. This activity also leads to new possibilities that were not available in individual datasets, such as the detection of full factorial designs using graph theory tools developed on top of the global dataset.

The study by Mahmoud et al [1] has thus multiple dimensions:

• The description of the solutions given to this data assembly challenge.
• The illustration of the usefulness of such procedure in a case study of 37 field experiments on cereal-legume associations. The dataset is publicly available [3], while some results obtained from it have been independently published elsewhere [e.g. 4].
• The description of an algorithm able to detect complete factorial designs.
• An informed discussion of the merits of global datasets compared to alternatives, in particular meta-analyses
• A documented reflection on scientific practices in the era of big data, guided by the principles of open science.

I was particularly interested in the promotion of the FAIR principles, perhaps used a little too uncritically in my view, as an obvious solution to data sharing. On the one hand, I am admiring and grateful for the availability of these data, some of which have never been published, nor associated with published results. This approach is likely to unearth buried treasures. On the other hand, I can understand the reluctance of some data producers to commit to total, definitive sharing, facilitating automatic reading, without having thought about a certain reciprocity on the part of users and use by artificial intelligence. Reciprocity in terms of recognition, as is discussed by Mahmoud et al [1], but also in terms of contribution to the commons [5] or reading conditions for machine learning.
But this is another subject, to be dealt with in the years to come, and for which, perhaps, the contribution recommended here will be enlightening.

References

[1] Mahmoud R., Casadebaig P., Hilgert N., Gaudio N. A workflow for processing global datasets: application to intercropping. 2024. ⟨hal-04145269v2⟩ ver. 2 peer-reviewed and recommended by Peer Community in Mathematical and Computational Biology. https://hal.science/hal-04145269

[2] Wickham, H. 2014. Tidy data. Journal of Statistical Software 59(10) https://doi.org/10.18637/jss.v059.i10

[3] Gaudio, N., R. Mahmoud, L. Bedoussac, E. Justes, E.-P. Journet, et al. 2023. A global dataset gathering 37 field experiments involving cereal-legume intercrops and their corresponding sole crops. https://doi.org/10.5281/zenodo.8081577

[4] Mahmoud, R., Casadebaig, P., Hilgert, N. et al. Species choice and N fertilization influence yield gains through complementarity and selection effects in cereal-legume intercrops. Agron. Sustain. Dev. 42, 12 (2022). https://doi.org/10.1007/s13593-022-00754-y

[5] Bernault, C. « Licences réciproques » et droit d'auteur : l'économie collaborative au service des biens communs ?. Mélanges en l'honneur de François Collart Dutilleul, Dalloz, pp.91-102, 2017, 978-2-247-17057-9. https://shs.hal.science/halshs-01562241