A Long Journey into Reproducible Computational Neuroscience

Meropi Topalidou (1, 2, 3) , Arthur Leblois (4) , Thomas Boraud (1) , Nicolas P. Rougier (1, 3, 2)

1 IMN - Institut des Maladies Neurodégénératives [Bordeaux]
2 LaBRI - Laboratoire Bordelais de Recherche en Informatique
3 Mnemosyne - Mnemonic Synergy
4 UMR 8119 - Centre de neurophysique, physiologie, pathologie

Computational neuroscience, a relatively recent field, has gained fast ground and modelling is now widely 1 used to better understand individual and collective neuronal dynamics, and to propose new functions relying on 2 neural substrate. While the development of a model is initially tightly linked to the specific question asked by a given 3 group of researchers, further development of the model in different contexts is often possible and desired. When such 4 further development relies on new players, the continuity of the work requires proper validation, reproduction and 5 sharing of the models original equations and code. However, as they are published today, computational neuroscience 6 papers rarely include the sufficient material for the reproduction and sharing of the underlying model as a whole. 7 Here, we aim at showing the full extent of the problem, as well as state-of-the-art solutions, through the detailed story 8 of the reproduction of a computational modelling study by Guthrie et al. (2013) investigating the dynamics of basal 9 ganglia circuits and their function in multiple level action selection. In collaboration with the authors of the original 10 work, we first explain the difficulties encountered during the reproduction and validation of the initial model and 11 results. These difficulties led us to completely rewrite the model enforcing best software practices, relying on previous 12 attempts to provide a common framework for reproducible computational science and software sustainability. We 13 hereby detail these practices in the face of our practical example: the reproduction of the results from Guthrie et al. 14 (2013). In particular, these practices include: (i) a template for formal description of the model in a single table, 15 (ii) a public repository for shared software a proper version control, and (iii) an easy interface to run the underlying 16 code and reproduce figures. We finally propose new formats for communicating results allowing the replay of a code 17 while reading a computational study, in order to get a deeper understanding of the concepts being introduced. 18 Author Summary Computational sciences, such as bioinformatics, computational biology or computational neuro-19 science, are gaining fast ground in modern Life Sciences and new discoveries can be made thanks to computational 20 models or numerical simulations and analysis. The picture is not all bright however. The computational part in 21 computational sciences implies the use of computers, operating systems, tools, frameworks, libraries and data. This 22 leads to such a large number of combinations (taking into account the version for each components) that the chances 23 to have the exact same configuration as one of your colleague are nearly zero. This draws consequences in our 24 respective computational approaches in order to make sure models and simulations can be actually and faithfully 25 reproduced. If reproducibility is the hallmark of Science, computational sciences seems to be still in their infancy 26 in this domain, even though things have started to improve. This article highlights the extent of the problem based 27 on the reproduction of a model in computational neuroscience and show how to circumvent it using recommended 28 practices.

Mots clés: Neuroscience; Basal Ganglia; Reproducible Neuroscience; Python

Domaines: Neurosciences

Meropi Topalidou, Arthur Leblois, Thomas Boraud, Nicolas P. Rougier. A Long Journey into Reproducible Computational Neuroscience. 2014. ⟨hal-01109483⟩