We showcase the role of federated queries in reproducible science, by attempting to replicate an existing study to predict essential proteins through an integrative approach. More precisely, we compute orthologous scores of proteins using OMA, as well as the expression breadth of their orthologs using Bgee, which we federate via SPARQL, in order to rank candidate essential proteins across model organisms. We highlight challenges of this attempt, in particular the granularity of the data available in RDF, performance limitations, but also the current absence of a protein-protein interaction dataset in RDF.
reader to [
We provide the federated SPARQL query covering this use case in our github repository at https://github.com/dssib/swat23-data-in-use. Furthermore, in the accompanying Jupyter notebook in the repository, we also provide a step-by-step construction of the query.
The federated query we considered here is merely a starting point for reproducing the analysis in the initial study and is currently limited by the performance of servers involved (i.e., we can only currently run the query using limits). In the extended discussion in our Jupyter notebook, we also highlight the dificulty in writing federated queries with aggregations. However, compared to the methodology of the study, ours would have a few important advantages: the SPARQL queries are in theory fully reproducible, allowing any reader to directly obtain the datasets of interest from the most recent versions of the databases. Moreover, the intersection of these databases is performed implicitly, via the federation, as opposed to the manual work involved in merging datasets downloaded from disparate resources (e.g. InParanoid, Gene Expression Omnibus in the case of our considered study), an efort that would need to be repeated by every reader, given that the finalised dataset is not references in the paper.
Ranking candidate genes by expression level and expression breadth could be an interesting direction for future work. The availability of a Protein-Protein Interaction database in RDF, through a public SPARQL endpoint, would increase the potential to reproduce the results from the study by incorporating network topological features, as well as the co-expression within interaction networks. All in all, federated SPARQL queries remain an interesting, currently under-explored avenue for the future of reproducible research, in particular in the case of integrative analyses. We are working towards compiling a catalog of questions from scientific publications that involve multiple databases jointly, with their corresponding federated queries.