Without appropriate metadata, data-sharing mandates are pointless

Article by Mark A. Musen: “Last month, the US government announced that research articles and most underlying data generated with federal funds should be made publicly available without cost, a policy to be implemented by the end of 2025. That’s atop other important moves. The European Union’s programme for science funding, Horizon Europe, already mandates that almost all data be FAIR (that is, findable, accessible, interoperable and reusable). The motivation behind such data-sharing policies is to make data more accessible so others can use them to both verify results and conduct further analyses.

But just getting those data sets online will not bring anticipated benefits: few data sets will really be FAIR, because most will be unfindable. What’s needed are policies and infrastructure to organize metadata.

Imagine having to search for publications on some topic — say, methods for carbon reclamation — but you could use only the article titles (no keywords, abstracts or search terms). That’s essentially the situation for finding data sets. If I wanted to identify all the deposited data related to carbon reclamation, the task would be futile. Current metadata often contain only administrative and organizational information, such as the name of the investigator and the date when the data were acquired.

What’s more, for scientific data to be useful to other researchers, metadata must sensibly and consistently communicate essentials of the experiments — what was measured, and under what conditions. As an investigator who builds technology to assist with data annotation, it’s frustrating that, in the majority of fields, the metadata standards needed to make data FAIR don’t even exist.

Metadata about data sets typically lack experiment-specific descriptors. If present, they’re sparse and idiosyncratic. An investigator searching the Gene Expression Omnibus (GEO), for example, might seek genomic data sets containing information on how a disease or condition manifests itself in young animals or humans. Performing such a search requires knowledge of how the age of individuals is represented — which in the GEO repository, could be age, AGE, age (after birth), age (years), Age (yr-old) or dozens of other possibilities. (Often, such information is missing from data sets altogether.) Because the metadata are so ad hoc, automated searches fail, and investigators waste enormous amounts of time manually sifting through records to locate relevant data sets, with no guarantee that most (or any) can be found…(More)”.