Licensing of data and software

The open science movement rests on a foundation that research data, software code, and experimental methods should be publicly available and transparent unless privacy or confidentiality is an issue [8]. The first foundational component of DaaS is straightforward, permissive, human- and machine-comprehensible licensing. Licenses need to be simple to reduce confusion [33–36] and designed to allow automated access to data. A call to license data and software is not new, but licensing, copyright, and data use agreements are poorly understood [33,37,38], delaying their application. A restrictive license, in this context, is any license that places additional requirements or caveats on use of the data. Investment in data citation, data publication, microannotation, and nanopublication [39–44] will reduce the need for the restrictive licenses and nonstandard use agreements that are often in place to track reuse and impact. A global system of interconnected data with automated methods of tracking use and apportioning credit requires standardized, machine-readable licensing and data use agreements.

Data integration is still a largely manual task

As with people and ideas, data have been siloed within discipline-specific, project-specific, or lab-specific environments. The key to integrated data among silos is the universal adoption of community-developed standards [45] (Box 2). Even with most current standards, substantial manual effort can be required to map overlapping schemas or transform data from one context to another. Standards which describe data and metadata, document formats, content, protocols, and vocabularies are essential for automated integration [46]. With appropriate standards, the transformation workflow can be automated, it can include multiple quality checks, perform transformations, promote reproducibility, identify provenance, reduce manual errors and inconsistencies; all at a lower cost. The automation of integration will add machine-actionable links across repositories that will result in a network of data repositories. An example that integrates across repositories is Biolink, a graph-oriented data model that was developed for a biomedical use case but is being extended to the rest of biology [47]. Domain standards are a concrete way to increase interoperability and integration across repositories given that the basic elements (i.e., semantic types) of biology and the relationships between them are identified consistently across resources. Yet, the context-dependent nature of data transformation is not well represented by existing standards. A solution may lie with microschemas, highly localized data models to ensure accurate context-dependent data transformation similar to the GA4GH schemablocks concept [48] that can facilitate automation of highly contextual data like total cholesterol or ocean acidification.

The second foundational component of DaaS is standards to support machine actionable metadata and corresponding algorithms for automated standardization and integration.

Automated data integration requires standards. The incentive structures for most academics do not include the development of standards or scientific software as intellectual outputs. A lack of data standards impedes progress, and it is now timely to acknowledge efforts to improve standards and the tools and services which support their use [15,49].

Metadata are underappreciated

Despite the importance of metadata, their creation is still neglected [50]. Collecting metadata at the same time as data they describe is a recognized best practice [9], but this does not always happen and substantial amounts of data have been collected without standard metadata. The third foundational component of DaaS is algorithms for automated metadata creation and standardization with documentation and provenance. New tools are needed to automate the generation of metadata across data types and scales, where possible [46]. Machine learning (ML) and artificial intelligence (AI) can enhance metadata with standards and detect appropriate protocols for data normalization. High-priority automated tasks include named entity recognition, data and semantic typing, and protocol detection. Algorithms for semiautomated crowdsourced curation will benefit quality control and other tasks that cannot be fully automated [51]. Some entity recognition algorithms already exist [52–54] but have not yet received wide adoption because of problems with usability, sustainability, and discoverability of the tools; or because of the need of changes to work practices. Without a strong user community, it will be hard to recruit resources to create and improve these tools. One perception is that metadata preparation and documentation is altruistic and without significant impact [19,50]. While not a universal view [23,55], better professional rewards that value metadata creation and associated tools are needed.

The quality of data and metadata is variable

Issues relating to quality include social and technical aspects of the data, metadata, data providers, aggregators, and repositories [15,56–59]. Numerous studies explore trust in data sets and the expectations of users [15,42,43,60,61], but there is no widely adopted, formal process for judging data set quality. The peer review system for publications, even with its flaws [62], can provide a starting point for an assessment of the quality of data sets [42]. The fourth foundational component of DaaS is a simple, predictable, and transparent system for peer review of data. While some repositories have a review process for submitted data sets that may include automated checks, and data publication journals can review data set documentation, routine rigorous peer review of data has not been implemented. One deterrent is that the pool of likely reviewers is already overburdened. If peer review of data sets is to have any hope of implementation, an infrastructure that puts reviews to good use and apportions credit for conducting reviews is needed. A supplementary approach is to use annotation technology to enable feedback on data sets and data atoms by end users [63].

Data use and contributions are hard to track

Researchers typically use citation metrics of publications as a measure of the impact of their career. The other types of activity, such contributions to an infrastructure, or third-party reuse of data, are often neglected because, historically, no comparable systems exist to track other endeavors. DataCite [40], an advocate for data access and discovery, developed a data set citation standard with DOI assignments that is used in several disciplines, but additional supporting infrastructure is needed to fully understand what a data citation, and the resulting metrics, means for the career of the producer and the value of the data themselves [64]. A roadmap for a system that supports standardized data-level metrics [65] was developed by MakeDataCount.org and is available for implementation. This roadmap fills many of the important technical gaps but requires resources to increase adoption across repositories, publishers, institutions, and researchers in order to create a system of data metrics comparable to publication metrics.

The mutable nature of data sets raises issues with identifiers and versioning that do not apply to publications [43]. For reproducibility, a published analysis must point to a persistent version of the data that were used, even if a small change was made after publication. In addition, credit needs to be apportioned appropriately when data sets are curated, subdivided, combined, vetted, and upgraded in ways that manuscripts are not. This raises several provenance and attribution issues that can only be addressed by well-documented versioning with a robust chain of provenance (for an example system, see [66]). When a data set is downloaded for analysis, that chain is usually broken, making it nearly impossible to communicate usage metrics back to the source and other agents in the data supply chain. Mechanisms and infrastructure that bring analyses to the data will better reveal the entire workflow so that it can be reproduced and refined. The fifth foundational component of DaaS is a transparent pathway for preserving provenance and attribution within analytical environments. Many computing environments for large data sets comply with this component because most researchers do not have the local resources to manipulate very large data sets. Researchers with small data sets that can be handled by spreadsheets are less likely to preserve these metadata. Collaborative environments with a support infrastructure similar to git, GitHub, or FilteredPush [63] engage all participants in data stewardship and to make the pathway of content flow, value-adding, and analysis visible.

Good data managers and curators are scarce

If we are to make full use of rapidly changing technology, we need data expertise coupled with in-depth biological knowledge [67]. People with such skills are rare. Increasingly, biologists will require data training, but this is not sufficient to create new advanced tools. Rather, we require a professional development structure for a community of biologists with advanced expertise in data management, curatorship, software development, and information science. The sixth foundational component of DaaS is a method that attributes and credits the work of professional data managers and curators that is equal to manuscript citation and can accommodate microannotation and nanopublication. The Contributor Attribution Model (CAM) [68] used in combination with microannotation or nanopublication [39,69,70], wherein metadata are associated with individual data atoms (smallest usable elements), can underpin a system of attribution tracking where individual credit cascades through the long pathway of content flow [37,39]. CAM builds on the work of groups like CASRAI [71] by using CRediT [72] to inform the Contributor Role Ontology—the source of role terms in CAM [73]. (Domain-specific groups like CASRAI are an integral part of developing the community standards discussed above.) There are a few existing systems for recording attribution such as ORCID [74], OpenVIVO [75], and rescognito [76] that have begun to tackle the issue of credit for data work. With more transparency, the investment in making data more reusable becomes more measurable and removes the disincentive of working hard for no reward [19,22,60].

Sustainability for data remains elusive

The current strategy for funding scientific research leaves most data unsupported after completion of the project. An essential but often overlooked aspect of data integration is long-term preservation. Repositories, including museums and libraries, have the knowledge and expertise to provide sustainable preservation of data [27], but many data repositories accommodate only a single subdiscipline or data type (with a few exceptions, e.g., [29]). CoreTrustSeal promotes best practices for repositories committed to the long-term preservation and curation of data [77].

Fitness of data can corrode with time, and this requires maintenance of the schemas, metadata, and even data (Box 2). An example is when names of and concepts for taxa change [37,56], but their use as metadata remains uncorrected. While many repositories regularly update their content, including the creation of new data products (e.g., [78]), others lack the resources or disciplinary skills to make these updates quickly. This leads to dissatisfaction with the current ecosystem of long-term data support [15,56]. One reaction is for researchers to maintain data locally; but the probability that project-oriented data environments are available for reuse decreases by 17% per year [79]. The seventh foundational component of DaaS is low-cost, reproducible workflows that convey data and metadata from creation to an accessible trusted repository that delivers data that are fit for purpose in perpetuity. Lack of preservation resources places much of our collective digital knowledge in jeopardy, is dismissive of the investment in creating data, threatens future insights that might be drawn from data, and decreases our ability to engage in reproducible science.

Our vision of a reintegrating biology

It is inevitable that an extensive integrated data environment will emerge. With it will come new opportunities for discovery, devices to address problems with greater scale and scope, and the quality of insights will improve [80]. There are several leaders in this developing space, including CyVerse—an open science workspace for collaborative data-driven discovery that offers an infrastructure that supports data throughout its life cycle regardless of platform [81]. CyVerse supports access to high-performance computing and the packaging of data and software together in reusable “Research Objects” such as those used in the Whole Tale Environment [82]. A DaaS model can promote the emergence of a more extensive network of curated, interlinked data within an environment that is rich in tools and interoperable services. Progress is impeded because much of the required self-reinforcing infrastructure is absent. We emphasize 2 barriers to achieving the foundational components discussed here. First: motivating the sustained community participation that is needed to develop and implement discipline-specific data integration solutions—especially in respect of discipline-specific and domain-specific standards that enable the automated components that make large-scale integration tractible. Second: the data citation and peer review infrastructure (beyond DOIs) needed to motivate professional participation in data-centric activities does not yet exist. The interconnected nature of these problems means that partial solutions, which are easier to fund, will not have the desired impact. The role of publishers and aggregators of intellectual output, like ORCID, in making this vision a reality cannot be overstated [83,84]. Some of the early progress with incentivizing data sharing were led by the requirements of publishers [85], and they remain a major driver of data sharing behaviors [22]. Publishers, researchers, and repositories will need to collaborate to adopt and enforce a standard of data citation and peer review that, combined with infrastructure supporting provenance, microattribution, annotation, and versioning proposed here, can perpetuate credit across a workflow. Well-formed attribution metadata can make essential data-related tasks just as visible as traditional publications. This is key to improving the academic incentive structure that currently demotivates investment in data-centric research practices.