The reuse of public datasets in the life sciences: potential risks and rewards

Data reuse as a part of the ‘big data’ revolution

Data reuse is an essential component of open science and has been facilitated by the ‘big data’ revolution. The transition from (hand) written notes to datasets stored on hard drives can be viewed as the first step on the road to effective data reuse in the life sciences (Fig. 1), allowing the generation of multiple copies at almost no additional cost. The second step was improved connectivity, which was enabled by the internet. Together, these technological advances in data storage and transfer enabled a worldwide exchange of ‘big data’, which is common in biology (e.g. sequence data). This technological basis made data sharing possible (Fig. 1). Next, it needs to become convenient for researchers to share data through increased accessibility. Obligations and benefits lead to established sharing behavior, which results in more datasets becoming available, which can then be reused in turn. Finally, it becomes common practice and a habit to share all data, resulting in a positive feedback loop. Data sharing is already common in several disciplines, including genomics, neuroscience, geoscience and astronomy, and an increasing number of studies reuse shared data (Pierce et al., 2019; Tenopir et al., 2020).

For clarity purposes, we distinguish here between fair reuse (for novel purposes, e.g. meta-analysis), reproduction of previous studies with available data (a vital component of open science), and unjust reuse (dual publication and plagiarism). Alongside the reproduction of studies to test findings, only fair reuse should be considered a valid component of open science and is the type of reuse our discussion focuses on.

To establish data sharing as a norm it had to be introduced through obligations and promoted through benefits for researchers (McKiernan et al., 2016). Numerous funding agencies, publishers (e.g. Nature: (Announcement, 2016), NSF: (‘Biological Sciences Guidance on Data Management Plans’, https://www.nsf.gov/bio/biodmp.jsp), PLOS: (‘PLOS ONE—Data Availability’, https://journals.plos.org/plosone/s/data-availability)) and private foundations require all data be made publicly available within a certain time frame, with an indication that this leads to increased transparency in the field (Parker, Nakagawa & Gurevtich, 2016; Tenopir et al., 2020). Many international data sharing guidelines like Findable Accessible Interoperable Reusable (FAIR) (Wilkinson et al., 2016), TOP (The Transparency and Openness Promotion) (Nosek et al., 2015), Open Data in a Big Data World (Open Data in a Big Data World, 2016) and the Beijing Declaration (CODATA, 2019) have emerged by the necessity of the ‘big data’ revolution. The sharing of datasets leads to statistical robustness and allows re-analysis of existing datasets underlying authors’ claims (Open Data in a Big Data World, 2016), while enabling the discovery of novel patterns through meta-analysis (Duvallet et al., 2017). Such data reuse leads to cost reduction, reproducibility and accountability of research, scientific discovery and detection of novel biological information, and can also be helpful in other areas like education, business and government (Safran, 2017; Pasquetto, Randles & Borgman, 2017; Porto, Pires & Franco, 2017; Leonelli et al., 2017).

Despite these measures and benefits, data reuse is still not ubiquitous, for which there may be different explanations. In 2017, an analysis of 318 biomedical journals revealed that only 11.9% of them explicitly stated that data sharing was required as a condition of publication (Vasilevsky et al., 2017). Additionally, a survey of 100 datasets associated with ecological and evolutionary research showed that 56% of the databases were incomplete, and 64% were archived in a way that partially or entirely prevented reuse (Roche et al., 2015). Thus, if the publicly available datasets are not widely re-examined (either checked for quality and/or re-analysed), enforcement of open science through policy may not be sufficient to harness the full power of global sharing (Pasquetto, Randles & Borgman, 2017). The main causes of researchers refraining from reusing publicly available datasets are (1) concern about the quality and reliability of data (often warranted), (2) a lack of awareness about the potential in big data or (3) insufficient bioinformatics knowledge to mine the data (Denk, 2017). Regardless of the cause, the resulting ‘backlog’ of under-utilised reliable datasets leads to unnecessary experiments (e.g. extensive repetitive sequencing increasing costs) and likely hides useful undiscovered patterns.

Types of reusable data

There are numerous different types of datasets which harbour reuse potential (Fig. 2). These include but are not limited to (1) publications which are accessible to text mining; (2) sequences of genomes, single genes or plasmids or whole sets of sequence reads, (3) annotations of sequences (e.g. sequence motifs collected in JASPAR (Sandelin, 2004)), (4) chromatography results and mass spectra, (5) information about the structure of proteins, (6) biochemical parameters of enzymes (e.g. affinity or reaction speed), (7) geodata (e.g. coordinates of observations), (8) images of biological material or geographical regions as well as plots and diagrams, (9) algorithms and software (e.g. code), measurements of automatic sensors and global-scaled observatory networks (Soranno et al., 2015), and (11) phenotypic data (Arend et al., 2016).

Generally, datasets can be classified as primary and derived. Primary can be defined as including direct, experimentally obtained data’ and derived as including meta-analyses and processed results. Statistical data is an example for the fact that data cannot always be classified into ‘primary’ and ‘derived’. Results of a specific experiment can be directly statistically evaluated (primary), whereas results of multiple studies can be assessed and compared in a statistical manner too (derived). Further, it has to be considered that specific fields require the integration of data of various types, formats and abundance (Leonelli et al., 2017), which is hard to achieve by a single database, and therefore requires cooperation to encourage data reuse.

Purpose of the review

Education about the opportunities, challenges, limitations and techniques of data reuse is an important topic that has not been sufficiently addressed in the literature. This review aims to provide an entry point to the discussion of data reuse as part of open science, accessible to a wider audience of life scientists, not only bioinformaticians. The benefits of data sharing (and reasons for refraining from it) have been extensively reviewed in fields other than bioinformatics, for example ecology (Hampton et al., 2013; LaDeau et al., 2017), medicine (Wade, 2014; Krumholz, 2014; Safran, 2017; Hulsen et al., 2019) and cell biology (Dolinski & Troyanskaya, 2015). However, the existing reviews do not summarise the common benefits and challenges that arise from data reuse in the life sciences.

Here, we highlight the potential of data reuse as well as hurdles which need to be overcome. The presented benefits, challenges, risks and potential solutions range across different fields and aim to illustrate the inherent characteristics of data reuse. We define ‘successful reuse’ as the use of previously published data to enable novel scientific findings, usually resulting in a peer-reviewed publication. As showing case studies could reduce the initial barrier to reuse by demonstrating its value in a more practical manner (Curty et al., 2017), we use examples of successful reuse of different data types (genome, transcriptome, proteome, metabolome, phenotype and ecosystem) to illustrate its enormous potential. In addition, we provide a checklist of questions for biologists without extensive experience in handling public datasets to consider when determining whether a particular dataset is fit for reuse. We focus on the reuse of primary data, especially different sequences (as the data type characteristic of the life sciences and continuously producing vast amounts of information). The benefits for individuals and the scientific community as a whole make a strong case for reuse of data but only when the risks and limitation have been taken into account.

Preventing information loss

Making data publicly available for reuse is an elegant way to prevent information loss stemming from different issues. Data can be subject to irretrievable loss in case of storage solely on private computers and servers, which may not convert the data to the currently used format and that are subject to failure. This is avoided when data are shared with the public and stored in adequately funded databases with backup mechanisms. There are already numerous public repositories for genomic and gene-expression data, such as the Sequence Read Archive (SRA)/European Nucleotide Archive (ENA) and Gene Expression Omnibus (GEO), respectively. Recently, GEO has been used in a case study to improve dataset reusability with a literature recommendation system (Patra et al., 2020) and is highly recommended over other databases for the submission of RNA-Seq datasets (Bhandary et al., 2018). In medical research, information loss stems from large amounts of gathered data remaining inaccessible for reuse by a wider audience (sometimes even the authors) after the initial publication (Wade, 2014). Moreover, the development of new tools and methods leads to the possibility of extracting more information from a given dataset than was feasible at the time of publication. A notable example of such extraction of new information, is the basecalling step when working with nanopore sequencing data derived from Oxford Nanopore Technologies devices. Enhanced algorithms allow higher accuracy or even the identification of DNA modifications (Liu et al., 2019a, 2019b). Further, meta-analyses (e.g. the prediction of specific genomic features) and machine learning approaches require large amounts of data which are already available, and therefore should not and cannot be produced again. This is especially important in fields where data is complex, like videos in organismal biology (Brainerd et al., 2017). Such examples illustrate the importance of sharing and maintaining existing datasets, alongside supplementing them with new data, in order to prevent information loss.

Expanding scientific knowledge

Reuse of data from different sources holds immense potential for scientific discovery and therefore generally enhances scientific progress (Curty et al., 2017). Integration of data from different sources, for example, of exRNA metadata, biomedical ontologies and Linked Data technologies can facilitate interpretation and hypotheses generation by providing independent biological context (Subramanian et al., 2015). In medicine, reuse of data collected as a by-product of health care has the potential to transform the practice of medicine and its delivery, which is a compelling argument of reuse benefits outweighing the risks (Wade, 2014; Safran, 2017). For example, reuse can help eliminate bottlenecks in biomedical research at all translational levels and data-mining (the hypothesis-free search for patterns in data) can reveal potential starting points for experimental medical research (Wade, 2014). Another method of reuse is meta-analysis, which can elucidate new patterns and produce novel hypotheses—inaccessible from the analysis of any individual dataset. In turn, gene expression studies generating large datasets tend to look for general patterns and thus an in depth inspection of single genes can provide additional insights (Bhandary et al., 2018). Temporal and spatial limitations of a single experiment can be overcome by new combinations of existing data and applications integrating different disciplines can be made possible (Curty et al., 2017), leading to more interdisciplinary collaboration (Tenopir et al., 2020). When examining communities in ecology, metagenomics, metatranscriptomics, metabarcoding, and metaproteomics provide insights into community composition and function (Ten Hoopen et al., 2017). Crucially, publicly available data can lead to the development of new experimental designs (Grace et al., 2018) and can be connected with complementary knowledge and reused in novel experiments (Martens & Vizcaíno, 2017) that contribute to expanding scientific knowledge.

Maximising time, labour and cost efficiency

Reuse of data saves time and money, thus is more economical, and can offer the opportunity to overcome the restraints of limited experiments, high costs and technical difficulties (Raju, Tsinoremas & Capobianco, 2016; Curty et al., 2017). In medicine, the reliance on expensive experimental research when a wealth of existing data is available has been criticised and reuse of existing data and information was presented as a solution (Wade, 2014). In metagenomics, where datasets tend to be in the gigabyte range, appropriate archiving of workflow intermediates for reuse can decrease the costs of re-analysis (Ten Hoopen et al., 2017). Additionally, many datasets deposited in sequence repositories like GEO were collected at enormous effort and used only once and so reusing them greatly increases their utility (Patra et al., 2020). The labour efficiency of reuse is also illustrated by the famous eGFP browser (Winter et al., 2007) which presents the content of RNA-Seq datasets to biologists in a simple way. The alternative would be downloading and analysing raw RNA-Seq datasets from the SRA, requiring a substantial amount of bioinformatics expertise and computational power during the analysis. Valuable computational resources are provided by international and national cloud computing services like Elixir, CyVerse or the German Network for Bioinformatics Infrastructure (de.NBI) as well as by commercial organisations. Data reuse is a profitable choice for researchers to lessen expenses and shorten the research process as data collection was already performed by others (Curty et al., 2017). Cutting such costs through reuse (Fell, 2019) enables groups with small budgets to harness extensive datasets and thus enhancing equality.

Benefits for authors

While the scientific community and society are profiting most from public datasets, there are additional benefits of open access for authors themselves (McKiernan et al., 2016; Leitner et al., 2016; Ali-Khan, Harris & Gold, 2017). Researchers can build a reputation by generating high quality and well-documented datasets. Although compliance with data standards might be seen as an additional burden in some cases, the chances of reuse occurring are increased by providing data in the proper format (Rocca-Serra et al., 2016). Dataset sharing may also increase visibility of the associated research and results in additional citations; an added encouragement for authors (Piwowar, Day & Fridsma, 2007). There are even reuse examples which are possible without biological context, like benchmarking of bioinformatic tools or the identification of patterns (Bhandary et al., 2018). It has been shown that there is a robust, statistically supported citation benefit from ‘open data’ in comparison to similar studies without publicly available data (Piwowar & Vision, 2013). This especially aids early career researchers, who are outsiders of the scientific establishment and likely experience more barriers to other aspects of open science (National Academies of Sciences, Engineering, and Medicine, 2018), yet are highly involved in data collection and analysis (Farnham et al., 2017).

Benefits for databases

The reuse of sequence data is of increasing importance due to the large and rapidly growing size of the databases storing them (Fig. 3). The size of the SRA alone increased from 3,092,408 entries to 6,243,265 entries within 2 years (September 2016–September 2018) (NCBI Resource Coordinators, 2017; Sayers et al., 2019a). This growth rate continues to increase exponentially (Fig. 3). GenBank comprises a total of 3,677,023,810,243 sequences (2018) with an increase of 39.52% in comparison to 2017 (Sayers et al., 2019b). Approximately 120 million sequences and annotations of proteins were available within UniProtKB/TrEMBL in 2018 (The UniProt Consortium, 2019). Since managing an exponentially growing database has numerous challenges (Lathe et al., 2008), it is important to consider how they can be addressed through changing practices, including data reuse.

The issue of the rate of nucleotide and proteomics data generation growing faster than storage capacity (Cook et al., 2016) can also be partially addressed through data reuse. Reusing available data instead of producing new and redundant datasets (i.e. when a large number of datasets that are in consensus is already available) results in a lower number of duplicates (Grace et al., 2018) and keeps databases concise. Since it takes an enterprise to maintain and upgrade the largest and most-used databases (meaning they must be adequately financed to survive), limiting redundancy through reuse is beneficial. One example of a sustainable business model is The Arabidopsis Information Resource (TAIR) database (Lamesch et al., 2012) which is funded by subscriptions from academic and non-profit institutions but allows a limited number of accesses by individuals. The public availability of datasets also allows the development of effective algorithms to tackle the bottleneck of data processing, all without the need to perform any sequencing (i.e., uncoupling the problem from access to sequence technology and allowing participation of e.g. computational fields). Additionally, not only the storage of a large number of datasets but also the actual reuse of the available data might increase funding of public databases and therefore ensure the long-term existence of these infrastructures.

Unknown quality

For successful reuse appropriate data quality is required in order to be reliable for the user. There are inherent quality differences between (1) user-submitted public datasets, (2) carefully curated databases for specific organisms, (3) ones with inherently small holding sizes (like PDB or SwissProt) and (4) phenotype databases. Additionally, information regarding experimental design, methods and conditions is often incomplete and results in datasets unsuitable for reuse. Mislabelled or swapped samples alongside intrinsic errors, like missing technical replicates, pose a problem that is almost impossible to identify when accessing a public dataset. Precise documentation of the workflow is crucial to clearly indicate limitations of the generated data product (Soranno et al., 2015). Despite this, the peer-review process rarely reaches the datasets and their descriptions (Patra et al., 2020). There is also a trade-off between the collection of detailed metadata during submission and high submission numbers. However, additional requirements for data submission should not result in fewer publicly available datasets (Rung & Brazma, 2013). Ultimately, the limitations of each study (and dataset) are best known by the primary investigators and not by the community accessing the data—a trade-off that studies based on reused data must consider.

It is also difficult and time-consuming for the user to discover available data which is relevant and suitable for the analysis and further to ensure sufficient quality of these datasets (Curty et al., 2017). Even if complete metadata are provided, accessing numerous different webpages to collect all information associated with a combined dataset can be tedious (Bhandary et al., 2018). This leads to a risk of wasted time and effort on flawed data, thus requiring trust in data producers and their methods and techniques (Curty et al., 2017). Moreover, simply using a large amount of publicly available datasets does not inherently lead to correct patterns. Despite the importance of trends revealed from large datasets, it is not necessarily the case that a large number of reads/replicates, with an associated low noise, means that the emerging trend is true. Conversely, one can imagine a trend with low noise and low deviation, produced from a large dataset, but with the data coming from one author/group that has an undetected systematic error. Equally, low noise and the use of a small dataset that is believed to be of ‘higher quality’ (i.e. has been thoroughly checked for errors), may hide a trend or show a nonvalid one. Ultimately, the low availability of necessary metadata in standardised formats, with insufficient additional information leads to a lack of reproducibility and can result in misuse and wrong assumptions (Rung & Brazma, 2013; Curty et al., 2017).

Denormalization

Of particular concern is the circular reuse of data. It can, for instance, lead to heavily denormalized annotation in databases, i.e., the same data is stored multiple times in the same database under different names/identifiers (Bell & Lord, 2017). When such data duplication is not recorded and the user is not made aware of it, the data distribution in the database does not reflect the true data distribution. For example, sequencing and annotation errors can be propagated by reuse and not eliminated by additional published sequences that would reveal it to be statistically insignificant. For annotations, it has been shown that it is possible to detect low-quality entries (resulting from this denormalization) by looking for specific patterns of provenance in the database (Bell, Collison & Lord, 2013). With respect to gene models, this problem could be addressed in the future by integrating RNA-seq datasets in the annotation of new genome sequences. In terms of functional annotations, this issue persists because the experimental characterisation of numerous genes in a diverse set of species cannot be expected in the near future.

Comparison and integration of datasets and databases

The comparison and integration of datasets from different sources remains a challenge of reuse (Pasquetto, Randles & Borgman, 2017). In metagenomics, when communities (that have been studied independently) are compared several issues arise, including differences in workflow, unrecorded variables, non-unified presentation format, and relevant raw data not being publicly available (Ten Hoopen et al., 2017). The same issue is illustrated by the enormous differences in annotation provided by the different databases, for example on NCBI (Genome), ENSEMBL, and Phytozome for the same species. In plant phenotyping, reuse and meta-analysis is challenging as data comes from different experimental sites, plant species and experimental conditions, while including many different data types (Papoutsoglou et al., 2020). This non-comparability is also repeated in medical research (Wade, 2014). Therefore, for any valid comparison between datasets from different databases or for integration of databases themselves, conditions specific to the data type and field have to be satisfied.

Different file types and data structures pose a universal challenge for integration which can be overcome by the use of standards if these are de facto accepted by the community (Rocca-Serra et al., 2016). For communication between disciplines, an interdisciplinary team with ‘brokers’ has been recommended for the setup of new databases (Soranno et al., 2015). Consistent use of controlled vocabularies and standardised languages like XML (Soranno et al., 2015) also enhances the value of data collections substantially. For example, in meta-analysis of sequences an established and unified file standard is crucial (Ten Hoopen et al., 2017). FASTA (Pearson & Lipman, 1988), FASTQ (Cock et al., 2010), and SAM/BAM (Li et al., 2009) are famous examples of file standards that have allowed the effective exchange of information between numerous groups involved in the earliest sequencing projects (Leonard & Littlejohn, 2004; Ondřej & Dvořák, 2012; Zhang, 2016). Any disparities in the sampling method also have to be taken into account when biological material is concerned, so it is essential they are recorded appropriately (Ten Hoopen et al., 2017). When appropriate, unified workflow reporting standards (like described by Ten Hoopen et al., 2017) within a field would largely remove these inhibitions to reuse. It is for these reasons that guidelines like FAIR, which (Wilkinson et al., 2016) promote universal metadata standards across-the-board, are essential to allow comparison and integration of datasets.

Re-analysis as a possible solution

Re-analysis of publicly available data is one way to tackle the issues of unknown data quality, data denormalization, and database integration which arise with reuse. This can be achieved through curation and self-correction, with both being difficult to directly enforce. In the same manner that the re-examination of public biodiversity data leads to error correction (Miller et al., 2015; Zizka et al., 2019), so should sequence repositories reflect changes in the field’s consensus (e.g. on specific gene annotations). A way to address some of the risks of reuse through re-analysis would be investing in a controlled environment containing extensively peer-reviewed datasets (Spertus, 2012) and manually-curated databases. A defined, suitable environment or database could also include follow-up data for a detailed understanding of the primary data and the corresponding results. Further, re-analysis and reproducibility might be improved by conventions for standardisation, documentation and organisation of analysis workflows (Lowndes et al., 2017). This includes detailed records using open science tools like shared GitHub repositories (Lowndes et al., 2017). Crucially, reuse cannot occur if produced datasets are not widely released to public. Re-examination of databases, like that by Grechkin, Poon & Howe (2017) of SRA and GEO, to automatically identify datasets overdue for release are vital in this effort.

Re-analysis has proven to be efficient with some data types but is not practical in all cases. An excellent example of how investing in manually curated databases eliminates many issues are ‘expression atlases’ with annotated sequences checked for quality and re-analysed using standardised methods (Kapushesky et al., 2010). However, regarding the enormous and still increasing amount of sequence data, this is hardly an option for all data types. An analysis of reused data types already indicates that studies often rely on previously identified differentially expressed genes or calculated gene expression values instead of processing raw data again (Wan & Pavlidis, 2007). Therefore, different strategies might work for different data types or different communities. In all cases, specific standards and formats for data reuse should be applied (Pasquetto, Randles & Borgman, 2017), lest “the wealth of data becomes an unmanageable deluge” (Parekh, Armañanzas & Ascoli, 2015).

Metainformation as a possible solution

The trade-off between public access and unknown quality can be partially resolved by the publication of metadata (the information about the acquisition, processing and presentation) associated with a particular dataset. So far, most researchers (almost 60%) share their data and metadata using institution specific standards only or even without general metadata standards at all (Tenopir et al., 2020). Further, the metainformation necessary to make data findable (Tenopir et al., 2020) and to enable successful reuse differs between data types and between fields (Parekh, Armañanzas & Ascoli, 2015; Brainerd et al., 2017; Ten Hoopen et al., 2017; Papoutsoglou et al., 2020). The amount of metadata that can feasibly be recorded also varies by field, for example, the metadata about a single blood test on a patient includes countless variables (Safran, 2017).

Submitters need to be aware of the importance of providing accurate and complete metadata, but also that a controlled vocabulary is required to facilitate automatic identification of relevant studies/samples (Bhandary et al., 2018). People accessing the dataset also need to be aware of missing information about a dataset—an issue that can be resolved by including a ‘completeness of metadata’ search criterium in database search engines, like implemented at NeuroMorpho.org (Parekh, Armañanzas & Ascoli, 2015). The MassIVE Knowledge Base aggregates proteomics data including statistical controls and records of the data origin to ensure high quality of the datasets (Doerr, 2019) and is thus an example for a database providing data fit for reuse. The recently updated MIAPPE metadata standard for plant phenomic databases is another example of reuse facilitation through metadata formatting and was developed to address the shortcomings of the previous guidelines preventing FAIR-complying reuse (Papoutsoglou et al., 2020).

Additionally, methods for reconstructing existing databases are already being investigated. This can be done by curating existing metadata or extracting more of it through natural language processing techniques (Patra et al., 2020) and through metadata predicting frameworks (Posch et al., 2016). Many sequence databases, such as ENA, are handling this elegantly and submitting users can provide a very basic set of metainformation or provide comprehensive details about their study. There are also easy-to-follow instructions for the submission process (European Nucleotide Archive (ENA), 2020). Despite such incentives, many datasets still lack descriptions that would allow them to be re-sorted according to their metadata (Patra et al., 2020).

Publication of metadata is a practice already routinely implemented in data papers and data journals. Data papers (already common practice in Astronomy (Abolfathi et al., 2018)) have been indicated as a solution to the quality-check problem (Chavan & Penev, 2011) of reuse by providing descriptions of methods for collecting, processing, and verifying data (Pasquetto, Randles & Borgman, 2017). Widespread publication of such metadata in data journals (Figueiredo, 2017) is vital to the construction of high quality, peer-reviewed datasets. Consequently, data-focused journals, for example GigaScience, Scientific Data, and F1000, emerged during the last years. As long-read sequencing became affordable and paved the way for numerous high continuity assemblies, genome announcements describing new genomic or transcriptomic resources became popular. These publication types provide an elegant solution for data reports if a valuable dataset should be shared with the community but does not meet all criteria for publication as a full research article.

Ethical considerations

Ethical considerations of data reuse are important to inspect in all life sciences fields and have been previously discussed in the context of clinical studies (Wade, 2014; Safran, 2017). Informed consent may not have been given with the knowledge that personal data will be utilised in more than the primary study. When sharing medical data, patients must not be identifiable, even if advanced methods are applied. This can require modification or masking of the dataset elements, for example defacing of brain images (Milchenko & Marcus, 2013). Similarly, genomic data (a part of modern medicine) is not covered by HIPPA and a parent’s genomic data could create a privacy risk for their children (Safran, 2017). Thus, there are ethical considerations specific to reuse of data in human research.

Research ‘parasitism’

The use of the same dataset in several different studies by the same author could be considered a type of dual publication in some circumstances (Beaufils & Karlsson, 2013). However, such reuse is not contentious to the same extent as plagiarism if it reveals novel findings and is not reused only to boost the number of publications. The trend of publishing from publicly available data (‘The parasite awards—Celebrating rigorous secondary data analysis’, https://researchparasite.com/; Longo & Drazen, 2016; Pucker & Brockington, 2018; Frey & Pucker, 2020) points to the crux of the matter of research integrity reservations about data reuse that some have. At the far end of this spectrum, there are authors exclusively using publicly available data (not generating their own to cross-check the quality), often choosing research topics/systems-of-interest based on the quality of data and not vice versa. This practice is associated with numerous advantages, including intensified use of existing datasets which effectively increases the ratio of value drawn from it compared to the costs of generating it in the first place. While multiple studies can benefit from reuse, long term risks might include funding bodies expecting reuse and rendering the acquisition of financial support for new experiments more challenging. Despite some expressed concerns regarding such ‘research parasitism’ (Longo & Drazen, 2016), including the fear of exploitation when acquiring the data was particularly expensive or labour-intensive (National Academies of Sciences, Engineering, and Medicine, 2018), the practice of reuse seems to prevail in the open science culture. The above-mentioned ‘Parasite awards’ use this tongue-in-cheek name to reward such practices, after the term was introduced to reprove of them. Due to the numerous benefits of reuse for the scientific community, we believe the term ‘research parasitism’ is unwarranted when fair reuse has been employed.

Recognising the value of data reuse by recognising the data producer

As perceived efficacy and efficiency of data reuse strongly influence reuse behaviour, it could be encouraged by demonstrating its value (Curty et al., 2017). So far, there have been extensive efforts to promote and develop standards for data sharing, but less effort to show the real value of data sharing or to recognise, cite or acknowledge the contributions of data sharing (Pierce et al., 2019; Tenopir et al., 2020). Adequate recognition of data producers would accelerate data sharing by eliminating the main barrier: the need to publish first (Tenopir et al., 2020).

Both data providers and creators of databases deserve recognition for their work. To a certain degree, this issue can be tackled with data publications, which might also prevent the splitting of a coherent dataset over multiple publications. Soranno et al. (2015) recommend specifically to describe the content of new databases in one paper with all contributors listed as co-authors and a description of the methods for the database development in an additional publication with all researchers involved in the process as co-authors. This is a practice that shares all its benefits with data papers (discussed above). The preservation and documentation of data provenance is crucial in aknowledging the support of data providers (Soranno et al., 2015). By supplying a citeable source of the dataset, credit is given to the data producer, which eliminates the concern about ownership by providing an official academic record of provenance.

As barriers for data sharing include concerns about loss of credit (Tenopir et al., 2020), the assimilation of sharing and reuse could be positively influenced by recognition (e.g. awards or other compensatory means, like co-authorship) of the expansion of scientific discovery through studies reusing available data (Piwowar & Vision, 2013; Curty et al., 2017; Tenopir et al., 2020). A way to recognise the producers of reusable data would be the connexion of an identifier for each researcher (e.g. ORCID) with an identifier for the dataset (e.g. DOI) which would then have to be cited in each new publication reusing the initial dataset (Piwowar & Vision, 2013; Pierce et al., 2019). In addition, publishers would have to ensure that these citations are available in a searchable system (e.g., Crossref) to establish a real link between data generator and publication (Pierce et al., 2019). Such a system, which recognises researchers regularly for generating data, could substantially influence the assessment of the value of scientific data by academic institutions, funders, and society (Pierce et al., 2019).

Establishing a reuse culture

Despite the reservations highlighted above many facets of data reuse provide incentives for the individual to practice it. Through open science initiatives (including but not limited to those listed in the Introduction), modern biologists are encouraged to make use of publicly available sequence repositories and mine data generated by others. Further, not comparing one’s dataset to publicly available analogues can be considered akin to ignoring replicated experiments (Denk, 2017). In fields where scientific progress has immediate and measurable positive impacts, such as medicine, the benefits of data reuse to society quickly outweigh the risks (Safran, 2017). Due to these advantages, there is an argument to be made that data reuse is an ethical obligation in the life sciences.

The statistic that the metagenomes of 20% of papers published between 2016 and 2019 are not publicly accessible (Eckert et al., 2020) demonstrates that there is still a long way to go until data sharing becomes routine. Therefore, open science incentives and database contribution guidelines should require the inclusion of metadata in all submissions to public datasets. Gamification via the implementation of fancy statistics about the data connected to a personal profile for each researcher might be another way to encourage researchers to share their data. A reusability score assigned by the community could increase the quality of the provided metadata. Such practices would not only encourage authors to collect data with reuse in mind (Goodman et al., 2014) but enable productive and valid re-analysis. Reinforcement could include spot-checking of provided metadata by funding agencies (Bhandary et al., 2018), as they have a monetary interest in the data quality of projects they support.

Only a combination of obligation and encouragement from the publishing and educational spheres are likely to ensure a future of successful and fair data ‘recycling’. The prevailing culture of positive publication bias (Mlinarić, Horvat & Šupak Smolčić, 2017) leads to ‘missing studies’ (the desk-drawer effect) and could also introduce a bias into analysis based on existing datasets (Wan & Pavlidis, 2007). For this reason, young researchers should be encouraged to share their data (if it is of sufficient quality) even if the outcome was ‘negative’. Additionally, bioinformatics should be integrated into the education of next generation life scientists to increase data re-use capacities (Soranno et al., 2015). As one learns about a new method of data collection so should one learn about the metadata that must accompany it in publication to render the data useful to others.