Standards arising from bioinformatics research for human biology

It is now sufficiently clear that Web Services can play an effective role in this context, but, in order for them to achieve a widespread adoption, standards must be defined for the choice of Web Service technology, as well for semantically annotating both service functions and the data exchanged; furthermore, a mechanism for discovering services is needed. However, experience is now overwhelming that the real standards used in biology emerge from the community.

Intended interactions between EDAM and BioXSD are shown in Figure 1.

Another project funded in EU FP7, was GEN2PHEN [11]; it pioneered the data, database and workflow challenges in collecting and sharing human genotype and phenotype data.

Finally, Open PHACTS [12] is a major knowledge management effort launched under the Innovative Medicine Initiative (IMI) framework. It is widely supported by pharmaceutical companies, but it also moves beyond the pharmaceutical realm. It is the first project that yielded a widely used infrastructure based on Semantic Web technology. The project attracted multiple associated partners and the beta version had more than one million hits. A rapidly increasing number of public and private partners adapt their services to use the content and the data model (described in Resource Description Framework (RDF) with rich provenance) of Open PHACTS. The technology developed by this project is generic and will increasingly be adopted in other fields of the life sciences.

Beyond the human species

A variety of bioinformatics software solutions, analytical methods and common procedures and standards purposely devoted to the human species are available. Conversely, non-human life science communities exist with different degrees of scientific coordination, and some areas have already agreed on ontologies and common procedures and standards. But only few efforts have gone into the wider task of harmonizing the research efforts of these communities.

By adapting existing technologies from the field of human bioinformatics and developing them further, it is possible to build working infrastructures for bioinformatics within the non-human fields of life sciences. AllBio [13], a EU FP7 KBBE project, coordinates efforts to make the human genome related technologies operational in the fields of microbial, plant and livestock. Partners in AllBio collaborate on subprojects such as the design of ontologies for data and methods, and the choice of common interoperability standards.

Cyber-infrastructures

The BioMedBridges project [15] was funded to investigate the creation of bridges between the ESFRI and other research infrastructures in order to prevent the development of data silos and non-interoperable tools. Currently, more than 10 ESFRI projects started to coordinate efforts in BioMedBridges. This shows a natural tendency of countries and international projects towards sharing the burden of data stewardship and management. Optimal sharing of best practices, data sets, tools and infrastructures within disciplines of biology, but notably also across the human, animal, plant, nutrition and biotechnology disciplines, will be driven by scarcity and scalability of resources. Most notably, ELIXIR [16], with the tagline "data for life", is a candidate for a coordinating role. Communities should monitor the development of ELIXIR and, where possible, align, coordinate and share local expertise with this international environment.

New integration and search challenges for Genomic Computing

Management of data generated by NGS technologies is a paradigmatic illustration of the so-called "big data" challenge. The integration and search of genomic information is a problem of its own and serves as a leading example, as we expect comparable quantum leap developments in other "-omics" technologies and imaging as well. Current formats and standards for the representation of NGS data are inadequate to support efficient and high-level search of information, as they are mostly concerned with the encoding of DNA-related information and not with its use within query and search systems. The interoperability standards, such as the Distributed Annotation System (DAS) [17], appear similarly inadequate to efficiently support interoperability at the level that is required by the breadth and complexity of exchanged information. In reality, the "data deluge" generated by NGS technologies has not been matched by corresponding progress in data query, integration, search and analysis, thus creating a gap in the potential use of NGS data.

Covering the gap is not easy. On the one hand, data integration is known to be a hard problem in all fields, as it requires coping first with different data semantics across data sources and then with efficient data sharing in the presence of replication and errors. In NGS, problems are amplified by the lack of standards for exposing data semantics at a level where it can be well understood and appreciated. Indeed, while the bioinformatics community has made enormous progresses in the description of several general-purpose ontological sources, similar attention has not been given to a high-level description of experimental and annotation data. Several databases and ontologies, as well as tools, exist for describing the general features of experimental data [18, 19], or for connecting to annotations and displaying the corresponding DNA regions [20], but little emphasis has been put on describing experimental results going beyond data formatting. Hence, the focus on specific DNA regions of experimental data sets, in order to "read" the experiment from a particular biological or medical perspective, is not adequately covered; but such capabilities are key ingredients to the support of biological research and personalized medicine. Indeed, the most important data integration involves the human genome, which description undergoes frequent updates of alignment references. his has been discussed for a period of over a year by a group of engineers and computer scientists from the Politecnico di Milano (Polimi) with biologists and bioinformaticians of the Istituto Europeo di Oncologia - Istituto Italiano di Tecnologia (IEO-IIT). The conclusion of this group is that the field requires a revolutionary, data-centred approach [21]. Practices established in various research laboratories, involving data alignment and data analysis pipelines, cannot be easily interfered with. Yet, there is a need for interpreting experimental data at a high level of abstraction, in terms of specific properties of genomic regions. Such interpretation is facilitated by the presence of few well-understood physical data formats (e.g. FASTQ, SAM, BAM, BED, bigWig, etc. [22]) that are suitable for data extraction via simple wrapping technologies. Thus, it is possible to extract region-aware data in high-level format from experimental or annotation data sets (e.g. height, width and probability of peaks in a ChIP-Seq experiment which satisfy a given threshold of extraction, relative to the genome region where peaks are expressed). Such information is much more compact and semantically rich than the one that can be expressed in the BAM, BED or bigWig data formats. Furthermore, it can be processed by high-level programming languages.

A joint Polimi/IEO-IIT effort is ongoing towards the definition of a "genometric data model" and a "genometric query language" which can be used to describe the information contained within each data set. The data model associates a semi-structured collection of metadata with each experimental data file; moreover, each data set is transformed into genomic regions, each having coordinates relative to a reference assembly and associated with specific data (e.g. describing mutations, gene expressions, transcription sites, etc.). The genometric query language is capable of high-level operations such as comparing experiments, extracting their most interesting regions and mapping a given region description to another within the "genometric space". It thus provides a good starting point both for pattern-based queries, e.g. by extracting experiments or regions that exhibit specific data patterns, and for data analysis, e.g. by constructing region-to-region or experiment-to-experiment networks highlighting their similarity or relationships. While this research project is not the only one dealing with high-level query languages for NGS data (e.g. see [23]), it aims at preserving the way in which data sets are produced and primarily analysed in experimental laboratories, and operates on top of these primary analyses.

A high-level description of genome information is the starting point of content-based indexing and search. Genomic information can be indexed by information content and significance in much the same way Web pages are indexed by word content and significance. If research centres make their NGS data available in some form of Web interface (ranging from basic DAS 1.6 or DAS 2.0 versions, up to direct exposure of an Application Programming Interface (API) to genometric queries), it will be technically possible to implement a process very similar to Web crawling, extending access to all the research centres who agree to share (part of) their NGS data.

By coupling indexing to crawling, we come to a vision of the "Web of Genomes" as a powerful infrastructure supporting the future of Genomic Computing. We can initially assume simple search patterns, such as finding experimental data related either to a given pathology (based on metadata) or to a panel of mutations localized on DNA regions (based on region information). Search patterns may then grow in complexity, up to encompassing similarity search with specific genome regions, which characterize a given experiment. The identified and extracted insights from highly variable original data files could then be stored as associations in a computer readable and interoperable format, such as the provenance-rich RDF, i.e. as nanopublications (see next section), that would seamlessly connect them to all core legacy information in the same format.

How far is this vision? We still need a number of well-traced technological achievements, based on well-established practices that have been already applied to other fields and primarily to the Web as we know it. This vision arises from strong previous expertise in generic data and Web management. It can progressively turn into reality through coordination, co-operation and support towards reshaping the field of data search and integration for Genomic Computing.

From data to information

As soon as data sets have undergone a first level of pre-processing and analytics, even before the actual biological interpretation and knowledge discovery start, the "associations" and "assertions" about how concepts addressed in the data sets may relate to each other become apparent. This can range from simple associations, such as co-location or co-expression, to entirely fleshed out assertions about how a given post-translational modification influences the 3D structure of a protein. Such associations, in essence, follow the model of "subject-predicate-object" (SPO) triples, such as those operated in RDF. Notably, certain assertions need more than one SPO to become meaningful. For instance, a single nucleotide polymorphism (SNP) in a given position (triple 1) and in a given species (triple 2) may cause a certain protein change (triple 3). Therefore, in many cases a small named graph is needed to form a minimal assertion. Following this principle, a "nanopublication" has been defined as the smallest possible meaningful assertion published in RDF [24]. It also relates to the concept of "Research Objects" [25], which has been introduced in order to capture the need for more formal modeling in biology in the broadest sense. A Research Object is an aggregation object that bundles together experimental resources that are essential to a computational scientific study or investigation. If we define a research object, ad interim for the purposes of this position paper, as "any identifiable artifact produced in the activities of research and formatted for computational studies", such research objects cover formal data models of level 1 (such as nanopublications and micropublications), level 2 (such as a formal pathway in, for instance, WikiPathways), level 3 (any system biology qualitative model) and level 4 (for instance well documented and reusable workflows).

All concepts in a nanopublication graph should ideally refer to a well established vocabulary, so that linking to ontological knowledge is possible and computers understand exactly what is meant by each URI in the graph. It is important that nanopublications can also be expressed in a human readable language, based on a correct linking of the URIs in the graph and the terms used in a narrative (in different languages) of the concepts in question. Next, the nanopublication needs provenance to be placed in context. Not only minimal information about the conditions under which the assertion has emerged and those under which it is considered "true", but also all other metadata that are usually associated with a classical narrative research article (such as authors, publisher, etc.) should be associated with the nanopublication. In fully compliant nanopublications, these parts of the connected graphs forming the entire nanopublication are also modeled in RDF (see guidelines and examples at [26]).

Nanopublications have now been created from different data types, such as locus specific databases [27], the Fantom 5 data set, GWAS data [28], chemistry and pharmaceutical databases [29], UniProt and neXtProt. In principle, each data source containing assertional information can be republished in this format, which is both machine interoperable and human readable, with relatively limited effort, without distorting neither the original data format nor the legacy database. Nanopublications can also be re-created by text mining, although they suffer from the same challenges as all text mining approaches [30]. An increasing number of narrative sources (including PubMed) are now being "nanopublished", i.e. published as nanopublications, and major international publishers are investigating how they can expose the scientific conclusions and evidences contained in their narrative collections in this format.

Data curation: from ignoring to cooperative

Many of us search or browse bioinformatics online resources. While doing so, we occasionally activate a link that unexpectedly breaks or points to absurd content. We mentally complain about it, but we usually ignore it and resume browsing. Some of us do spend the couple of minutes necessary to report this broken or mistaken link to the development team (if still in existence) and thereby spare the trouble to other users. In this case, "contributing" is pointing out errors but not solving issues. Indeed, too few of us envisage on-line resources as community wealth to which contribution would mean definite improvement and added value, i.e. a form of curation, which benefits all.

The activity of biological data curation has evolved over the years to a point where there is now an organised International Society of Biocuration [35] within which the question of community-based curation is debated and promoted, among other themes. The need for a coordinated action in this domain was emphasised for instance when the Swiss-Prot team introduced in 2007 the "adopt a protein" scheme [36], encouraging specialists of a given protein to oversee the update of the corresponding UniProtKB/Swiss-Prot entry. As it seems, scientists are not born protein adopters and the initiative could not be sustained. In the same period, a more sophisticated wiki based attempt was made in WikiProteins: the paper calling for a million minds [Mons 2008] from 2008 has meanwhile collected more than 120 citations, but is not in line with the number of community annotations in WikiProteins, and the attempt was discontinued. Other wiki-based approaches, such as the GeneWiki (in the context of WikiPedia) [32] and for instance WikiPathways [33], have met with slightly more traffic, but to the best of our knowledge the only community annotation effort that really took off to a level of satisfaction is ChemSpider [37]. However, with these lessons learned, an alternative invitation to contribute was devised a few years later in the human protein-centric knowledge platform neXtProt [38]. The neXtProt scheme promotes users' participation through the specific input of a selected network of specialists. Experts contribute by submitting experimental data sets and defining metrics for quality filtering in agreement with the neXtProt team. Very recently, curated associations on, for instance, Post Translational Modifications from NextProt have been formatted as nanopublications; this will allow the community contributions to certain snippets of information to be fully recognized (see next section).

In essence, biological data curation history tends to show that direct contribution may not be the ultimate strategy for gathering quality information and attracting potential contributors when it is limited to the addition of comments or facts in a Web page. Instead, guided input, so as to capture and shape information upon criteria that were previously and collectively agreed upon, seems more of a realistic approach. Future tools should rely on social interfaces encouraging users' cooperation in a constructive and targeted manner. Some efforts have already been made in this direction, e.g. for collaborative ontology development [39] and for interactive knowledge capture by means of Semantic Web technologies [40]. Yet, this important future area of scientific contribution suffers from the same roadblock as the "data-based science" discussed before. Unless a culture develops where these contributions (when measured perfectly) influence the career of the next generation of scientists, community contributions will always be limited to the "altruistic few" [41].

Exchange, access, provenance and reward models

An outstanding issue is the social award system and the perceptions prevailing around data sharing. In white papers advocating data sharing, authors usually emphasise technical challenges rather than the actual process of data sharing, although there are as many social challenges associated with actual data sharing as there are technical challenges. Obviously, technical challenges come first: data can only be shared if they are interoperable in format or have been captured with proper metadata attached.

Making data Open Access is clearly not enough; data accessibility and reusability by others than the data generators, is what really matters. As stated in previous sections, reuse of valuable data sets will support e-science discovery processes. In this context, provenance is the key for users planning to include existing data in a meta-analysis. Prior to adding a data set to the analysis mix, an e-scientist needs to evaluate the set, its overall relevance, quality and the underlying methods. For this crucial decision step, the metadata, including rich provenance, are needed.

In many cases, data can be excluded or included from/in an analysis workflow by properly instructed machines. For instance, all data on genes of a given species, e.g. mouse, can be automatically discarded, as long as sufficient provenance is associated with each candidate data set. It is thus very important that the concept "mouse" as Mus musculus is associated with a data set based entirely on mouse experiments, and properly referred to with a computer readable identifier. But it is also important that such identifier is at the appropriate position in the metadata fields, or in a RDF graph; this, for instance, to allow for the distinction between an occasional mention of the concept Mus musculus in a table or graph, as opposed to the statement "this entire set was generated on 'mouse' experiments". However, this ideal situation is currently far from reality. Even with Digital Object Identifiers (DOIs) for data sets and initiatives such as FigShare [42], Dryad [43] and the Research Data Alliance [44], we will need many years before each valuable data set can be properly judged and interpreted by others than its creators. This becomes even more pertinent in multi-scale modeling and the associated multi-omics and multi-technology data sets that increasingly dominate contemporary biology. It is not enough to "find a data set of potential relevance", because soon there will be too many, or to see some metadata on how the study was performed, although this is a conditio sine qua non. For real e-science approaches in biology, we need to see the provenance of each individual data element as it may appear, for instance, as a crucial edge in a graph-based hypothetical discovery interface.

Nanopublications and micropublications are here important. It is clear that entire complex data sets with hundreds of thousands, and sometimes millions, of interesting associations can be published as nanopublications; so, they are no longer lost in a huge number of hyperlinks to remote repositories, but each and every individual association becomes a research object in its own right, and is discoverable by computers and humans alike, all across the Web.

Assuming nanopublications and micropublications can solve the data integration issue, the next challenge is to "build a market" for such small information units. They do not intrinsically carry enough evidence to be trusted. The decision to "trust" them or not is taken on the basis of the source information, the associated methods and the reasoning that led to the claim in question. Again, additional steps (in fact, a form of annotation) are needed to create such computer and human readable units with rich enough provenance. This raises the question of finding the means of improving scientists' motivation to spend extra time on annotation.

Naturally, we lack much of the technical infrastructure that is needed to make this all reality, but, in practical sense, these needs are "easier" to fulfil than breaking through the science ecosystem hurdle to make "preparation for sharing" of data a core activity for every data creator and publisher. In fact, what we need is "desktop publishing" of data and information, very much like today authors carefully pre-format their papers according to guidelines for authors. Modern publishers should become data publishers, as well as narrative publishers, and should assist scientists in the curation and shaping of newly published data sets and of their provenance, much the same as they currently do for narrative.

Especially for sensitive data, both in terms of privacy and competitiveness, a trusted party status for the needed data publishing and stewardship infrastructure is a conditio sine qua non. Therefore, such a data exchange environment can only be built effectively as a federated and "approved" infrastructure, serving national as well as international data driven projects, and as a public-private partnership.

For purposes of clarity, in Figure 2 we have summarized, and grossly oversimplified, the basic workflow of data driven science. What is needed for e-science is, in fact, a completely new way of publishing, using, searching and reasoning with massive data output, in an open, software-driven, interactive environment.

Relevant scientific data, such as open source publications (e.g. Public Library of Science (PLoS) or BioMed Central (BMC)), individual assertions from closed access publications, abstracts (e.g. PubMed) and relevant legacy data sources (e.g. ChEMBL, UniProt), that constitute a central core of biological information requested by almost all domains, should be made available in an interoperable format to make their direct integration, comparison and modeling with new data possible (Figure 3). Currently, only a small percentage of information in databases, for instance SNP-phenotype associations, can be recovered by text mining from abstracts, or even the entire narrative part, of full text articles. Many of these types of associations are included in tables and figures, which escape ordinary text mining algorithms, and in supplementary data, which are ignored by text mining. It is therefore crucial to move to a situation where massive numbers of associations can be published in a "discoverable" and interoperable format, with proper references to the producers of the data and associated narrative elements in order to allow award of the efforts.

Notably, publishers have already played a role by imposing annotated data submission and, some of them, by being involved in the definition of related standards, e.g., MINSEQE [45] or MIAPE [46] standards that are governing corresponding data repositories: ArrayExpress [47] for MINSEQE or PRIDE [48] for MIAPE. However, a major roadblock at this point in time is that many grant and manuscript reviewers still do not recognise the value of studies that do not entail the production of new experimental data, but only exploit results from data repositories. Without challenging the sustained importance of proving a biological hypothesis with sound experimental data, it should nonetheless be admitted that validation does not necessarily impose being the creator of the data used as evidence.

Finally, social hurdles for data sharing are not limited to the conservatism of publishers and funders, which could be overcome hopefully soon. Additionally and more importantly, there is no "scientific" reward for sharing, i.e. acknowledgement of its value as a scientific product. If no mechanisms exist for any generally acknowledged reward for sharing and making own data discoverable, well annotated, principally interoperable and citable, a routine of data sharing is not likely to be established. Movements like Altmetrics [49] are crucial to raise a discussion and to demonstrate technical feasibility of a fine grained judgement about an individual's contribution to a scientific record. However, until the "reward" reaches a steady and wide acceptance by reviewers, funding bodies and publishers, nothing will change. They have the means to push researchers make a proper data stewardship part of their natural workflow. It is only since recently that we need to take the "reusability" of the data that are being generated into account in the study design. Several funders already require a well-drafted data stewardship plan for any proposal that will generate significant data sets. This practice should be encouraged; proper standards, best practices, guidelines and reward systems should be implemented and made easily findable, so that biologists with little or no affinity with bioinformatics or data sharing can still participate. Only if all these prerequisites for data sharing are in place, the culture may change and a genuine open data exchange culture in the life sciences can be established.