This paper presents insights gained utilizing semantic technologies as a FAIR implementation profile for clinical metadata standards at Roche Pharma Product Development Data Sciences. It highlights the journey undertaken to democratize FAIR data authoring and achieve higher-order FAIRification all while addressing the associated challenges. The paper includes a specific use case from the clinical data collection standards domain, illustrating the automation potential of semantic model-driven development for implementing FAIR user-centric applications. The authors aim to share valuable lessons learned and challenges faced, providing insights for organizations seeking to advance FAIR practices and transform their data landscapes in similar endeavors.
Semantic Web Technologies, with their approach to linked metadata, ofer substantial potential
as a fit-for-purpose FAIR implementation profile [
Breaking that barrier and unlocking the potential of semantic technologies as a FAIR implementation profile was an objective we have set for ourselves over the last couple of years as we embarked on a journey of building a semantic-model-driven development framework for • Democratizing the FAIR data authoring process by embedding semantic applications in user-centric operational contexts. • Aiming for higher-order FAIRification; not merely retrofitting metadata but leveraging semantic inference and automation to proactively incorporate FAIR considerations into the design and development of data systems and applications.
In this paper, we discuss our experience developing authoring applications based on ontologies. These applications are smoothly incorporated into the everyday workflows of clinical standards subject matter experts with limited or no background in semantics. We provide an overview of what we denote by semantic model-driven development, highlighting challenges and specificities. Additionally, using a use case from the clinical data collection standards domain, we demonstrate how the automation capabilities brought by semantic inference facilitate higher-order FAIRification by accelerating the creation of FAIR user-centric applications.
Our starting point was an existing metadata repository known at Roche as the Global Data Standards Repository (GDSR). GDSR is Roche’s homegrown multi-tenant metadata repository for clinical data standards, be it the industry-wide standards such as the standards from the Clinical Data Interchange Standards Consortium (CDISC) or any Roche-specific data standards. The metadata standards in GDSR are expressed as RDF knowledge graphs, stored in a triple store, cataloged, versioned and then served to users via a web browser and to applications downstream via a ReST API.
The GDSR system architecture inherently adhered to the FAIR principles in its metadata
content [
At this point, It was clear that reaping the full FAIRification benefits GDSR ofered was predicated on having the right native RDF authoring and governance tools in the hands of its content owners. This is how a key component in the GDSR ecosystem, the GDSR Workbench, was born.
Our decision to pursue a model-driven approach stemmed from the need for a systematic and eficient development process for the envisioned GDSR Workbench. One that leverages automation not only to cover existing data standards models but also to accommodate future ones, ensuring adaptability and scalability.
Similarly to model-driven development in an object-oriented paradigm [
The development process (Figure 1) for a typical FAIR authoring application using the GDSR
Workbench framework consists of:
1. If the underlying ontology is not in place already, an ontology creation phase where an
ontologist works closely with a domain expert on a set of competency-based questions to
define the ontology as a formal machine-understandable representation of the domain of
interest.
2. An artifact configuration phase that covers:
• Writing SPARQL queries as a catalog of parameterized named queries. The queries,
with metadata annotations, are then used for populating the various user interface
views, for quality control and report generation or for downstream consumption.
• Writing SHACL shapes on top of the ontology classes exposed to the end users. The
shapes are used for validating the graph data mutations according to the ontology
definitions and for automatically generating user interface views.
• Adding the necessary URI redirect rules in Roche’s identifiers service for domain
registry and resource dereferenceability.
built at a time when the more standard GraphQL mutations did not exist yet [
While adopting a semantic model-driven approach for developing FAIR authoring applications introduced numerous automation possibilities and a systematic means to standardize editing interfaces, it also came with its own set of hurdles and challenges. In the following sections, we highlight some of the noteworthy ones and the solution approach we have taken to tackle them.
SHACL shapes provided us with a very good base to describe the diferent authoring forms
and editing constraints an object is subject to but we still encountered some expressiveness
limitations when exclusively relying on SHACL constraints semantics [
To fully automate the user interface generation process starting from the SHACL shapes, more expressive power was needed to complement the shape definition with the additional constraints to dictate aspects such as these in the aforementioned questions. It was for that reason that we defined our own Roche extension to SHACL namespace. An example SHACL shape with some Roche-specific extensions is shown in Listing 1. 1 @prefix schema: <https://id.roche.com/example-domain/schema/> . 2 @prefix gsh: <https://id.roche.com/gdsr-shacl/> . 3 # Other standard prefixes are omitted 4 5 :ExampleShape 6 a sh:NodeShape ; 7 sh:targetClass schema:ExampleClass ; 8 # a gsh extension to determine if resources of the target class are mutable 9 gsh:editable true ; 10 # a gsh extension to specifiy the graph in which permutations will be saved 11 gsh:targetGraph <https://id.roche.com/example-domain/example-graph> ; 12 # a gsh extension prob. to determine if resources should be indexed for full text search 13 gsh:searchable true; 14 # a gsh extension to calculate user-friendly dynamic label expressions 15 gsh:labelExpression [gsh:concat([sh:path schema:propA] [sh:path schema:propB])] ; 1CRUD is a commonly used acronym in software development that stands for Create, Read, Update, Delete.
Listing 1: An example of a SHACL shape definition with additional Roche-specific extensions
In some cases, the expressiveness gap manifested when the standard SHACL shape definition resulted in a non-deterministic translation of the user transaction into a graph mutation. An example we can give here is the standard sh:alternativePath used in SHACL complex property paths. To avoid such non-deterministic scenarios, we equipped the GDSR Workbench engine with the necessary shape validation rules to report faulty shape configurations (in the example here, an alternative path used in a mutable property definition).
Like other languages in the Semantic Web sphere, SHACL is inherently declarative, emphasizing the ’what’ over the ’how’. This serves the purpose of knowledge representation and validation perfectly: declaratively describe the problem domain, its entities, their relationships and what constraints apply to a given solution. The systematic procedural logic of how the solution is found is left to OWL reasoners, SPARQL query execution engines, SHACL inference engines, etc. In our case, however, adopting a pure declarative approach would have fallen short of the expectation. In many cases, we were faced with situations where a user action in the interface should trigger a series of mutations to be handled in a transaction where synchronous system-to-system communication was involved. In these scenarios, and many similar ones, we needed to have more imperative control over the how, and when, not just the what.
For that purpose, we introduced the concept of an action shape, a subclass of the standard SHACL shape but with the additional semantics of having an action component attached to it. The procedural logic behind the action component itself can be written using any programming language of choice as long as it is encapsulated and reachable via an HTTP(s) endpoint. A simplified example of action shape to notify a remote system of an event (e.g.: a certain mutation to a resource) is shown in Listing 2 below.
When we started configuring the authoring interfaces using SHACL shapes, SHACL’s closed-world assumption was just what we needed to validate transactions on top of linked data in an operational context. We relied on it to define custom uniqueness constraints and rules applicable for cascade deletion. As we covered more and more configuration scenarios, however, we realized that the closed-world assumption was too restrictive in certain cases where the related linked resource resided in a diferent “world” such as a neighboring integrated system or an archived version of the metadata. After careful analysis of the diferent false positives SHACL violations we collected over time, we finally adopted a hybrid approach to validation. By default, the close-world assumption prevails and in the cases where falling back to an open-world assumption was needed, we declared it in the object property in question using a dedicated predicate from Roche’s SHACL extension as shown in Listing 3.
Listing 3: An example of a shape definition where the closed-world assumption is amended for remote properties.
The gsh:location here tells the workbench engine that the linked resource resides in a remote location. Depending on the relationship type, that remote location can be dynamically resolved at runtime so that the applicable cascade validation rules are enforced.
To ofer a more tangible perspective on the opportunities explored in this paper, we present a use case from the clinical data standards domain. The use case scenario presented below, along with the subsequent solution sections, should illustrate the automation potential of our semantic model-driven development framework and how it helps achieve higher-order FAIRification by accelerating the creation of FAIR user-centric applications.
Like every step in any clinical trial, data collection is a process that lends itself to harmonization. At Roche, collected data should conform to collection standards that are based on the CDISCcontrolled terminologies or their Roche sponsor extension. This means that every variable’s permissible value is bound to fall within pre-established enumerated value domains. The controlled terminologies and the underlying schemas can vary depending on whether the variables are collected as part of a questionnaire in an electronic Case Report Form (eCRF) or as one of the multitude of other non-CRF data structures such as lab assays, digital measures, CT-scans, ophthalmologic grading parameters, etc.
The majority Roche’s metadata standards are largely aligned with the existing CDISC standards but when it comes to data collection, Roche maintains its own set of global metadata standards that covers a wide range of therapeutic areas and collected data formats2.
While dictating the permissible values for a given non-CRF variable in isolation from others is readily achievable using the existing standard models, dictating the permissible combinations of values for variables collected together poses a more significant challenge. It requires maintaining Value Level Metadata (VLM) according to schemas that vary dynamically following the diferent data collection settings and the applicable version of the standards at a given point in time.
Given the current scope of Roche’s non-CRF standards, populating and maintaining dozens of diferent VLM schemas seemed infeasible. Before the introduction of the semantic modeldriven framework, and despite the FAIR metadata representation of the underlying collection variables and controlled terminologies, maintaining combinations for all VLM schemas in a machine-readable manner required a dedicated solution for each schema. Defining the schemas themselves in a semantic form was relatively straightforward, but the challenge lay in maintaining permissible combinations as populated knowledge graphs. We were faced with the choice of either having to provide the data standards managers with new custom-built authoring tools for each VLM schema; or having to handle the population of all VLM knowledge graphs on our end. Neither option justified the resource investment. Consequently, only three selected VLM schemas were maintained in the metadata repository, while all other schemas were manually kept in a tabular sheet, disconnected from the FAIR source, posing the risk of possibly outdated copies circulating freely and non-compliance issues encountered downstream. 2Roche’s eCRF and non-CRF standard models were established before the CDASH collection standards counterparts from CDISC were publicly released.
Leveraging the semantic model-driven framework we have progressively implemented, and
drawing inspiration from concepts in the meta-programming paradigm [
Right from the start, we took an abstract perspective on the problem of modeling the diferent
non-CRF VLM schemas for diferent therapeutic areas and non-CRF standard models. Instead
of developing a new ontology for diferent occurrences of the VLM problem, we developed a
more abstract meta-VLM ontology. Similar to how a meta-program possesses knowledge of
itself and can manipulate itself [
Every time the data manager declared a new VLM schema in the user interface, the VLM ontology automatically expanded with the help of semantic inference, creating new classes and properties. For these classes and properties just-created, a few more inference rules automated the artifacts generation process including the SPARQL and SHACL required for creating the GDSR editor user interfaces. The resulting end-to-end automated pipeline for the artifact development process is shown in Figure 3.
To address the legacy VLM combinations maintained in tabular sheets, we implemented a separate dedicated parser component that dynamically understood the VLM schemas created. This component performed a series Figure 3: The end-to-end automated pipeline to of reverse look-up operations to the metadata create VLM authoring interfaces using repository, translating from free-text values the GDSR Workbench framework into the URIs of the linked entities.
Having undergone the automated pipeline outlined above, the legacy tabular content is now fully converted to machine-readable metadata. The resulting knowledge graph is appended to the overarching clinical standards knowledge graph in GDSR where the semantic link to the non-CRF standard variables and controlled terminologies is restored. This not only enforces data integrity and opens new possibilities for standards compliance assessment but also serves as a significant eficiency booster and time-saver for the data standards manager maintaining the VLM combinations. It practically eliminates the need to manually propagate changes to the legacy tabular VLM sheet following an update to the standard source, thereby eliminating the risk of the VLM not being aligned with the metadata repository.
The insights presented in this paper were obtained working hands-on in the specific context of
Roche’s metadata repository and its satellite projects. Rather than being research-oriented in
nature, our eforts were geared toward tackling the hurdles and bottlenecks faced scaling our
metadata FAIRification initiatives across the diferent areas of Roche’s clinical trial development.
That being said, we consider our work to be largely aligned with the viewpoint of the research
community and that of our industry peers; especially those underlining the importance of
adhering to the FAIR principles and the role Semantic Technologies play in there [
On a more technical note, a thorough review of the literature is yet to be conducted but to our
knowledge, most of the work on augmenting SHACL shapes with procedural logic is attributed
to Knublauch et al. at TopQuadrant, Inc [
This paper documented our eforts to harness semantic technologies as a FAIR implementation profile for metadata standards applications in life sciences organizations. We stressed the importance of proactively incorporating FAIR considerations into the design and development of the FAIR data systems. Using a real-world use case scenario, we illustrated how we leveraged automation to build native FAIR authoring tools seamlessly embedded in user-centric contexts. By sharing the challenges we encountered, we aim to provide some insights for other organizations pursuing similar eforts to advance FAIR practices and transform their data landscapes.
We appreciate the lack of open access to the work behind the insights shared in this paper. Open-sourcing a selection of our standard ontologies and SHACL namespace extension is potentially possible but pending a feasibility assessment due to known inter-dependencies with other components of the GDSR ecosystem. Nonetheless, a more elaborate exchange of ideas and experiences with the community remains a privilege we are always after and we encourage interested readers to reach out via email with any feedback or questions.
The insights shared in this paper are the outcome of the long and dedicated eforts of a team comprising semantic experts, information architects, software and QA engineers. We express our sincere gratitude to all those who have contributed to and supported this endeavor. In particular, we’d like to thank the following key contributors for adding to the GDSR story over the years:
Adam Kozac , Celso Pereira , Huw Mason , Ivan Robinson , Jacek Caban , Javier Fernandez , Marcelina Kasprusz, Michal Kalinowski , Michal Pietrusinki , Nelia Lasierra and Szymon Tyka.
We equally appreciate the great support of our sponsors and the continued fruitful collaboration with our key stakeholders. In particular, we’d like to call out the whole Data Standards and Governance Group at Roche Pharma division.