Biomedical data management is increasingly complex due to the variety of storage systems and evolving data models. This heterogeneity presents obstacles to data integration and querying, crucial for advancing biomedical research and healthcare. A possible solution is to employ a recent idea in the database field: polystores. A polystore is a DBMS designed to integrate and manage multiple heterogeneous data stores, allowing for eficient data processing and querying across diverse data models and storage systems. Nevertheless, polystore systems may difer profoundly one with the other, both in their structure and in the interface provided to the users: this is usually due to the diverse landscapes where polystores may be applied, and thus the diverse nature of data available in diferent fields, and to the information needs that users may have. These limitations impede the adoption of existing polystores in the context of biomedical data. Moreover, as far of our knowledge, there exist no system ofering an integrated viewpoint to biomedical data by means of a graph data model, which would instead provide a sharpened representation of this domain. In this paper, we outline the research challenges and the initial steps towards the development of a polystore system that provides eficient access to multiple heterogeneous biomedical data sources, addressing also critical privacy concerns by tracing data oflw and ensuring the privacy and anonymity of individuals.
Managing biomedical data poses significant challenges. Practitioners are confronted with numerous disparate storage systems, each employing diferent data models that may evolve over time. This leads to a high degree of heterogeneity, where the same data domain can be represented using various models such as relational, hierarchical or graph-based. In particular, graphs are highly used for biomedical data, because their structure naturally mirrors the complex and interconnected relationships found in biological systems. Moreover, these systems are often managed by diferent Database Management Systems (DBMSs). Additionally, the data is described by diverse metadata schemas, leading to more heterogeneity. This low level of integration hinders the possibility of querying them together and infer new knowledge, unless experts manually integrate them, by defining a unified data model, achieving consensus among data stakeholders, manually matching existing data to this new model, migrating data accordingly, and then modifying applications to adapt to changes in the used query language. All these steps are time-consuming and expensive.
At the European level, numerous contexts exist where the aforementioned challenges are relevant. Specifically, the Department of Information Engineering at the University of Padova leads the EU project HEREDITARY, which involves collaboration with three medical centers, encompassing heterogeneous multimodal clinical and genomic data, which require integration. HEREDITARY underscores the pressing demand for a robust and eficient system capable of seamlessly integrating diverse biomedical datasets.
A possible solution is to employ a recent advancement in the database field: polystores. A polystore is a system that integrates multiple heterogeneous databases (SQL, NoSQL, graph, document stores, etc.) under a unified query interface. It enables querying data across diferent types of databases without physically moving or transforming the data. Polystores uses a query federation approach, where queries are translated and executed across diferent underlying databases, leveraging their native query engines.
The state of the art is represented by BigDAWG [
Considering other polystore solutions, among the most interesting and actively maintained academic
projects is Polypheny [
Another relevant aspect is how integrated data is modeled. As mentioned before, a graph-based
data model is necessary to represent the intricate pattern of such a complex domain: this translates
into the need of an ontology through which data is accessed. The Ontology-Based Data Access
(OBDA) paradigm [
This paper’s sections are organized as follows: section 2 and 3 describe in details both the approach and the proposed framework, analyzing how its structures allows to tackle our challenges. Section 4 derives instead open research questions coming from the framework weaknesses, and how we plan to address them; section 5 concludes the paper.
Recent advancements in the research proposed a novel approach to polystores: the Ontology-Based Data
Federation (OBDF) paradigm [
implicit results. In many OBDA implementations (e.g. Ontop [
Identifying functional requirements is essential to shape the research design, workflow, and system architecture. These requirements are fundamental to shape the system architecture and to drive the choice of the diferent components. To address them, we explore the detailed design components essential for setting up a system that meets the project objectives. In Figure 3, we show how diferent components plug together: the SPARQL endpoint, the OBDA Data Integration platform (with ontology and mappings) and the Data Virtualization and Federation layer. In this section we will inspect each of these layers from a bottom-up perspective and we will discuss the implementation choices. There exist diferent solutions implementing a VDB such as Denodo 2, which although it is very robust, is an enterprise solution. An alternative is Dremio3, which supports several sources, and allows for the development of custom adapters for unsupported sources. Considering that Dremio is open-source, we will focus on it as our federation component. Going into details, Dremio is a VDB with several features: it is as a data virtualization system that enables seamless access to data across various storage systems like RDBMSs, NoSQL databases and cloud storage without duplicating data, thereby optimizing query performance. The platform also includes a user management system that allows administrators to control access and manage permissions efectively, ensuring that data is accessible only to authorized personnel and facilitating collaboration. Additionally, Dremio incorporates a robust logging system, 2https://www.denodo.com/en 3https://www.dremio.com/
MAIN SERVER
OBDA DATA FEDERATION
HERO Genomics MAPPINGS
Network
Relational Columnoriented
Hierarchical which is essential for recording user activities and troubleshooting. This logging capability helps in analyzing usage patterns and tracking data flow.
Diferent OBDA systems have been implemented within time to support SPARQL-to-SQL translations.
However, recent studies focused on comparing performances between Mastro and Ontop, as they are
among the few OBDA systems which support ontology reasoning. In general, Mastro shown faster
responses in scenarios [
Between these, an ontology must act as a shared dictionary. As a first step towards the integration of
heterogeneous data, the HEREDITARY consortium provided the genomic component of the HEReditary
Ontology (HERO) [
Considering the connection points between HEREDITARY challenges and OBDF key features, our plan is to adopt it as the backbone of our polystore. Nevertheless, this does not come without challenges. In this section we present preliminary results from an initial implementation of an OBDF-based polystore system, focusing on bottlenecks and hurdles we encountered in the process we aim to tackle.
Query Optimization We set up a federated environment with two data sources hosting genomic data, and we federate them in an OBDF setup, choosing Ontop as the OBDA layer, and Dremio as the Data Federation layer. Then, we derived 4 query of interest in the domain of genomics, in particular mimicking those provided by Beacon V24 for variants discovery. We ran each of these multiple times, recording timings for each execution phase, and we then computed average and standard deviation. Figure 4 showcases a portion of our results. As it is instantly evident, there is a huge bottleneck in the last phase, namely the Dremio Running phase. What we observed is that queries that come out from the OBDA layer present an unnecessary amount of self joins, that could possibly collapse into a single interrogation, grouped by identical queries. Moreover, cross join between federated sources are performed, retrieving uselessly the whole databases from the sources multiple times, often risking
to fill the available memory. This occurs because OBDA systems have traditionally been designed for
single-source scenarios and have not been optimized for federated settings. Recently, new research
advancements [
Mapping Accuracy How mappings between the ontology and the VDB are defined is paramount. Although automatic mapping bootstrapping tools are available, they usually operate in a classical OBDA scenario. Hence, in a federated context, especially in a sophisticated domain such as the biomedical one, it is essential to have an human agent manually defining them. This process is inherently error-prone, and thus it will be necessary to put into action accuracy estimation techniques. Recent advancements in the research [11] proposed a novel approach to estimate KG accuracy using sampling strategies and providing estimates within confidence intervals. We plan to extend these techniques also in the domain of Virtual Knowledge Graphs, to assess their accuracy with respect to the original data sources, and thus to determine mappings quality.
System Accessibility The aim of having a Federated Analytics infrastructure is providing to project partners all the tools and the services to get access to the data in a streamlined way. This means that on top of the architecture, an user-friendly interface should be available to the audience. We plan to do this by means of a web application. This application will not only ease data retrieval, but will allow to export query results in many formats. Additionally, a comprehensive logging system will monitor data access, contributing to explainability. As a preliminary result, we efectively set up a web portal allowing to interact with the aforementioned embryonal OBDF-based setup5 for genomic variant discovery.
In this paper, we have discussed the challenges posed by the management and integration of heterogeneous biomedical data. We have critically evaluated existing polystore and OBDA solutions, highlighting their limitations in meeting the complex requirements of the biomedical domain. Consequently, we proposed adopting the OBDF approach, combining OBDA techniques and Data Federation technologies. We described an initial implementation leveraging Dremio as a federated virtualization layer and Ontop as the semantic integration engine, showcasing preliminary results within the genomic domain. Our analysis identified critical performance bottlenecks, inaccuracies in ontology-data mappings, and the need for improved system accessibility. In future work, we aim to address the identified challenges. Specifically, we will investigate query optimization techniques to reduce computational bottlenecks in federated query execution, develop methods to rigorously estimate and improve mapping accuracy, and design intuitive user interfaces complemented by logging and privacy-preserving mechanisms. By addressing these aspects, we will significantly advance toward an eficient and accessible polystore solution tailored for biomedical research requirements.
This project has received funding from the HEREDITARY Project, as part of the European Union’s Horizon Europe research and innovation programme under grant agreement No GA 101137074.
Rules and Reasoning, Bucharest, Romania, September 16-18, 2024, volume 3816 of CEUR Workshop Proceedings, CEUR-WS.org, 2024. URL: https://ceur-ws.org/Vol-3816/paper73.pdf. [11] S. Marchesin, G. Silvello, Eficient and reliable estimation of knowledge graph accuracy, Proc.
VLDB Endow. 17 (2024) 2392–2404. URL: https://www.vldb.org/pvldb/vol17/p2392-marchesin.pdf. doi:10.14778/3665844.3665865.