Highlights from the workshop “Unlocking the potential of data analytics in collaborative research”
Data availability, transparency, reliability, and management for bio-based innovation
On 4th June 2026, the BIORING, BIOSAFIRE, BioPhenom, DESIDERATA, SuperBark, SurfToGreen, and DATA4CIRC projects successfully hosted an online workshop focused on data analytics and predictive modelling. The event brought together over 40 participants, including researchers and industry representatives, data providers and end users, to exchange experiences and insights from various EU-funded projects.
Objectives
The workshop aimed to promote awareness and to share views and experiences on data analysis from both researchers, and data management experts and practitioners. After a brief introduction from Andrea Porcari (AIRI – Italian Association for Industrial Research) of BIORING, Claudia Mabile (Process Design Centre – PDC) from BIOSAFIRE presented an overview of the gap analysis on data management and predictive modelling.
Then, experts dived into the most relevant topics: Tassos Papadiamantis (Entelos Institute and DESIDERATA) talked about responsible data and AI in cooperative research, Barry Hardy (Edelweiss Connect) presented how evidence-based workflows support product innovation, development and marketing, Vince Hop (PDC and BIORING) explored the potential of data analytics and modelling tools for cooperative projects, and Álvaro Ojeda (Idener and Data4Circ) focused on specific data gaps while producing models.
The workshop ended with an open discussion among all participants, split into four groups to facilitate sharing of feedback and comments on the most relevant aspects emerged from the presentations.
Watch the plenary session here
Highlights
Claudia Amabile presented the relevance of gaps, challenges, and opportunities in data analytics in collaborative research: collaborative research generates large amounts of raw, unprocessed and non-documented data, difficult for other users to access, share, and reuse or interpret it.
A proper data management is the pillar helping to overcome this, indicating who uses and can access the data, ensuring data quality, accuracy, completeness, consistency, and reusability, in compliance with relevant regulations or FAIR principles. Another pillar is predictive modelling, which allows to see the innovation process as a workflow from data generating to impact. These two pillars increase data value and knowledge transfer: firstly, creating, collecting, documenting, storing, and sharing data; then, feeding it into models (statistical, AI, or machine learning) to translate it into insights and to help informing decisions. The highest the quality of generated data, the most valuable the return information. Translating this process into practice faces some challenges, though.
Technical failures are common obstacles in data management, like the negative perception of gathering and reporting data while creating them, the lack of understanding of the correct terminologies, the variety of vocabulary, formats, and software. In addition, not-well documented data (e.g. without an data-update tracking system) or lack of FAIR guardrails. This results in less interoperable data, hindering data analysis.
The most common obstacles are the governance failures of lack of resources, process and planning allocated to data management, as well as data inconsistencies that makes the process modelling results mathematically meaningless: biassed data set will result in biased results, especially when generative models could make hidden assumptions to fill data gaps. To properly set-up, monitor, and understand modelling results, experts in the field are required.
FAIR foundations for data management and AI
Tassos Papadiamantis outlined the bases for an effective data governance, which is essential for trustworthy and reusable scientific research, particularly in multi-partner projects involving diverse disciplines and data environments. Data have a lifecycle, and with several steps of data generation, use, and modification, everything is interconnected: differences in terminology, formats, metadata, tools, and tools licensing can hinder data integration and reuse, while inadequate governance reduces traceability, trust, and value of research outputs. Specific metadata are required to create a record of the entire data lifecycle, so that the user next in the chain will understand what the data is.
Thus, the FAIR principles provide the required practical framework and common understanding, but remaining a guidance and not a standard. Such principles, originated for data themselves, nowadays need to be extended to research software and workflows, especially when AI is involved. Their successful implementation requires clear governance, defined roles and responsibilities, comprehensive metadata, origin, and user-friendly tools integrated from the beginning of the research lifecycle.
This extension of FAIR principles has been one of the key outcomes of the DESIDERATA project, demonstrated through expert validation, predefined metadata schemes, automated data harmonisation, and machine-actionable metadata publication, while keeping data confidentiality. Overall, a FAIR-by-design approach can strengthen data quality, interoperability, transparency, reuse, and confidence in scientific results and AI tools.
Evidence-based workflows to further trust, innovation, and regulatory acceptance
Barry Hardy, from the experience of several collaborative research initiatives and projects (like EU-ToxRisk, RISK-HUNT3R, BioPhenom, and SSbD4Chem), deepened how effective scientific and industrial decision-making requires the integration of evidence-based data, documented workflows, and human judgement. Data alone is insufficient: document the thinking behind the experiments, and the approach in developing the models and assessing uncertainty and conflicting evidence are fundamental activities to ensure understanding from other stakeholders. Scientific knowledge infrastructures should adopt FAIR principles, harmonised metadata, interoperability, sustainable repositories, and well-documented, reproducible workflows.
AI can enhance these infrastructures by supporting documentation, data organisation, evidence retrieval, knowledge graphs, and workflow development; however, it should be computational, a complement to the main human actions like interpretation, judgement, discussion, and complex decision-making.
The objective is to have clear data that can be used by end-user tools and software, but also frameworks and workflows that can be adapted to different sectors, to allow evidence-based innovation. Moreover, this increasing evidence-based generation approach can strengthen regulatory acceptance of new risk-assessment and SSbD frameworks. This is how we’ll maintain curiosity, collaboration, and a trustworthy knowledge.
Modelling can make experimentation more efficient, saving time and costs
Vince Hop brought the vast PDC experience, also applied in BIORING, explaining how modelling supports process development. The aim is to improve the individual process steps in a safe and sustainable way by applying data analytics and digital technologies, while also gaining a cost-effective outcome.
Process development requires a lot of information and selection criteria, usually from cheap literature review (that can though be inconsistent, due to contradicting articles or different methods) and reliable lab experiments (which though are time- and cost-consuming). Modelling is a time-saving alternative, but it’s mostly assumption-based and requires some software (often expensive) or coding expertise. However, there are modelling experts with verified databases that can cooperate with researchers. Among the several modelling options, like statistical, supervised learning, unsupervised learning, forecasting, and rule-based modelling, the last one if implemented properly is very reliable; PDC applies it, with a proprietary software based on decision trees and large language models.
Reliable process design requires the appropriate combination of quality data, suitable models, validation, and qualified expertise. Whichever modelling is chosen, it requires a final experimental validation, but guide towards the most relevant experiments and efficient process design.
Tool and approaches to fill data gaps
Álvaro Ojeda Roldán presented the approach and tools Data4Circ applied to face three data gaps in different fields to produce models: data scarcity, data reporting, and limited data sharing.
The limited availability of real data at early stages of the project can delay the data-driven solution, with a high risk of generating a biassed model and poor performance. A mitigation action can be the synthetic data generation, a pipeline emulating realistic data flows, in 4 steps: data research, model selection and training (also with AI), sampling and generation, and validation (comparing synthetic and real-world data).
Heterogeneous IT landscapes and data sets between test environments complicate the creation of a common knowledge base. A mitigation action could be to develop a modern data engineering tool improving the semantic and interoperability of reported data, ingesting and processing them to serve them to another tool.
At a major scale, is the high reluctance to share data between companies. A mitigation action can be to develop a secure fair data space with based on Zero-Trust Security principles (i.e., data anonymisation, attribute-based encryption, multi-party computation) to provide a safe data sharing environment based on the principle of “data sovereignty by design”, allowing for the smooth steps of data sharing: discover, negotiation, and data transfer.
From this, Data4Circ distinguish two types of intelligence: a collaborative intelligence, between companies, and local intelligence, for internal training and use of tools.
Group discussion: awareness, reliability, and heterogeneity
Inside the two macro-topics of data management and predictive modelling, the gaps and obstacles identified by the speakers have been further detailed and divided into four overarching categories: data gathering and reporting, data analysis, use of results, and governance. Workshop participants have been divided into four groups to face such issues, addressing questions, providing experiences and learnings, and suggesting good practices.
Transparency and reliability have been the most discussed issues, especially when dealing with metadata or confidentiality in collaborative research. A clear dialogue between all actors involved, including possible external agreement about data utilization with modelling experts, has been stressed out as a key factor in co-work and trust building. Moreover, cross-checking and the utilization of different data sources (e.g. literature, experiments) are required to validate the modelling, but define the minimum of concordant sources needed to use is a case-by-case parameter, like the quality checks to catch bad inputs. Finally, trust in AI can be strengthened by consulting experts whose models have been trained and validated over many years.
Lack of dedicated resources and of awareness among scientist have been indicated as ones of the main factors hindering design modelling application. A stronger advocacy on advantages of data management and analytics is required, but a proper governance, with responsibilities and workflows defined since the early stages, can decrease costs; it must be clear, though, that it is an iterative process improving itself along the way, and that mismanage of uncertainties can thwart assessment (e.g. SSbD or LCA). Again, looking for experts makes the whole process more efficient and reliable.
Yet, modelling have been confirmed to vary a lot in case-by-case scenarios, giving some participants’ questions on how it could help in validating properties of new molecules, how to choose between literature, modelling, and experiment for a given property, or how much training data is needed before machine learning methods could beat simple human-made correlations for property estimation.
