Abstracts & Speakers

SESSION I: DATA ANNOTATION AND FUNCTIONAL ONTOLOGIES

Epigenetic Legacy: The Role of Sperm miRNAs in the Paternal Inheritance of Diabetes and Obesity Development

Evidence shows that metabolic diseases such as obesity and diabetes are more pronounced in the offspring of malnourished parents. Studies suggest that predisposition can be inherited via epigenetic information in gametes. This has sparked growing interest in small regulatory RNAs in sperm as carriers of epigenetic inheritance. However, the functional annotation of dysregulated sperm microRNAs (miRNAs) in obesity and diabetes remains limited. This work addresses this gap by analysing publicly available datasets of diet-regulated sperm miRNAs and linking them to genes functionally associated with obesity and diabetes with an in silico approach. We systematically identified diet-responsive sperm miRNAs and overlapped their predicted targets with genes associated with metabolic phenotypes, as catalogued by the International Mouse Phenotyping Consortium (IMPC). With a sequence-based approach we uncovered 11,272 and 6,528 potential target genes for miRNAs regulated by the acute and chronic HFD interventions, respectively. By overlapping these predicted target genes of sperm miRNAs with our IMPC-derived list of obesity and diabetes associated genes, we identified more than 1,000 HFD predicted response genes. To support further research, we provide the field with the ShinyFatSperm App, which facilitates the functional interpretation of diet-regulated sperm miRNAs and enables users to explore their roles in the intergenerational transmission of metabolic disease risk. Our findings reinforce the concept that paternal dietary exposures can influence offspring health through epididymal- and sperm-borne miRNAs. This work provides a roadmap for hypothesis-driven investigation into the intergenerational inheritance of metabolic disease and highlights the urgent need for translational strategies to interrupt this cycle.

T-cell workflow for immunotherapy safety and tumour response optimisation via integrated discovery – a regulatory-grade NAM for decoding the pathogenic TCR-antigen landscape

This work presents a human, data-rich platform that links high-throughput T cell receptor (TCR) discovery to mechanistic, partially in silico modelling of immune-mediated cardiotoxicity and tumour control. Starting from single-cell TCR and transcriptomic profiling of patients receiving immune checkpoint inhibitors, we computationally cluster and prioritise pathogenic and tumour-reactive clonotypes, then resolve their cognate antigens using automated ex vivo screening against large peptide and cDNA libraries in HLA-engineered induced pluripotent stem cell–derived antigen-presenting cells.

For immune-related cardiovascular toxicities, we apply this workflow to historically collected cohorts of 200 patients with checkpoint inhibitor–associated cardiac dysfunction to identify up to 200 high-priority pathogenic TCRs and systematically map their cardiac self-antigens. These antigen–TCR–HLA axes are then functionally tested in advanced 3D human cardiac organoids co-cultured with TCR-engineered T cells, generating multiparametric readouts (contractility, electrophysiology, cell death, cytokine release) that can feed into quantitative, in silico models of tissue injury and patient risk.

By integrating harmonised protocols, multi-site validation and AI-based analysis of the resulting multi-omics and functional datasets, our platform provides a regulator-ready New Approach Methodology for predicting both therapeutic benefit and off-target toxicities of antigen-specific TCR interventions, while enabling formal human-versus-mouse benchmarking of predictive performance. This directly supports the INFRAFRONTIER 2026 agenda of bridging experimental disease models with AI-enhanced, in silico frameworks for human disease, and offers a concrete path to reduce animal use by replacing large parts of TCR safety and efficacy assessment with human-centric experimental–computational pipelines.

Towards controlling and imaging stem cell differentiation one cell at a time in organoids.

In this presentation, we will explore the role of spatially distributed growth factors in directing stem cell differentiation and tissue organization. I will introduce an experimental strategy that employs 3D colloidal scaffolds biofunctionalized with growth factors and embedded within spheroids of human stem cells. These scaffolds act as artificial organizing centers, guiding the spatial architecture of cardiac and liver organoids. I will demonstrate how the localization of these artificial centers influences the intrinsic secretion and deposition of tissue-specific extracellular matrix, which in turn directs cell differentiation and morphogenesis through the self-organized formation of local microniches. I will present results showing how this approach enables the in vitro recapitulation of periportal area development in the human liver and promotes the guided elongation of millimeter-scale bile ducts. Finally, I will discuss the potential of these colloidal scaffolds to support the assembly and spatial integration of multi-organ assembloids.

Advanced behaviour phenotyping for neurogenetic and preclinical studies in the laboratory mouse.

Behavior is the ultimate output of the nervous system, yet linking altered behavior to altered neural circuit function, and ultimately to altered genetics, remains challenging. The Kumar Lab uses advanced behavior analysis across multiple biological domains to build tools and statistical models for high-throughput, objective, and precise phenotyping in mice. We developed the JAX Animal Behavior System (JABS), an open-source integrated hardware and software platform that uses deep learning to extract rich phenotypic measures—including locomotor activity, grooming, stride-level gait and posture, nocifensive behavior, body mass, and biological frailty—from video recordings in a standard open field apparatus. We have also led the development of JAX Envision, a commercial home cage monitoring system, that enables continuous observation of multiple mice for weeks to months. Our work spans neuroscience, gerontology, rare diseases, and oncology. I will discuss our findings across these domains and the future development of our methods, including the application of genome-wide association and cross-species genetic mapping to connect computationally derived mouse phenotypes to human disease. I will also discuss our efforts in the democratization of machine learning and computer vision methods, lowering the barrier to entry for advanced behavioral phenotyping in the research community.

SESSION II: DATABASES AS ENGINES OF FUNCTIONAL GENOMICS

IMPC as a foundational dataset for understanding human disease

This presentation will explore how the IMPC data is created, improved and validated by experts, mapped to human rare and complex disease and how the IMPC data and materials are used in exploring human disease and pre-competitive drug discovery. It will consider future modes of access for IMPC and and the global and long term impact of the data resource.

Functional Annotation of the Mouse Non-Coding Genome for Understanding Human Diseases

Clinical genetics has revealed that approximately 60% of rare human diseases remain undiagnosed and cannot be explained solely by variants in protein-coding sequences. Moreover, genome-wide association studies (GWAS) of common diseases have indicated that over 90% of disease-associated variants are found in non-coding regions. These findings highlight cis-regulatory elements (CREs) as a critical yet underexplored layer in disease biology. However, modeling such variants in mice is particularly challenging due to their abundance and complexity of gene regulatory mechanisms.

Recent large-scale consortia have now catalogued millions of candidate CREs across mammalian genomes in a cell-type-specific manner. Although it is difficult to predict their impacts on gene regulation and phenotypic outcomes, advances in emerging technologies and large-scale datasets provide an opportunity to refine the identification of key CREs and address this challenge.

In this talk, I will introduce an ongoing International Mouse Phenotyping Consortium (IMPC) initiative led by RIKEN BRC to address this challenge: (i) systematic prioritization of candidate CREs, (ii) genome editing in mice to model non-coding variants, and (iii) establishment of interdisciplinary working groups to integrate genomics, epigenomics, and phenotyping expertise. By moving beyond coding sequences, the IMPC has the unique opportunity to define the functional and phenotypic impact of non-coding variants at scale.

Exploring large-scale mouse phenomic data to better understand genetic underpinnings of neuropsychiatric disorders

Objective: Neuropsychiatric disorders, with complex genetics, diverse pathomechanisms, and lack of objective biomarkers, remain a challenge to diagnose and treat. Our work harnesses the vast potential of large-scale phenomic data from the International Mouse Phenotyping Consortium (IMPC) and advanced computational approaches to uncover conserved genetic mechanisms and systemic pathways, bridging preclinical research and human psychiatry. By leveraging these novel strategies, we also aim to illuminate the intricate interplay between brain, body, and behavior.

Methods: We analyze IMPC phenotypic data to identify mouse endophenotypes relevant to schizophrenia, autism spectrum disorders, and other psychiatric diseases. These endophenotypes are mapped to human disease traits using advanced tools, including machine learning, hierarchical clustering, and co-expression network analysis, combined with transcriptomic and multi-omics datasets. This integrative approach reveals novel candidate genes, disease pathways, and stratification frameworks that address the genetic and phenotypic heterogeneity of psychiatric disorders.

Results: Our analysis reveals conserved endophenotypes that link genetic disruptions to pathomechanisms in neuropsychiatric disorders. By uncovering key genes associated with sensory processing and behavioral traits, we highlight the critical role of specific brain region and brain-body interactions in disease risk and progression. These findings offer novel insights into biomarker discovery, patient stratification, and therapeutic targeting.

Conclusion: Large-scale preclinical datasets and cross-species approaches can drive breakthroughs in precision psychiatry. Thus, leveraging cutting-edge methodologies and mouse phenomic data paves the way for objective biomarkers, transformative diagnostics, and personalized treatments for psychiatric disorders.

A standardised mouse schema and integrated analysis tool for genotype-phenotype associations

Phenopackets provide a standardised framework for representing human phenotype and genotype data, facilitating consistent data sharing and analysis across studies. Genotype–Phenotype Statistical Evaluation of Associations (GPSEA) builds on this representation to identify statistically significant associations between genetic variants and clinical features. Together, these approaches enable comprehensive human genotype–phenotype association analysis and advance precision medicine research.

However, no directly equivalent structured framework or analytical toolkit exists for mouse data. The heterogeneity of mouse knockout models, including genetic background, perturbation strategy (knockout approach, allele design), and zygosity, introduces context-dependent variability that complicates the interpretation of genotype–phenotype relationships. This challenge is particularly relevant in the study of essential genes, where discrepancies in viability outcomes are observed and may be driven by underlying genetic context. Here, we present a novel Mouse Genotype–Phenotype Schema that standardises genotype and phenotype information derived from the Mouse Genome Informatics and the International Mouse Phenotyping Consortium. By harmonising the representation of mouse genotype and phenotype data, this schema improves interoperability across resources and supports robust computational analyses, allowing more consistent and scalable exploration of genotype–phenotype relationships in mouse models. We further introduce an integrated analysis tool that enables flexible selection of combinations of genetic features and phenotypic traits to explore statistically supported associations. We apply this framework to lethal phenotypes to assess its potential to uncover associations with viability outcomes and improve the interpretation of mouse knockout models.

FAIR and AI-ready image data resources at Euro-BioImaging

The rapid growth of biological and biomedical imaging generates complex, high-volume datasets that require robust stewardship to remain usable, reproducible, and valuable. At Euro-BioImaging, a coordinated ecosystem of services, community initiatives, and European projects supports researchers in implementing FAIR (Findable, Accessible, Interoperable, Reusable) principles across the full image data lifecycle. These efforts combine practical tools, expert guidance, and community-driven standards to enable FAIR and AI-ready imaging data and metadata. Euro-BioImaging’s Image Data Services provide hands-on support through data stewardship consultations, training, and openly accessible resources, helping researchers plan, structure, annotate, and share their data effectively. This includes guidance on metadata standards, data management planning, and deposition into trusted repositories such as the BioImage Archive and Image Data Resource, ensuring interoperability and long-term reuse. Community activities, including workshops and the Image Data Community Days, foster exchange of best practices and real-world experiences, bridging gaps between theory and implementation.

In parallel, Euro-BioImaging contributes to major European initiatives such as AI4Life, EUCAIM, and EVOLVE, which advance AI-ready datasets, harmonized metadata standards, and scalable infrastructures. These projects highlight the critical role of high-quality annotations and standardized metadata in enabling machine learning workflows and reproducible research. This presentation will share insights from these community experiences, emphasizing challenges, lessons learned, and emerging solutions. It will showcase how coordinated infrastructure, training, and collaboration empower the imaging community to produce FAIR and AI-ready data, accelerating discovery and enabling cross-disciplinary reuse in the life sciences.

scATLAS approaches and multimodal data integration for robust annotation of the immune system in preclinical models

Accurate annotation of immune cell populations in preclinical models remains a major challenge due to cellular heterogeneity, species-specific differences, and limitations of single-modality datasets. Recent advances in single-cell technologies, combined with reference atlas frameworks (scATLAS), provide new opportunities to standardize and refine immune cell classification across experimental systems. In this presentation, we explore how scATLAS-driven approaches, together with multimodal data integration, enable robust and reproducible annotation of immune compartments in preclinical settings.

Building on recent efforts to construct comprehensive single-cell immune atlases and to integrate transcriptomic and proteomic layers, we will highlight strategies for leveraging well-annotated reference datasets to improve cell type identification. We will discuss computational methods for mapping query datasets onto reference atlases, addressing batch effects, different cytometry technologies and transferring annotations across species. Particular attention is given to the integration of complementary modalities, such as scRNA-seq, and protein-based measurements, which together enhance resolution and reduce ambiguity in immune cell classification. We further illustrate how these approaches can uncover conserved and divergent immune cell states between model organisms and humans, facilitating translational research. By combining scATLAS frameworks with multimodal integration, researchers can achieve more consistent, operator independant and biologically meaningful annotations, ultimately improving the interpretability of preclinical studies.

These advances pave the way toward standardized annotation in immune profiling pipelines, enabling better cross-study comparisons and accelerating the development of immunotherapies and disease models.

Data Welfare is Animal Welfare: Leveraging the FAIR3R Initiative and WELLFAIR Concept

The successful transition towards predictive in silico models of human diseases depends heavily on robust, reproducible data, often initially derived from in vivo animal models. However, fragmented data sharing practices frequently lead to substantial data attrition, hindering reproducibility and compromising the 3R principles (Replacement, Reduction, Refinement).

To address these ethical and regulatory challenges, we recently introduced the WELLFAIR concept in Neuroscience Applied, establishing the paradigm that “data welfare is animal welfare.” WELLFAIR conceptually merges the FAIR principles with the 3Rs, arguing that ethical animal research mandates rigorous, transparent data stewardship.

Translating this concept into actionable practice, we present the FAIR3R initiative. FAIR3R acts as a comprehensive, turnkey solution for structured data sharing. It provides researchers with a ready-to-use, standardized framework to seamlessly capture, annotate, and share in vivo data in strict compliance with FAIR principles. By eliminating the technical and structural barriers to data management, FAIR3R ensures that valuable experimental results are systematically transformed into Findable, Accessible, Interoperable, and Reusable digital assets.

By empowering researchers with an accessible tool for structured sharing, FAIR3R directly tackles the reproducibility crisis and transforms standard databases into powerful engines of functional genomics. Ultimately, the synergy between the WELLFAIR philosophy and the FAIR3R turnkey solution creates a highly effective, ethical bridge connecting traditional animal models to the next generation of mechanistic in silico disease modeling.

SESSION III: MECHANISTIC IN SILICO DISEASE MODELING AND PREDICTIVE MEDICINE

AI in Human Genetics: Enhancing Discovery and Mapping Disease Mechanisms

I will discuss opportunities for artificial intelligence in human genetics to connect disease-associated genetic variation with cell- and tissue-level phenotypes and inform interpretable models of human disease. I will first introduce HistoGWAS, a framework that integrates histology foundation models for automated morphological phenotyping with scalable variance-component methods to perform genome-wide association studies of latent tissue imaging features. Coupled with generative modeling, HistoGWAS enables visual interpretation of how specific genetic variants influence tissue architecture. I will then discuss extensions of these concepts to other imaging and genomic modalities—for example, fine-tuning retinal foundation models to enhance discovery and interpretation of retinal disease loci, or identifying causal cell types and states using single-cell data. Finally, I will present opportunities to bridge sequence-based and population-level models, highlighting a Bayesian framework that integrates pathogenicity scores from sequence language models to improve discovery power in rare-variant association studies. Together, these approaches illustrate how AI-driven analysis of large-scale biological data can move the field beyond statistical association toward interpretable and predictive models of disease biology.

From Cells to Chips to Digital Twins: Statistically Sound Parameter Estimation for Agent-Based Models in Organ-on-Chip and Cancer-on-Chip Systems

Organ-on-Chip (OoC) and Cancer-on-Chip (CoC) platforms are increasingly paired with digital twins to turn rich experimental data into predictive, mechanistic models of transport, growth, interaction, and treatment response. Agent-based models (ABMs) are especially attractive in this context because they represent individual cells, heterogeneity, local rules, and emergent behavior, all of which are central to microphysiological systems. While model construction is straightforward, if complicated, the major obstacle to making these models scientifically useful is obtaining statistically correct parameter estimates.

OoC and CoC experiments typically produce noisy, multiscale, and only partially observed data, such as time-lapse images, cell trajectories, concentration-time curves, and endpoint assays. ABMs are stochastic, computationally expensive, and usually have intractable likelihoods, so that classical estimation methods are often unavailable or unreliable. Moreover, different parameter combinations can generate very similar observable outputs, creating serious practical non-identifiability and making visually good fits potentially misleading.

Against this background, we discuss observation models, summary statistics, identifiability, uncertainty quantification, and validation, and review methods suited to simulator-based inference, including discrepancy minimization, Approximate Bayesian Computation, synthetic likelihood, sequential Monte Carlo, and newer simulation-based Bayesian approaches. The goal is to move beyond ad-hoc tuning and toward defensible parameter estimates and predictive distributions for trustworthy OoC and CoC digital twins.

Leveraging mathematical model to advance vaccination against SARS-CoV-2: using virtual mice cohorts

Rapid vaccine development requires efficient preclinical evaluation of an increasing number of vaccine strategies, while addressing the ethical imperative to reduce animal use in research. We aim to develop a computational modelling approach that provides a simulation platform for vaccine research by enabling the construction of virtual mouse cohorts from experimental data.

We build on a previously published mechanistic model (Wilding et al., 2025), developed across several vaccine platforms and pathogens, which describes the immune response from antigenic stimulation to B-cell proliferation and differentiation into plasma cells and memory B cells, ultimately leading to antibody production. We adapted this framework to a dataset consisting in 13 mice vaccinated with monovalent (5) or bivalent BNT162b2-based (7) SARS-CoV-2 regimens. Mice were followed recording longitudinal antibody measurements for 490 days, with blood samples collected monthly and, at selected early time points, weekly.

We show that the resulting model generates predictions that can be interpreted as digital twins of individual mice and can be used to quantify and compare immune responses to various vaccination strategies, allowing to generate virtual mice cohorts. This framework enables in silico simulation of counterfactual scenarios under alternative vaccination schedules, such as alternative or additional booster doses that were not administered experimentally. Finally, we illustrate how this platform can be used to optimize the design of future experiments, particularly with respect to the sample size, the sampling times and follow-up duration for investigating the persistence of the immune response.

Designing biological molecules with generative AI: from RNA structure to synthesisable drugs

Recent advances in generative artificial intelligence are beginning to transform how biological molecules are designed. Across RNA engineering, protein representation learning, and drug discovery, new models can now generate sequences and structures that are not only novel but also experimentally viable. For example, structure-conditioned RNA language models can design complex RNA folds and catalytic molecules that match or exceed human expert performance in laboratory validation, enabling the automated creation of functional ribozymes and structured RNAs. In parallel, synthesis-aware generative frameworks for small molecules incorporate chemical reactions directly into the generation process, producing drug-like compounds together with feasible synthetic routes rather than purely theoretical structures. Complementary work on large-scale protein representation learning shows that training on massive structural datasets improves the ability of models to capture relationships between sequence, structure, and function, providing foundations for data-driven biomolecular design. Together, these studies highlight a shift toward integrated AI pipelines that couple generative design with structural reasoning and experimental validation, moving the field closer to programmable biomolecules and practical AI-driven biotechnology.

Towards AI-Generated Virtual Control Groups in Preclinical Research: Insights from the IHI VICT3R Project

The VICT3R project, funded by the Innovative Health Initiative (IHI), aims to reduce animal use in preclinical studies by replacing the concurrent control groups with Virtual Control Groups (VCGs). By leveraging historical control animal data through machine learning and advanced statistical modelling, VCGs promise a scientifically rigorous and ethically responsible alternative to traditional study designs, in line with the 3Rs principles (Replacement, Reduction, and Refinement). This talk presents the AI methodologies being developed and qualified within VICT3R for VCG generation, addressing key challenges such as data extraction and curation, generation of VCGs that cope with the multivariate data distribution and match the characteristics of the study design, and achievement of regulatory acceptance. We will discuss how access to large, well-curated repositories of historical data is a critical enabler for this work, and how collaboration with research infrastructures could accelerate the development and broader adoption of VCGs.

SESSION IV: ANIMAL MODELS TO DIGITAL HUMAN DISEASES

Multiscale modeling of Hippocampal CA1 networks

Understanding hippocampal function requires modeling approaches that span multiple scales of biological organization. In this talk, I will present work from our laboratory ranging from the detailed characterization of single-cell properties to the reconstruction of large-scale networks, and further to the analysis of signal propagation along the hippocampal circuit. I will also show how this multiscale framework can be extended to translational applications, offering new insights into both healthy hippocampal processing and neurological dysfunction.

SESSION V: ETHICAL, REGULATORY AND REPRODUCIBILITY CHALLENGES

Beyond Modelling: The Future of Regulatory Evidence in Pharmaceutical Innovation

Pharmaceutical innovation is approaching a fundamental inflection point. For more than half a century, regulatory science has been built around a relatively stable paradigm in which biological evidence is primarily generated through experimental systems and subsequently interpreted through pharmacological, statistical, and clinical frameworks. This paradigm has enabled extraordinary therapeutic progress and shaped modern drug development. Yet it is increasingly challenged by the growing complexity of human disease, persistently high clinical attrition, rising development costs, the emergence of precision medicine, and the global scientific, ethical, and regulatory imperative to move toward more human-relevant approaches. At the same time, computational science is undergoing a transformation of its own. What began as a discipline focused on mechanistic understanding, simulation, and predictive modeling is now evolving into something far more consequential. In silico models are no longer simply tools for interpreting biological data. They are increasingly influencing critical decisions across the pharmaceutical lifecycle: from dose selection and safety assessment to clinical trial design, patient stratification, regulatory strategy, and ultimately lifecycle decision-making across approval, post-marketing evidence generation, and pharmacovigilance.

This evolution creates one of the defining tensions of the next decade: scientific innovation is advancing faster than regulatory frameworks. And this raises one of the most consequential questions facing our field: when does a computational model stop being a scientific tool and start becoming regulatory evidence?

This talk proposes a new strategic perspective on the regulatory evolution of evidence: from traditional animal-centric paradigms, to model-informed drug development, and now toward hybrid evidence ecosystems. Using real-world regulatory case studies, some computational approaches that achieved regulatory maturity will be discussed. Particular attention will be given to recent regulatory developments, including the 2026 NAM draft guidance released by FDA and early dialogue tools. Building on this perspective, the talk will introduce a regulatory vision as a conceptual framework for evaluating the regulatory readiness of next-generation computational evidence. Beyond scientific performance, future regulatory adoption will increasingly depend on interoperability, FAIR-aligned data ecosystems, cross-platform reproducibility, and distributed evidence generation across research infrastructures. In this context, Europe is uniquely positioned to lead the next era of regulatory science through its research infrastructures, translational innovation ecosystems and institutions.

Supporting NAMs innovation: why standardisation and validation matter!

New Approach Methodologies (NAMs) are transforming the landscape of scientific research, regulatory science, and safety assessment by offering alternatives to traditional animal-based testing. These methods, which include organ-on-chip, 3D bioprinting, high-content screening and computational models, promise improved human relevance, scalability, and ethical acceptability. However, the translation from technological and scientific innovation to industrial and regulatory use needs to be supported by two closely related activities: standardisation and validation. This talk will focus on the principles and processes that are on the basis of validation and standardization, on current advancements of NAMs used for chemical safety and medicinal development, and on best practices and tools that should be used by scientists and developers of NAMs.

Beyond Replacement: Transparent Integration and Communication of NAMs and Animal Research

The past decade has seen significant advances in New Approach Methodologies (NAMs)—including biobanking, organoids, organs-on-chips, computational models, and human-based data— alongside ongoing discussions about the future role of animal research. This joint talk brings together two complementary perspectives: one from the European Animal Research Association (EARA), focusing on the transparent communication of both NAMs and animal research, and another from a researcher who integrates these approaches to address complex questions in biobanking, biodiversity and reproductive biology.

The session will explore the capabilities and current limitations of NAMs, particularly in studying complex biological systems. It will highlight their role as valuable tools that can refine experimental design, reduce variability and contribute to improved animal welfare in line with the 3Rs principles.

A central focus will be the responsible communication of biomedical research methodologies. This includes clearly acknowledging the strengths and limitations of NAMs and animal models, avoiding oversimplified narratives and promoting transparent, evidence-based dialogue. Emphasis will be placed on selecting the most appropriate model, or combination of models, based on the scientific question, rather than prioritising methods based on external pressures.

Drawing on practical research experience, the talk will illustrate how integrating NAMs and animal models support the generation of robust and translatable data. It will also reflect on the importance of continued investment in diverse methodologies, recognising that many areas of biomedical research still rely on animal models and complementary approaches to address complex scientific challenges.