Research - Dr. Gennady Roshchupkin

Opening Vision

Human health is written across genomes, images, and lifetimes of exposure and care. In our group, we study how these layers fit together in large population cohorts such as the Rotterdam Study and Generation R, and how they can be translated into models that are both predictive and biologically meaningful. We connect molecular variation to organ-level structure and function, and population context to risk, using integrative methods that combine multi-omics, quantitative imaging, and longitudinal clinical and lifestyle data with international consortia. The aim is not prediction alone, but a clearer path from population evidence to mechanism, and from mechanism to prevention and precision medicine.

The group consists of PhD candidates, postdoctoral researchers, and research engineers working at the interface of epidemiology, AI, and imaging. Projects are typically embedded in large cohort studies and international collaborations, combining methodological development with applied research.

Imaging and computation

Embedded collaboration with the Biomedical Imaging Group Rotterdam (BIGR): large-scale imaging science, GPU computing, and infrastructure for AI-driven discovery.

Population cohort depth

Prospective resources including the Rotterdam Study (ERGO) and training through NIHES, extended via meta-analysis and international harmonization.

Our Research

1. AI for Population Biology

Large population cohorts are one of the few places where genetics, imaging, and clinical data come together at scale. We use resources such as the Rotterdam Study and international consortia to model disease risk across the lifespan, combining multi-omics, imaging, and longitudinal clinical data. A key challenge is that these datasets are heterogeneous and evolve over time, so our models explicitly account for population structure, cohort differences, and longitudinal dynamics.

2. Interpretable and Causal AI

High capacity models can fit data without explaining it. We prioritize interpretability as a scientific output: attributions and structured representations that researchers can interrogate, not only dashboards. Epidemiological reasoning is embedded in model design and evaluation—explicit attention to confounding, selection, and measurement error—so that performance gains translate into claims that survive scrutiny. In practice, this means building models that remain stable across cohorts, imaging protocols, and definitions, and that can be interpreted by domain experts rather than only optimized for performance.

3. From Association to Mechanism

Genetic association is a starting point. We connect population signals to function—especially in the non-coding genome—by combining statistical genetics with models that highlight where and how variation may act. Quantitative brain MRI (structure and connectivity), musculoskeletal imaging, and related traits give a concrete bridge from variants to intermediate phenotypes. From there we move toward hypotheses about pathways and contexts that can be tested across data types and cohorts.

4. Scalable and Federated AI

Science advances when data can be used without being centralized. We build federated and privacy-preserving pipelines—including production-style workflows on NVIDIA Clara–based infrastructure—so models can be trained and validated across sites under governance constraints. Coordination within major international networks keeps methods aligned with shared scientific goals. Open software and reproducible workflows remain part of the research product: tools other groups can adopt, extend, and benchmark.

What We Deliver

Open-source AI frameworks (e.g., GenNet)
Population-scale analyses across international cohorts
Methods for interpretable and confounder-aware machine learning
Translational models linking genetics, imaging, and disease

Flagship Projects

GenNet — Our open-source framework for interpretable deep learning in genomics, spanning biologically informed neural network design, transparent phenotype prediction, and downstream translation to imaging, confounder correction, and functional interpretation.
Federated learning for population multi-omics — A federated stack (including NVIDIA Clara–based workflows) focused on multi-center genomics and related outcomes, prioritizing CHARGE and EADB-style collaborations where raw data cannot leave sites but scientific questions remain global.
Multi-modal integration across omics, imaging, and clinical depth — End-to-end pipelines that link genetics to quantitative brain MRI (structure, diffusion, connectivity, function), musculoskeletal and craniofacial phenotypes, retinal imaging, and other high-dimensional traits in cohorts such as ERGO and Generation R. Domains include deep imaging genetics for osteoarthritis, Face as a Health Mirror, and oral and craniofacial health.
Open tools for heritability, brain aging, and exploration — Public resources that translate methods into usable interfaces for researchers and educators (see below).
National AI R&I coordination (Netherlands UMC ecosystem) — Strategic leadership to align translational AI capacity across university medical centers, strengthening the environment in which cohort science and clinical translation meet.

GenNet

GenNet is our open-source framework for interpretable deep learning in genomics.

Key Outputs

Brain Age Prediction - Gray Matter Age Prediction as a Biomarker for Risk of Dementia (PNAS)
Latent-space confounder correction - AI-based association analysis for medical imaging using latent-space geometric confounder correction (code)
BRAIN-MAGNET - A functional genomics atlas for interpretation of non-coding variants (Cell, 2025)

We actively contribute to international consortia by developing AI methods for large-scale data integration, interpretation, and prediction across population cohorts—from multi-omics and imaging through to questions that matter for biology and clinical translation.

Interested in collaboration or joining the group? Please get in touch.

Methods in Practice

GenNet: network architecture

GenNet: explainable AI module

NVIDIA Clara framework for federated learning

GenNet pipeline for federated learning

Imaging and Quantitative Phenotypes

We derive and analyze endophenotypes from high-dimensional imaging: structural MRI, DTI, connectome and resting-state fMRI, OCT, 3D craniofacial and musculoskeletal shape, and related quantitative traits—linked to genetics and epidemiology in population studies.

Expertise

The group operates across the full pipeline: harmonizing complex data, building models with clear evaluation logic, interpreting outputs in genetic and phenotypic terms, and releasing software that others can run. Strength is not “methods for their own sake,” but end-to-end rigor—how a question is framed in a cohort, how bias is handled, and how results connect back to biology and practice.

Population genomics and multi-omics — Large-scale association, integration across molecular layers, and translation to interpretable genomic models.
Computational imaging and imaging–genetics — Quantitative neuroimaging, musculoskeletal and craniofacial phenotyping, and high-dimensional image-derived traits linked to genetics and epidemiology.
Statistical learning with epidemiological discipline — Validation strategies, confounder-aware modeling, and clarity about what a model can and cannot claim.
Distributed computing and open tooling — Federated workflows, reproducible pipelines, and interactive tools for dissemination.

Open Tools

We develop open source software and interactive resources so that methods can be explored, taught, and reused.

ImageNe Heritability Tool

Interactive exploration of heritability of brain imaging traits

Related paper: Multidimensional heritability analysis of neuroanatomical shape

Brain Aging Tool

Predict brain age from MRI scans

Collaboration & Ecosystem

We lead and coordinate within international research infrastructures, not only contribute analyses. The Machine Learning working group in CHARGE and bioinformatics leadership in GEMSTONE set shared standards for how AI is applied across cohorts. Partnerships span ENIGMA, eQTLGen, EADB, and EU COST networks, connecting genomics, imaging, and translational questions across countries. Within Erasmus MC, the program bridges Epidemiology and Radiology & Nuclear Medicine through embedded collaboration with BIGR, while co-creating community capacity via squAIre and contributing to imaging strategy and women’s health research leadership. This ecosystem connects detailed cohort data in Rotterdam with international efforts, enabling reproducible and comparable research across sites.

Partner networks and local collaboration

Within Erasmus MC

Epidemiology; Radiology and Nuclear Medicine
Neurosurgery; Plastic Surgery; Craniomaxillofacial Surgery
Internal Medicine; Psychiatry; Neuroscience
Genetic Identification

Consortia

EU COST Actions

GEMSTONE (musculoskeletal genomics)
ML4Microbiome

Closing Vision

Over the long term, we aim for a tighter loop between population evidence and clinical utility: models that are portable, interpretable, and grounded in mechanism, trained in settings that respect privacy and diversity. Success is measured by better understanding of disease architecture, fairer and more reliable tools for risk and progression, and a research culture where data, methods, and translation advance together.

Research program diagram: multi-omics, imaging, and population data flow into interpretable AI models; outputs include prediction, biological mechanism, and clinical impact; federated learning underpins scalable collaboration. — From multi-modal data and interpretable AI to mechanism and clinical impact—with federated infrastructure for privacy-preserving collaboration.