Research
We develop machine learning and statistical methods to make sense of complex, multi-modal biological data. Our work spans tool development, disease modelling, and collaborative data analysis, always with a focus on reproducibility and biological interpretability.
Multi-omic Integration
Modern biology generates data across many molecular layers: genome, transcriptome, proteome, and more. We build computational methods that jointly analyse these heterogeneous datasets to reveal biological patterns invisible in any single modality. Our recent tool Multiverse provides a reproducible benchmarking framework for multimodal single-cell integration, and MIMA is a multimodal variational autoencoder for learning unified cellular representations.
DeepHeart - Variant Prioritization for Congenital Heart Disease
Genome sequencing produces millions of variants per individual, the vast majority of which are benign. Identifying the small subset responsible for disease is a critical bottleneck in clinical genetics. As part of the DeepHeart project, we are developping deep learning models that integrate variant features, gene expression, and patient phenotypes to prioritize candidate variants in congenital heart disease, improving diagnostic yield for patients.
Explainable AI for Multi-omic Representation Learning
Building accurate models is not enough; we need to understand what they learn. We are developping explainability methods tailored to multi-omic settings, enabling researchers to trace model predictions back to interpretable biological features such as gene programmes, regulatory elements, and cell states. This work bridges the gap between black-box deep learning and actionable biological insight.
RNA Subcellular Organization
Where an RNA molecule sits inside a cell is not random, subcellular localisation is tightly regulated and functionally important. We develop deep learning models that map RNA localisation at subcellular resolution by fusing transcriptomic readouts with high-content microscopy images. Our tool Subcellspace provides an end-to-end pipeline for this analysis.
Collaborative Bioinformatics within LISCO
As members of the Leuven Single-Cell Institute (LISCO), we work closely with experimental groups to help them get the most from their single-cell and spatial data. A key focus is end-to-end analysis pipelines for marker gene identification, from raw count matrices to biologically validated cell-type signatures. These collaborations drive new methodological questions and ensure our tools are tested on real and diverse datasets.