Jupyter notebooks bundle executable code with documentation and output, making them a popular mechanism for sharing computational workflows. We assess computational reproducibility at scale for notebooks associated with biomedical publications — mining PubMed Central full texts, locating notebooks on GitHub, and re-executing them in environments as close to the original as possible. Our study covers over 27,000 notebooks from 2,660 repositories across two independent runs, identifying key factors that influence reproducibility and trends in notebook quality over time. Recent work extends this to automated containerization to close the reproducibility gap.

ReproduceMeON is an ontology network for the reproducibility of scientific studies, bringing together foundational and core ontologies that capture different aspects of scientific experiment provenance. The development uses a semi-automated approach combining ontology matching techniques to select and develop core ontologies for each sub-domain — including scientific experiments, machine learning, computational workflows, and microscopy — and links them to existing domain ontologies.

Machine learning is an increasingly important scientific tool, but ML experiments face a reproducibility crisis similar to other disciplines. This project investigates factors beyond source code and dataset availability that affect ML reproducibility, proposes ways to apply FAIR data practices to ML workflows, and develops methodologies for end-to-end reproducibility of ML pipelines — including custom loss functions for robust neural networks and systematic benchmarks for deep learning reproducibility in biodiversity research.

Deep learning models are widely used in scientific domains, but their internal mechanisms remain opaque, hindering validation and improvement. This project develops interpretability methods that leverage domain knowledge — particularly ontologies and knowledge graphs — to produce human-understandable explanations extracted directly from neural networks. Current focus is on plant disease classification for sustainable agriculture, with extensions to multimodal large language models.