October 1, 2025
Project overview
This project investigates how causal reasoning can support the analysis of relationships between the gut microbiome and health. Its context is personalized medicine: understanding which microbiome and patient profiles might benefit from an intervention. The observations combine microbiome analyses with self-administered health questionnaires covering lifestyle, symptoms and medical conditions.
Associations between bacterial abundance and questionnaire responses do not establish causality. A shared, potentially unobserved factor may explain both observations. Students will explore how causal models can help distinguish these situations, drawing on causal inference theory and research on causal discovery for the microbiome [1, 2].
The central objective is to connect observational data with scientific knowledge. Language models will be used to extract functional relationships involving bacteria from the biomedical literature and organize them into knowledge graphs. Students will study how these graphs can inform causal hypotheses and how their proposed structure can be assessed against the available observations. The supplied readings cover biomedical causal feature selection, language-model-assisted causal discovery and knowledge graph extraction [3-5], alongside personalized medication recommendation and an example of microbiome-related mechanisms [6, 7].
The experimental workflow will use synthetic scenarios to test basic mechanisms before exploring their application to the project data. Python, PyTorch and the PyWhy ecosystem [8] provide the technical environment. The aim is to demonstrate and evaluate the building blocks of a causal analysis workflow. Expected deliverables include a demonstrator, reproducible experimental code and a report documenting the hypotheses, experiments and limitations.
References and resources
[1] Pearl, J., & Mackenzie, D. (2018). The Book of Why: The New Science of Cause and Effect. Basic Books. ISBN 978-0-465-09760-9.
[2] Corander, J., Hanage, W. P., & Pensar, J. (2022). Causal discovery for the microbiome. The Lancet Microbe, 3, e881-e887. https://doi.org/10.1016/S2666-5247(22)00186-0
[3] Malec, S. A., et al. (2023). Causal feature selection using a knowledge graph combining structured knowledge from the biomedical literature and ontologies: A use case studying depression as a risk factor for Alzheimer’s disease. Journal of Biomedical Informatics, 142, 104368. https://doi.org/10.1016/j.jbi.2023.104368
[4] Susanti, Y., & Färber, M. (2024). Knowledge Graph Structure as Prompt: Improving Small Language Models Capabilities for Knowledge-based Causal Discovery. arXiv:2407.18752, version 3. https://arxiv.org/abs/2407.18752v3
[5] Mo, B., et al. (2025). KGGen: Extracting Knowledge Graphs from Plain Text with Language Models. arXiv:2502.09956, version 1. https://arxiv.org/abs/2502.09956v1
[6] Li, X., et al. (2024). CausalMed: Causality-Based Personalized Medication Recommendation Centered on Patient Health State. Proceedings of CIKM ‘24. https://doi.org/10.1145/3627673.3679542
[7] Fayt, C., Morales-Puerto, N., & Everard, A. (2025). A gut microorganism turns the dial on sugar intake. Nature Microbiology, 10, 270-271. https://doi.org/10.1038/s41564-024-01917-1
[8] PyWhy. Causal inference ecosystem. https://www.pywhy.org/