ESPE Yearbook of Paediatric Endocrinology

Nat Genet. 2025;57(4):797-808. doi.org/10.1038/s41588-025-02124-2

Brief Summary: This ‘Perspective’ article outlines how causal machine learning (CML) can be applied to single-cell genomics to uncover mechanistic insights into cellular processes. The authors discuss the limitations of current statistical learning approaches and propose causal models that can generalize across experimental conditions, interpret biological mechanisms, and capture temporal dynamics. They highlight three major challenges: 1) generalization to novel perturbations, 2) interpretability of learned models, and 3) modeling of dynamic cellular processes.

Integrating causal inference into single-cell biology offers the opportunity to develop models of gene regulation and cellular behavior, supporting drug discovery and experimental design. It encourages the use of interpretable models that align with known biological pathways, regulatory mechanisms, and disease processes.

Limitations: Current causal models often rely on assumptions, such as perfect interventions, that may not hold in real biological systems, which are typically more complex, involving multimodal and spatiotemporal dynamics. Additionally, there is a lack of large, standardized, and high-quality interventional datasets needed to effectively train and validate these models.