EMBL Explores Brain Cancer Cell States With AI and Single-Cell Technologies

Two companion studies in Nature from EMBL, the German Cancer Research Center and the Wellcome Sanger Institute describe how glioblastoma tumour cells move between distinct cellular states along predictable paths governed by gene regulatory networks. Using single-cell and spatial data plus deep learning, the teams also reprogrammed those states experimentally, pushing cells in a less malignant direction and improving survival in mouse models. EMBL News report.

The Plasticity Problem in Glioblastoma

EMBL describes glioblastoma as the deadliest and most common type of malignant brain tumour in adults. One reason it is almost always incurable, according to the report, is cellular heterogeneity: individual tumour cells differ from one another and shift between states with distinct molecular features.

Some cells resemble nerve cells, others stem cells, and others become trapped in an aggressive, injury-like state. Because therapies may target some states but not others, this plasticity helps tumour cells survive treatment. The researchers note that clinical applications may still be some way off.

Building the Molecular Atlas

Funded by a Wellcome Leap grant, Omer Bayraktar's team at the Wellcome Sanger Institute examined DNA, epigenetic marks and RNA in tumours from 12 patients, producing what the researchers describe as the largest glioblastoma molecular atlas to date. Analysis was done with Oliver Stegle's team at EMBL Heidelberg and DKFZ, who specialise in computational approaches.

Contrary to the prior view that state-shifting is essentially random, the data showed stereotypical, predictable paths between states, with different states occupying different spatial niches inside tumours. The researchers report this architecture appeared in every tumour and every patient studied, and that similar trajectories emerged in patient-derived organoids.

Regulatory Rules and Safeguards

Led by Moritz Mall's team at DKFZ, researchers examined gene regulation signatures in over 1 million patient-derived glioblastoma cells and identified a structured, predictable and hierarchical regulatory programme governing plasticity. They also identified safeguards: regulators that arrest cells in low-plasticity states and prevent transitions.

In patient-derived brain tumour cells, overexpressing the safeguard regulator MYT1L reduced cellular plasticity and tumour malignancy. Tumours modified in this way, when transplanted into mice, resulted in better survival than control tumours. Our editorial recommendation is to distinguish these experimental findings from clinical treatment evidence.

Where AI Enters the Workflow

Stegle's team developed scDORI, a deep learning method that infers gene regulatory networks at single-cell resolution. By comparing how regulators and their target genes change together across more than a million cells, it pinpoints regulators that steer cells toward particular states; those predictions seeded the reprogramming experiments.

A second machine-learning method mapped where each cell state sits within the tumour, revealing the spatial architecture. The researchers report that manipulating only a handful of regulators steered tumour cells into specific states, and say they are now applying the approach to other cancers and brain disorders.

What This Means for Practitioners

Editorially, the practical takeaway is not a therapy announcement. The work offers a rule-based framework connecting single-cell atlases, spatial data and machine-learning inference of regulatory networks, a pattern research organisations could assess for other tumours with high cellular heterogeneity.

Teams considering similar programmes should note the dependencies described: large-scale patient-derived atlases, organoid validation, and experimental testing of computational predictions. The report cautions that clinical applications may still be some way off, and the reported survival observations come from mouse models. Our recommendation is to keep those experimental findings separate from any claim about patient treatment and to evaluate computational predictions alongside the underlying laboratory evidence.