Deep learning method maps spatial architecture of brain cancer cells

Glioblastoma is the deadliest and most common type of malignant brain tumor in adults. One of the reasons it's almost always incurable is that individual tumor cells are very different from one another and shift dynamically between various 'states'. This plasticity helps them evade anti-cancer treatments.

Now, two new studies from researchers at EMBL, the German Cancer Research Center (DKFZ), and the Wellcome Sanger Institute, UK, show that this state-shifting follows predictable patterns, providing hope for future therapies targeting these cancers. The results were recently published in the journal Nature. 

Building a molecular atlas of glioblastoma cells

Researchers and clinicians have long been stymied by the high heterogeneity of glioblastoma tumors. Cells within these tumors exist in a variety of states, each with distinct molecular features. Some resemble nerve cells, while others look like stem cells or become trapped in an aggressive, injury-like state. The cells can also transition from one state to another, a phenomenon known as plasticity. Since therapies may target some cell states but not others, such plasticity helps the cancer cells survive treatment.

However, the prevalent view in the field has been that such state-shifting is essentially random and very difficult to predict. This is the view the researchers in the current study challenged by creating and analysing one of the most comprehensive molecular datasets derived from glioblastoma patients to date. 

"We asked – can we go back to the drawing board?" said Omer Bayraktar, Group Leader at the Wellcome Sanger Institute and a new professor at the UCL Cancer Institute. Bayraktar's team uses single-cell spatial technologies to study the cellular diversity of brain disorders. "Can we assume that we know nothing about this disease, and reconstruct how a cancer cell behaves in glioblastoma through quantitative approaches?"

Funded by a Wellcome Leap grant, Bayraktar's team examined DNA, epigenetic marks (a type of molecular 'annotation' of genes that regulates their expression), and RNA in tumors from 12 patients, generating the largest glioblastoma molecular atlas to date. To make sense of this highly complex dataset, they teamed up with Oliver Stegle's team at EMBL Heidelberg and DKFZ, who are experts in applying advanced computational approaches to fundamental life science problems. 

We wanted to figure out whether there is a specific rule-based system that determines how cells transition between these states. And we asked not only whether there were rules, but also what would drive such rules."

Oliver Stegle, EMBL Heidelberg

Together, the researchers found that, contrary to what was known before, glioblastoma cells followed certain stereotypical, predictable paths between cell states. Different cell states also occupy different spatial niches inside a tumor. 

"By integrating the single-cell data with spatial transcriptomic data, we found that there is a meticulous and very specific spatial architecture of cancer cell states," explained Bayraktar. "And this was present in every single glioblastoma tumor in every single patient we studied."

The team further confirmed this data using organoids – tumor cells taken from patients and grown in the laboratory on a three-dimensional scaffold, closely mirroring conditions inside the brain. They found that similar cell state trajectories emerged in patient-derived organoids as what they had deduced earlier from the glioblastoma atlas data. 

Understanding the rules and regulators that govern state transitions

Now knowing that glioblastoma cells follow certain stereotypical trajectories between states, the researchers then set out to investigate how this process is regulated. For this, they turned their attention to gene regulatory networks. These are the rules and molecular regulators that determine which genes are expressed in which cells and when. 

Led by Moritz Mall's team at DKFZ, the researchers examined gene regulation signatures in over 1 million patient-derived glioblastoma cells and used this to figure out how tumor cells transition between states. They identified a structured regulatory program that governs cellular plasticity in these cells, one that was both predictable and hierarchical. 

The team also identified a set of 'safeguards' – regulators which arrest cells in low-plasticity states, preventing them from transitioning to other states. This was of particular interest as these regulators could potentially be targeted genetically or pharmacologically as part of cancer therapies in the future. 

"One of our premises has always been that when we understand these rules, we can not only observe what the cells are doing, but we can also change what they are doing," said Mall. "We could perhaps reprogram them, and maybe push them into a state that is less malignant, less dangerous."

The researchers tested this hypothesis in patient-derived brain tumor cells, where they found that overexpressing MYT1L, one such 'safeguard' regulator identified by this study, could reduce cellular plasticity and tumor malignancy. If transplanted into mice, these 'modified' tumors resulted in better survival than control tumors.

"The results with MYT1L were interesting, but the findings go much beyond that," explained Mall. "Our approach predicts the regulators for all of the major tumor cell states, and we tested those predictions experimentally. We were able to steer tumor cells into specific states by manipulating just a handful of regulators." In other words, the predictions of the new model could be used to engineer the identity of a tumor cell.

AI turns tumor diversity into a source of insight

The diversity that makes glioblastoma so hard to treat also turned out to be the key to understanding it. Each tumor, and each cell within it, carries its own combination of genetic and epigenetic changes. Across a million cells, this adds up to a vast number of natural variations in how genes are regulated, far more than researchers could ever create in the lab.

To learn from this variation, 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, scDORI pinpoints which regulators steer cells towards one state or another. These predictions were the starting point for the reprogramming experiments carried out by Mall's team.

"A tumor generates an almost endless number of variations on its own regulatory program," said Stegle. "Instead of seeing this heterogeneity only as an obstacle, we can use AI to learn from it and extract the rules that determine how cells choose their fate."

A second machine-learning method mapped where each cell state sits within the tumor, revealing the spatial architecture described above.

The researchers are now applying this approach to other cancers and brain disorders. While clinical applications may still be some way off, this is already helping researchers understand tumor cells at the deep molecular level and spot common patterns. "With the help of this atlas and the new AI tool, we identified new regulatory sequences that enable us to potentially drive specific cell states in a tumor-specific setting," said Mall. "Now that we have this framework, we can apply it to many other questions and cancers."