May 9, 2023 | Whereas some detection and diagnosis tools search for mutations in certain genes or analyze the genome, an international group of researchers spanning Canada, the UK, and Australia has created an atlas of pediatric cancer using a machine-learning algorithm that considers the whole transcriptome.
One key finding emerging from their line of research is that pediatric cancers display greater levels of transcriptional variability, or the number of genes expressed in a cell, than adult forms of cancer. They were also able to find nuanced differences within subtypes of cancer.
Although pediatric cancers stem from fewer major tumor classes, they seem to display more complicated hierarchies, according to the research group. This suggests that many forms of pediatric cancer share a common ancestry and then diverge into an array of tumor subtypes.
From their vantage point, taking this type of variability into account can help improve the tools that are used to diagnose pediatric cancer.
“We know that fundamentally, childhood cancers are different than adult cancers,” said Adam Shlien, Senior Scientist and Canada Research Chair in Childhood Cancer Genomics at The Hospital for Sick Children (SickKids) in Toronto, Canada. This is not only because these cancers emerge earlier in life but also because a different spectrum of cancers is involved.
Many studies, including some by his lab, have examined the whole genome sequence of childhood cancers to attempt to find drivers and to time the emergence of those drivers in kids with cancer.
Some tumors found in childhood are “quiet” in that they do not have many drivers, and Shlien and colleagues have used this to their advantage. “We think it actually gives you a very clear, unobstructed view of what’s happening and what the key linchpins are in pediatric cancer,” he told Diagnostics World.
That said, Shlien pointed out that there has never been a detailed analysis of the transcriptional consequences of these mutations in pediatric cancer compared to adult malignancies across all cancer types.
“We know from other busy tumors that we’ve looked at, just because there’s a copy number change or a DNA alteration doesn’t mean that tumor has decided that that’s important enough to be acted upon,” Shlien explained. So they decided to set aside the genomes and focus on the transcriptional uniqueness of childhood cancers.
That’s where the idea for their study was born, he added, and the results have recently been detailed in a paper published in Nature Medicine (DOI: 10.1038/s41591-023-02221-x).
Transcriptomics, or “the study of all RNA molecules in a cell” as the National Institutes of Health (NIH) defines the term, is a way of learning how genes are turned on and how this could contribute to the development of diseases like cancer.
In some instances, the presence or absence of a gene isn’t enough to answer a biological question, explained Sandrine Miller-Montgomery, CEO of cancer-detection company Micronoma, who has worked with RNA for a large part of her career, including roles in both academia and industry.
“You may want to know if the genes are actively transcribed into a protein or, on the other hand, inhibited. To that end, studying RNA expression is a good way to see which pathways are active or not,” added Miller-Montgomery, who was not involved with the Nature Medicine study.
Clustering Cancer Using Transcriptional Uniqueness
During their research, Shlien and colleagues realized the need to develop an atlas of pediatric cancer. His then-postdoc, Federico Comitani, who is now a research associate in the Genetics & Genome Biology program at The Hospital for Sick Children, created a new algorithm.
“To develop molecular definitions of childhood cancers, we designed a method that reduces the complexity of RNA-sequenced tumors and then groups them into hierarchically organized clusters,” the researchers noted in their paper. “This was done in a way that would enable a deeper exploration of the transcriptional differences between and within tumor classes and would facilitate the discovery of new tumor subtypes,” they added.
The technology makes it possible to cluster and organize human cancer based on transcriptional uniqueness. Analyzing the entire transcriptome makes it possible to find a tumor’s key features and gather a clearer sense of the cancer activity that is specific to each person, Comitani said in a news release.
Although many clustering algorithms have been designed over the years, Shlien said his group’s tool fine-tunes the clusters and provides a nuanced hierarchy of cancer. “The beauty of the algorithm also is that it doesn’t use any of the labels or diagnoses that the tumors were identified as beforehand,” he added.
So they applied their algorithm to over 13,000 tumors in kids and adults to define different clusters. They then went into a deep dive to better understand what makes childhood cancers unique and to learn more about the various clusters they inhabit, Shlien explained.
After coming up with this high-resolution transcriptional atlas of pediatric cancer, the researchers wondered whether it could be used as a diagnostic aid, so they trained a set of neural networks against the atlas to define a model that is unique to each and every tumor, he added.
“So if a Ewing sarcoma patient comes into the clinic and their RNA sequenced, it can be then very quickly matched against the atlas. And then we get a probability that this is, in fact, Ewing sarcoma based on the transcriptional profile,” he said.
“Quiet” Genomes, “Noisy” Transcriptomes
Shlien and colleagues used these technologies—the hierarchy of cancer based upon a transcriptional clustering algorithm, and a set of neural networks that makes it possible to match ongoing patients to that—to explore aspects of pediatric cancers.
The researchers found, for instance, that sarcomas (bone and soft tissue cancers) are divided into two main subtypes that differ based on markers of immune infiltration and markers of stemness. Shlien said they also found, “quite strikingly,” that many pediatric sarcomas were distinct and clustered separately. “Some of the tumors don’t look like sarcomas at all. Ewing sarcoma is a great example, which just clusters into its own little universe compared to the others.”
In some instances they also found markers that bring together tumor types from different parts of the body—that you’d think would have no connection—into a common cluster. “And we're really deeply exploring that now,” he noted.
Yet another key finding to emerge from their research, Shlien pointed out, is that although one might assume childhood cancers would have a “quiet” transcriptome, their findings revealed just the opposite.
“The tumors coming from these young cells are prone to try different versions of the transcriptome,” Shlien explained. “There are more copies and greater variability from cell to cell.” In other words, “they’re actually quite noisy.”
For example, if you take leukemias in children compared to leukemias in adults, he explained, there is a “massive difference” in transcriptional disorder. This is in complete contrast to the “quiet” genomes of pediatric cancers.
Shlien also pointed out that he and colleagues found certain subtypes of cancer, like some types of bone cancer, that seem to have prognostic importance.
Researchers Envision Universal Diagnostic Potential
As more samples are added to the atlas and as it is validated with larger data sets, Shlien and colleagues see potential for their classifier to be used as a novel, universal system for diagnosing cancer.
The process, as he described it, would include taking the RNA from a person’s tumor, sequencing it, deriving a simple set of gene expression values, and uploading that data to a website. Patients could be matched to the atlas within a few minutes and a probability of the tumor type and subtype would be generated.
“We found in a validation cohort of 300 patients, and we’ve now done validations in another 300 patients, that we can match or exceed the pathologist classification in about 89 to almost 95 percent of cases,” he said. In some instances, there will also be an ability to refine a diagnosis based on the transcriptional clustering.
Shlien indicated that the tool is already being used for cancer diagnosis purposes as part of the SickKids Cancer Sequencing program. One afternoon each week, patients with hard to cure cancers—often from The Hospital for Sick Children but also from across Ontario—are discussed. As part of this program, the medical team reviews several cases where the RNA has been sequenced and mapped to the atlas, along with the tumor type and subtype probabilities presented by analysts.
“We’ve seen a number of different examples where this has reversed the diagnosis of a tumor, or caused us to look differently,” he said.
One example that came to mind involved a young boy from about a year ago. Although a genome alteration in his tumor was initially missed, a strong signal from the RNA nudged the team to explore further. The team eventually discovered a “cryptic alteration of the genome, and because of that, the child’s therapy was changed.” Shlien’s understanding is that the patient’s response improved due to that change in diagnosis.
In addition to the SickKids Cancer Sequencing program, some early adopter cancer centers are using this technology to enable physicians to compare their patients’ diagnoses to cancer types that the platform has identified.
“What’s really exciting to me is that now a lot of our colleagues around the world are starting to use this web application and our models to map their patients,” said Shlien. Initially, this is on a research basis only, he pointed out, “but this is quickly moving into the clinical sphere.”
Moving forward, there are plans to build upon this line of work in a number of ways. The researchers have gathered additional samples that will be added to the atlas. They intend to retrain their model and make it available to the web application so that users will always be able to map their patient samples to the latest version of the atlas. They are also working to refine and improve the web application, and to make it easier to use.
Beyond that, they are working with Industry Partnerships & Commercialization at The Hospital for Sick Children to explore commercial routes.
And although their recently-published study emphasizes childhood cancers, Shlien highlighted the vision for extending this technology even further. “This is something that I think will be part of the next generation of tests for cancer diagnosis and ideally for cancer prognostication as well—even beyond pediatric cancer,” he said.
Paul Nicolaus is a freelance writer specializing in science, nature, and health. Learn more at www.nicolauswriting.com.