Skip to main content
Premium Trial:

Request an Annual Quote

Transcriptomic Classification of Pediatric Cancers Shows Potential to Improve Clinical Diagnoses

Premium

NEW YORK – Computational biologists from the Hospital for Sick Children (SickKids) in Toronto have created two new computational methods that they believe will facilitate wide adoption of transcriptomics for clinical diagnosis of pediatric cancers.

RACCOON, which stands for Resolution-Adaptive Coarse-to-fine Clusters Optimization, automated the process of analyzing transcriptomes and organized the samples into an atlas of 455 tumor and normal classes based on similarities in gene expression.

The other method, OTTER, for Oncologic Transcriptome Expression Recognition, combines three convolutional neural networks trained on the atlas. "With that, you can do a lightweight and quick match of someone's transcriptome to the atlas," said Adam Shlien, associate director of translation genetics in the department of pediatric laboratory medicine at SickKids.

Federico Comitani, a postdoctoral researcher in Shlien's program and lead author of a new paper in Nature Medicine, designed RACCOON and the atlas of 455 different cancer subtypes. OTTER grew out of that work.

The SickKids team applied RACCOON to 13,313 transcriptomes to build an atlas to map how pediatric tumors vary based on the patient's age and how they evolve over time. Most of the data came from the publicly available Treehouse Childhood Cancer Initiative's dataset, and SickKids supplemented that with more than 1,000 of its own tumors as well as data from the St. Jude Children's Hospital Pediatric Cancer Genome Project.

Using transcriptomes rather than mutations or methylation signatures means that the SickKids assessments "represent the active state of the disease," according to the paper. The researchers found that the RACCOON algorithm can assess established and novel forms of pediatric cancer alike.

In the study, OTTER confirmed 82 percent of pathology diagnoses in the research cohort and clarified diagnoses for another 7 percent of cases, making it 89 percent accurate. "Collectively, this work both defines the transcriptional distinctiveness of childhood cancer and uses this to validate a novel, pan-cancer diagnostic assay," the authors wrote.

Shlien and colleagues said that pediatric cancers have more transcriptional variability than adult cancers. "The transcriptional variability of childhood cancers is in stark contrast to the quietness of their genomes, generally harboring fewer substitution mutations at diagnosis" than adult cancers, they wrote.

"[RNAseq] is very useful for running fusions, but the ability to subtype cancers based on their expression profile was a little bit tough," Shlien explained. "We set out to discover the transcriptional uniqueness of childhood cancer, and along the way we created this atlas of all of human cancer and then focused really deeply on pediatric cancers."

The transcriptional variability they found in childhood cancers shows "that there is a lot more heterogeneity in pediatric cancer than in adults, and that was just quite fascinating," he said.

The convolutional neural network training took a while even on SickKids' high-performance computing infrastructure, he added, but now that the algorithm has been trained, the classifier can run a match in three to five minutes.

OTTER currently runs on a server at SickKids, but it has been designed for easy migration to a cloud, should the need arise. The translational genetics program has set up a web portal for outside users to compare their own transcriptomic samples to the atlas, but the site clearly states that it is for research purposes only because it has not been cleared for diagnostic use by regulatory authorities.

The SickKids team called this research "the first iteration of an ever-learning tool," saying that RACCOON "has the potential to grow such that it provides diagnostic or prognostic utility to every child with cancer."

"Because a critical number of childhood tumor transcriptomes are or will soon be available, RNA-seq has the potential to become a standalone 'universal diagnostic assay,'" the researchers wrote.

Shlien, who has a particular interest in sarcoma, runs what he calls a computational oncology laboratory. He said that the laboratory medicine department plans on implementing the atlas in clinical processes at SickKids and the institutions of his collaborators.

These include Toronto's Mount Sinai Hospital, which is right across University Avenue from SickKids, as well as farther-flung institutions like the University of New South Wales in Australia and University College London, the University of Cambridge, and the Wellcome Sanger Institute in the UK.

Shlien said that SickKids is already using RNA-seq to detect fusions and splicing defects for many cancer patients.

A research study called SickKids Cancer Sequencing (KiCS) has recently expanded into the clinical realm with its cancer panel, and Shlien said that RNA-seq will be added soon.

Physicians and clinical geneticists treating any of the approximately 700 patients in KiCS take part in weekly molecular tumor board meetings to discuss findings that may assist in treatment or management of each case.

RACCOON and OTTER will allow the hospital to accelerate the expansion of transcriptomics-based diagnosis.

Shlien said that there is "definitely an interest" in working with the hospital's commercialization office to create commercial versions of the RACCOON and OTTER computational tools. "You can imagine that it would extend well beyond pediatric cancer and that this would be massively useful for rare entities in adults and kids," he said.

He also expects to develop a specific atlas for sarcoma that will be supplemented with perhaps 2,000 bone and soft-tissue tumors.

One potential stumbling block to universal adoption of the process is the fact that the SickKids team worked mostly with frozen tumors, while the norm is formalin-fixed, paraffin-embedded (FFPE) tumors.

The paper's authors wrote that "most molecular-based analyses seem incompatible with FFPE data" because DNA and RNA degrade in FFPE samples over time.

Katherine Janeway, director of clinical genomics at Dana-Farber Cancer Institute and a pediatrician at Boston Children's Hospital, also raised this issue.

"Do you need frozen tumor in order to do this? And if so, what's the process of changing pathology department policies and procedures to make sure that we obtain that tissue and properly preserve it at the time of biopsy?" she wondered.

Dana-Farber has mostly been using RNA sequencing to detect gene fusions, but has not really gotten into whole-transcriptome analysis because it is "not standard to pathology processing" in the US, according to Janeway.

Still, she applauded the SickKids team for advancing the subclassification of childhood tumors.

"This tool that they've created is a major advance that will help us get to using whole transcriptome in pediatric cancer diagnosis and cell classification," said Janeway, who was the lead author of a 2013 paper in the Journal of Clinical Oncology that discussed the "future of clinical genomics" for childhood cancer. That paper, which Shlien and colleagues cited in their work, looked at clinical trial design and drug development, not clinical applications.

"This preliminary RNA data that they present is quite compelling that transcriptome may be very helpful in subclassification in the future," she said of the new research. "Hopefully, this will inspire people to generate more whole-transcriptome data from pediatric cancers and to use the tool and the classification system that they've created to further assess the clinical contribution of this approach."

Janeway expects there to be other, similar transcriptome-based classification tools for clinicians to turn to in the very near future. She said that her program has a similar project that hasn't yet published to examine molecular diagnoses of rare and ultrarare pediatric cancers.

Jinghui Zhan, chair of computational biology at St. Jude, also said he is developing a similar classifier, but one that goes beyond the SickKids work by including "total" RNA-seq, not just mRNA-seq. The SickKids classifier was trained only on messenger RNA data.

Janeway expressed confidence that whole-transcriptome analysis will eventually become a standard part of diagnosis and subclassification of pediatric cancers.