Diagnostic classification of childhood cancer using multiscale transcriptomics

Acknowledgements: The KiCS program is supported by the Garron Family Cancer Centre at The Hospital for Sick Children through funding from the SickKids Foundation. A.H. received funding from the Canadian Institutes for Health Research (grant no. 162267) and is the Tier 1 Canada Research Chair in Rare Childhood Brain Tumors. D.M. is supported by the CIBC Children’s Foundation Chair in Child Health Research. A.S. is partially supported by an Early Researcher Award from the Ontario Ministry of Research and Innovation; by the Canada Research Chair in Childhood Cancer Genomics; and by funding from the V Foundation and the Robert J. Arceci Innovation Award from the St. Baldrick’s Foundation. We would like to thank the Centre for Applied Genomics, The Hospital for Sick Children, for assistance with RNA sequencing and the Treehouse Childhood Cancer Initiative, University of California, Santa Cruz, for access to their public data repository.
The causes of pediatric cancers' distinctiveness compared to adult-onset tumors of the same type are not completely clear and not fully explained by their genomes. In this study, we used an optimized multilevel RNA clustering approach to derive molecular definitions for most childhood cancers. Applying this method to 13,313 transcriptomes, we constructed a pediatric cancer atlas to explore age-associated changes. Tumor entities were sometimes unexpectedly grouped due to common lineages, drivers or stemness profiles. Some established entities were divided into subgroups that predicted outcome better than current diagnostic approaches. These definitions account for inter-tumoral and intra-tumoral heterogeneity and have the potential of enabling reproducible, quantifiable diagnostics. As a whole, childhood tumors had more transcriptional diversity than adult tumors, maintaining greater expression flexibility. To apply these insights, we designed an ensemble convolutional neural network classifier. We show that this tool was able to match or clarify the diagnosis for 85% of childhood tumors in a prospective cohort. If further validated, this framework could be extended to derive molecular definitions for all cancer types.