After accidents, cancer is the leading cause of child death in the developed world.  Choosing the right treatment for pediatric tumors relies on first having an appropriate, specific diagnosis to guide therapy.  Adult cancers are diagnosed on the basis of their appearance under the microscope (histopathology), clinical presentation, and molecular features.  Unlike adults, most childhood tumors are not caused by a lifetime accumulation of mutations or of toxin exposure.  Instead, pediatric tumors may be caused by changes in normal developmental programs, with activation of the wrong program in place and time.  Despite the fundamental nature of pediatric tumors as diseases of development, we lack systematic characterizations of the developmental programs active in each tumor.  Additionally, some pediatric cancers are difficult to diagnose altogether, either because they display features of multiple tumor types or do not have definitive features using current diagnostics.  New tools are needed for diagnosing pediatric tumors that incorporate the most pertinent information to guide therapy. 

Data science approaches such as machine-learning (‘artificial intelligence’) tools have provided new insights into a vast array of problems in human health and biology ranging from protein folding to image analysis, but have not been applied as widely in child health.  The goal of the present proposal is to systematically identify the developmental programs active in major types of pediatric tumors, construct a machine-learning classifier that uses this information for optimal diagnosis, and to apply the classifier to the most difficult diagnostic cases in pediatric oncology.  Developmental profiling may provide a novel approach for classifying pediatric tumors, and new machine learning tools will provide a useful adjunct to helping diagnose and treat pediatric cancers.