Predicting Response and Prognosis in Pediatric Cancers

Division Of Clinical Sciences - Nci

Investigators

Linked publications, trials & patents

Abstract

The project is ongoing, data is being collected and integrated with the other projects in the Oncogenomics Section. We are developing a fully integrated RMS risk prediction engine leveraging clinical, molecular, and histologic data. We will validate and improve previous convolutional neural networks (CNNs) models using hematoxylin and eosin (HandE) images of RMS to predict diagnosis, molecular classification, and outcome. We will harness advanced large language models (LLMs), including transformers, for HandE images to significantly enhance the CNN predictive models. Additionally, we will generate attention heatmaps to analyze the morphological features that play significant roles in the classification and prediction of RMS subtypes, the presence of driver mutations, or prognosis. We will develop multimodal artificial intelligence (AI) classifiers for EFS and OS prediction by integrating clinical variables, germline genetics, somatic alterations, CNN and LLM models of HandE images, and circulating tumor DNA analyses. Cox regression models will be applied to study relationships between the multimodal AI risk prediction and time to event. Building on our recent work, CNN-based classifiers trained on whole-slide HandE images of RMS tumors have already demonstrated superior predictive performance compared to current risk stratification schemas. Our prior studies showed that CNNs could distinguish high-risk molecular alterations-such as RAS-pathway mutations, MYOD1 and TP53 alterations-and stratify patients by event-free and overall survival with higher accuracy than traditional clinical or histopathological methods . We aim to extend these models to identify critical molecular features such as FOXO1 fusion status, MYOD1 mutations, and TP53 alterations-mutations shown to be strongly prognostic in both fusion-negative and fusion-positive RMS in recent large genomic analyses . These molecular features are frequently misclassified or missed altogether in routine diagnostics, yet they carry clear prognostic and therapeutic implications. Additionally, we are integrating germline genomic data from recent cohorts, which have revealed that approximately 7-8% of pediatric RMS patients carry pathogenic or likely pathogenic (P/LP) germline variants in cancer susceptibility genes (CSGs), including TP53, DICER1, NF1, BRCA2, and others . These germline alterations often co-segregate with early age of onset and can guide cascade testing, surveillance, and therapeutic decision-making. Incorporating these findings, our project will explore the predictive and prognostic synergy of somatic and germline genomic alterations. We are developing models that combine histomorphologic patterns extracted from HandE images with deep genomic insights to refine personalized risk stratification for RMS patients. Specifically, we will test whether germline P/LP variants, which are not currently part of clinical risk stratification, add independent prognostic value when integrated with molecular and imaging-based predictions. By unifying pathology image analysis, germline and somatic genomics, and liquid biopsy data via AI, this project aspires to build the most comprehensive risk prediction framework yet developed for pediatric rhabdomyosarcoma. Our approach will facilitate individualized risk stratification at diagnosis, which is urgently needed given the clinical heterogeneity of RMS-especially within the intermediate-risk group, where outcomes remain highly variable despite standardized treatment protocols . Ultimately, this work will inform prospective clinical trial designs and has the potential to identify new molecular subtypes of RMS defined by histopathologic, genomic, and AI-inferred features. These new subtypes may serve as the basis for novel therapeutic approaches, including trials of targeted therapies for MYOD1 or TP53 mutant tumors, or surveillance strategies for children carrying high-risk germline alleles.