ncibtep@nih.gov

Bioinformatics Training and Education Program

Javed Khan: Integrating Genomics and H&E Images to Predict the Molecular Subtype and Survival of Patients with Rhabdomyosarcoma using Deep Learning Algorithms - CANCELLED

Javed Khan: Integrating Genomics and H&E Images to Predict the Molecular Subtype and Survival of Patients with Rhabdomyosarcoma using Deep Learning Algorithms - CANCELLED

 When: May. 19th, 2022 1:00 pm - 2:00 pm

To Know

Where:
Online Webinar
Organizer:
BTEP
Presented By:
Javed Khan (NCI/CCR)
This class has ended.

About this Class

THIS EVENT HAS BEEN CANCELLED
Rhabdomyosarcoma (RMS) is the most common soft tissue sarcoma of childhood and are subdivided into three major histomorphologic subtypes: alveolar (ARMS), embryonal (ERMS), or spindle/sclerosing (SSRMS).  Patients with ARMS histology have a poor outcome relative to ERMS, and molecular studies have found recurrent chromosome rearrangements t(2;13) or t(1;13) which generate PAX3-FOXO1 or PAX7-FOXO1 fusion genes, respectively, in the majority of ARMS.  The PAX3-FOXO1 fusion gene is more common, is the main oncogenic driver, and is associated with poor outcome. For patients with metastatic disease or recurrent RMS, despite aggressive therapy, the 5-year survival rate remains poor. Beyond PAX-FOXO1 fusion status, no genomic markers are available for risk stratification. We first established an international consortium to study the incidence of driver mutations and their association with clinical outcome and identified 641 patients that had sufficient DNA for analyses. A median of 1 mutation was found per tumor. In FOXO1 fusion negative cases (FN), mutation of any RAS pathway member was found in greater than 50% of cases, and 21% had no putative driver mutation identified. We discovered that mutations of MYOD1, TP53, and CDKN2A were associated with  a dismal survival. We next utilized convolutional neural networks (CNNs) to learn histologic features associated with the driver mutations and outcome using hematoxylin and eosin (H&E) images of the diagnostic RMS tumors. The trained CNN could accurately classify ARMS with an ROC of 0.87 on an independent test dataset. CNN models trained on mutationally-annotated samples identified RAS pathway mutations and tumors with high-risk mutations in MYOD1 or TP53 with an ROC of 0.96 and 0.64, respectively. Remarkably, CNN models, were superior in predicting event-free survival compared to current molecular-clinical risk stratification. We thus identify mutations associated with adverse outcome in RMS, allowing for an improved risk stratification, and demonstrate that CNNs are a powerful tool for molecular and prognostic prediction of rhabdomyosarcoma from diagnostic H&E images.