PAX3/7::FOXO1 fusion detection and transcriptomic prediction from whole-slide images of rhabdomyosarcoma using attention-based deep learning frameworks: A multi-institutional study

Background: Rhabdomyosarcoma (RMS) is a highly malignant pediatric soft-tissue sarcoma where molecular subtyping, particularly PAX3/7::FOXO1 fusion status, drives prognosis and treatment. However, histology-based diagnostic approaches remain limited by subjectivity and the scarcity of comprehensive molecular annotations. To overcome these challenges, we improved our previously reported convolutional neural network learning models that predict PAX3/7::FOXO1 fusion status from whole-slide images (WSIs) while additionally trained models to infer gene expression profiles from histology, thereby linking morphology to transcriptomic signatures. Methods: A total of 826 independent WSIs from three sources [Children's Oncology Group (COG) biobanking protocols = 322, Kids First (KIDS) = 252, Childhood Cancer Data Initiative/Molecular Characterization Initiative (CCDI/MCI) = 252] were used to train and evaluate an Attention-Based Multiple Instance Learning (ABMIL) model using UNI2-h foundation features for fusion classification. For gene expression prediction, 135 RMS WSIs paired with bulk RNA-seq data were used to fine-tune a SEQUOIA transformer model, which was trained on TCGA UCEC/COAD datasets. Model performance was evaluated using the Matthews Correlation Coefficient (MCC), AUC, and Pearson's r correlation, with biological validation through pathway enrichment analysis. Results: The fusion detection model achieved robust and generalizable performance across independent test cohorts (MCC ≥ 0.80, AUC ≥ 0.94), with multi-institutional training improving external generalization (MCC up to 0.84). The gene expression model reliably predicted bulk transcriptomic profiles from WSIs (mean r > 0.6, p < 0.05), identifying biologically meaningful pathways including cell cycle and muscle development. Together, these models demonstrate the feasibility of integrating morphological imaging data to gain molecular insights that would not be possible with histology alone. Conclusions: This work presents the first large-scale validated deep learning framework for simultaneous molecular subtyping and transcriptomic inference in RMS. By combining digital pathology with molecular prediction, our approach offers a scalable, tissue-sparing, and generalizable tool for advancing precision oncology in RMS.