FRANK: a pan-cancer RNA-seq classifier for childhood tumors with data-efficient learning

Abstract

Accurate characterization of tumor subtype is of paramount importance in guiding treatment decisions in pediatric cancers. Here, we present FRANK (Fully-connected RNA-based Augmentation Neural Klassifier), a machine learning approach for classifying pan-cancer pediatric tumors. We curated a set of 11,467 transcriptomes to build a high-quality training dataset representing 181 distinct childhood tumor and normal tissue types. We used extensive data augmentation techniques with a hierarchical penalty term to train a neural network capable of detecting tumor types with as few as three training examples. To evaluate the generalization of the trained model, we perform validation on nine datasets spanning 16,398 samples, achieving an overall accuracy of 98.3%. Furthermore, we demonstrated that FRANK outperformed other published pediatric pan-cancer RNA-seq classifiers on an external dataset. To evaluate the robustness, we showed that FRANK remained accurate when applied to long-read RNA-sequencing data, as well as to noisy and low-quality data.