Transcriptome-based machine learning classification of pediatric germ cell tumors using nanopore RNA sequencing

Childhood cancer comprises a heterogeneous group of diseases, with disproportionately higher mortality in low- and middle-income countries (LMICs) due to limited access to molecular diagnostics. Pediatric germ cell tumors (GCTs) are rare and represent 3% of childhood cancers, showing substantial clinical and biological heterogeneity, which makes accurate subtype classification challenging. As part of an ongoing multicentric collaboration, we evaluated nanopore whole-transcriptome sequencing as a scalable approach for differentiating pediatric GCTs from other solid tumors and distinguishing their histological subtypes using a nanopore-based transcriptomic classifier. Nanopore RNA-seq was performed on 56 pediatric tumors, including 14 germinomas/dysgerminomas/seminomas, 13 yolk sac tumors, 11 mature teratomas, seven immature teratomas, eight mixed tumors, and three embryonal carcinomas. A machine learning model was trained using their gene expression profiles and followed by independent validation. Transcriptome data enabled molecular classification, achieving 98.9% overall accuracy in distinguishing GCTs from other previously sequenced pediatric solid tumors, and 100% accuracy above the established confidence threshold of 0.8. GCT subtype classification achieved 90.3% overall accuracy and 89.8% above 0.8 prediction probability. The model showed high F1-scores and AUC values above 0.86. Consistent performance was maintained in an independent cohort, and transcriptomic profiles derived from FFPE samples showed high concordance with cryopreserved tissues, supporting the method’s applicability in routine clinical specimens. These results demonstrate the diagnostic potential of the developed pipeline for GCT subtyping and support its future application in decentralized or resource-constrained healthcare systems. By enabling transcriptome-based classification with minimal infrastructure, nanopore platforms offer a promising solution for improving cancer diagnostics in LMICs.