Hierarchical Meta-Learning for Cancer Pathway Signatures: A Novel Framework for Few-Shot Cancer Type Discovery

Keywords: pan-cancer; classification; meta-learning; few-shot learning

TL;DR: A few-shot meta-learning framework for cancer type discovery

Abstract: Cancer subtype classification remains challenging due to the rarity of certain cancer types and limited labeled data. We introduce a novel hierarchical meta-learning framework that leverages pathway-level gene expression signatures to enable few-shot learning for cancer type discovery. Our approach employs a three-level hierarchy (organ system → histology → molecular subtypes) with pathway-aware attention mechanisms, enabling rapid adaptation to new cancer types with minimal training examples. We evaluate our method on 12,226 samples across 36 cancer types using 32 pathway signatures from The Cancer Genome Atlas (TCGA). Our hierarchical Model-Agnostic Meta-Learning (MAML) architecture achieves 70-100\% accuracy with only 1-10 training examples per cancer type, significantly outperforming traditional transfer learning approaches. Key discoveries include identification of highly discriminative pathways (oxphos\_program, Jak1\_vivo\_ko, proliferating) and quantification of cross-cancer transferability patterns with similarity scores ranging from 0.5-1.0. This work represents the first application of hierarchical meta-learning to cancer genomics, providing both technical advances for few-shot learning and biologically interpretable insights for precision medicine. Our framework enables rapid classification of rare cancer subtypes and discovers transferable pathway biomarkers with direct clinical applications.

Submission Number: 175