RARE MUTATIONS GRAPH NEURAL NETWORK MODELS FOR PREDICTING PATHOGENICITY OF RARE CANCER-ASSOCIATED GENOMIC VARIANTS

Abstract

Rare cancer-associated genomic variants are frequently under-determined due to small numbers of patient samples, heterogeneity in molecular characteristics and lack of functional annotations. In this paper, we propose Rare Mu'Graph, a graph neural network-based framework to predict the pathogenicity of rare genomic variants involved in cancer. The proposed approach abstracts both the local features of the variants and the broader relationships in the biological network by representing genes, variants, molecular interactions, pathway memberships and clinical annotations as interconnected graph structures. Rare Mu'Graph combines sequence-derived attributes, mutation context, gene–disease associations, protein interactions and knowledge of cancer pathways to further enhance pathogenicity predictions of variants that are not well represented in standard training sets. The graph neural network models are well suited to this task as they have a capability to learn from the relational dependencies and propagate the biological signals across connected nodes. The framework helps prioritize potentially pathogenic rare variants that can help researchers and clinicians identify candidate mutations for validation. In summary, this research underscores the promise of graph-based deep learning to revolutionize the interpretation of rare variants in precision oncology, particularly in cancers with limited genomic data.