DSAI-14 MULTI-OMICS GRAPH NEURAL NETWORK INTEGRATION REVEALS PREDICTIVE SIGNATURES OF BRAIN METASTATIC PENETRANCE

Abstract

Brain metastasis remains a major oncological challenge, where tumor cells evolve a specific, multi-layered genetic and epigenetic makeup to breach the blood-brain barrier (BBB) and colonize the central nervous system. Existing single-omics approaches fail to capture this complex molecular evolution, limiting their predictive and translational utility. To address this, we developed a graph neural network (GNN)-based multi-omics integration framework to predict brain metastatic potential using Cancer Cell Line Encyclopedia (CCLE) molecular profiles linked to MetMap annotations. Brain penetrance was modeled as a binary classification task by integrating five complementary modalities: RNA-seq, reduced representation bisulfite sequencing (RRBS) methylation, microRNA expression, damaging mutations, and transcriptome-derived gene-set enrichment scores. Following modality-specific preprocessing and principal component analysis to retain 50 components for high-dimensional data, unsupervised integrated embeddings were learned using the GNN framework. Downstream classification utilized logistic regression under five-fold stratified cross-validation, preserving penetrance balance and tissue composition. Evaluating 423 cell lines at the intersection of all modalities, our multimodal model achieved a mean test ROC-AUC of 0.710. Demonstrating robust generalization with a minimal train-test gap of 0.032, the integrated multimodal approach outperformed the best single-modality GNN baseline (RNA-seq alone: ROC-AUC 0.690). These findings demonstrate that deep learning can successfully decode the complex, multi-omics molecular patterns driving brain organotropism. Ultimately, this robust in silico framework has the potential to complement preclinical and clinical research; it can be used to develop translational tools that inform predictive decision-making based on patient-derived cell lines or in vivo animal models, identify critical molecular drivers of BM, and uncover targets for potential therapeutics.