Sparse weighted graph convolutional networks for colorectal cancer classification using Hi-C data

Abstract Colorectal cancer remains a leading cause of global mortality, driving demand for precise screening methodologies that leverage complex genomic architectures. The functional interplay between spatial chromatin organization and regulatory networks is increasingly recognized as an important component of malignant transformation. Here, we present the sparse weighted graph convolutional network (SW-GCN), a specialized architecture designed to exploit the quantitative and sparse nature of 3D genomic interactions. Our methodology constructs biologically informed sparse weighted graphs by identifying genomic bins with significant contact frequencies using a bi-square kernel and adaptive optimal bandwidth selection. This approach filters stochastic noise while preserving essential spatial patterns in high-throughput chromosome conformation capture (Hi-C) data. We evaluated SW-GCN using a dataset of 102 individuals, focusing on chromosome 18 because of its established association with recurrent structural variations in colorectal cancer. A rigorous nested 10-fold cross-validation protocol was employed to prevent information leakage and obtain unbiased performance estimates. SW-GCN achieved a pooled bootstrap accuracy of 92.2% [95% CI: 87.3%, 97.1%], F1-score of 94.4% [90.6%, 97.9%], and sensitivity of 94.4% [88.7%, 98.6%], outperforming conventional GCN while showing comparable performance to convolutional neural network, graph sample and aggregate, and graph attention network, with a balanced sensitivity-specificity profile. The framework also reduced training latency by approximately 36% compared with conventional GCN. These findings support biologically informed edge weighting and graph sparsification as a computational framework for leveraging 3D genome architecture and motivate future investigation of 3D-genome biomarkers in precision oncology.