Effectiveness of machine learning at modeling the relationship between Hi-C data and copy number variation
image: Overview of the study design
Credit: Wang, Y, Sun, Y, Liu, Z, Chen, B, Chen, H, Ren C, Lin, X, Hu, P, Jia, P, Xu, X, Xu, K, Liu, X, Li, H, Bo, X
Copy number variation (CNV) refers to an increase or decrease in the number of copies of a DNA sequence in a genome, which can subsequently be implicit in promoting aberrant gene expression patterns contributing to cancer initiation and progression. Identifying CNV from Hi-C data can help provide insight into how CNV affects the 3D interactions between genomic fragments involved in the expression levels of genes and regulatory factors. However, detecting CNV from Hi-C data is challenging due to the sparsity and complex spatial topological structure.
Recently, Quantitative Biology published an approach entitled “Effectiveness of machine learning at modeling the relationship between Hi-C data and copy number variation”, which employed machine-learning method to establish a relationship between Hi-C data and CNV based on the 1D interaction signal of each bin and the spatial information of chromosomes precisely and performed a series of experiments to comprehensively evaluate the utility and robustness of the graph convolutional networks model.
It predicts CNV events from Hi-C data by linear model and graph neural network for modeling the complex interactions and hierarchical relationships in a graphical structure, respectively (Figure 1). Specifically, it constructed the linear model and applied it to a chromosome (Figure 1A). One weight-shared linear model is applied to infer the CNV across the different chromosomes with the dimension reduction method (Figure 1B). For exploring the relationships between the Hi-C data and CNV to understand the mechanism behind chromosome spatial interactions, it trained a GCN-based model to effectively capture the features of the chromatin structure and predict the CNV of each bin on the Hi-C interaction map (Figure 1C) while performing a series of Hi-C data perturbation experiments to evaluate the ability of the GCN-based model to obtain critical features of the chromosomal structure related to CNV (Figure 2D).