Abstract
High-throughput single-cell Hi-C (scHi-C) technologies have opened new avenues for investigating cell-to-cell variability in the three-dimensional organization of the genome within individual nuclei. Despite their potential, analyses of scHi-C data are hindered by data sparsity, which varies substantially across cells. To address this challenge, recent methods aim to denoise scHi-C data and differentiate between two types of zero entries: structural zeros (SZs), which reflect true absence of contacts due to biological structure, and dropouts (DOs), which arise from insufficient sequencing depth. However, current dissimilarity measures used in downstream analyses, such as Euclidean distance and Kendall's tau, treat all zeros as equivalent, thus do not distinguish between SZs and DOs nor recognize the special role of SZs in capturing cell-to-cell variability. Such oversight limits the ability to accurately capture biologically meaningful differences between cells. In this study, we introduce structural-zeros-aware Kendall's tau (szKendall), a novel dissimilarity metric that explicitly incorporates the spatial structure of 2D contact matrices and leverages the presence or absence of shared SZs across cells. Through comprehensive simulations and analysis of real scHi-C datasets, we demonstrate that szKendall more effectively captures key structural features and achieves superior performance in cell clustering tasks compared to existing approaches. Our results underscore the importance of SZ-aware dissimilarity measures in advancing single-cell 3D genomics.