Fig. 4: Extensive performance comparisons at different resolutions based on diverse input datasets and gold-standard benchmarks support Tensor-FLAMINGO’s superior performance in reconstructing single-cell 3D chromatin structures.

a Performance comparison at 10kb-resolution across four different input single-cell datasets (GM12878 Dip-C dataset n = 14, K562 snHi-C dataset n = 16, mESC scHi-C dataset n = 8, and mESC snm3C dataset n = 351). The chromatin contact maps (Hi-C and 3D ATAC-PALM) profiled from the matching cell types are used as gold standards. Spearman correlations are used as the performance metric, including both global accuracy for all pairwise distances in the gold-standard datasets and the accuracy for the subset of pairwise distances also observed in the input datasets. The error bars represent the standard deviations across single cells. b Performance comparison at 30kb-resolution across four different input single-cell datasets (GM12878 Dip-C dataset n = 14, K562 snHi-C dataset n = 16, mESC scHi-C dataset n = 8, and mESC snm3C dataset n = 351). The predictions of NucDynamics provided by Si-C are directly used. The chromatin contact maps (Hi-C, GAM and 3D ATAC-PALM) profiled from the matching cell types are used as gold standards. Spearman correlations are used as the performance metric, including both global accuracy for all pairwise distances in the gold-standard datasets and the accuracy for the subset of pairwise distances also observed in the input datasets. The error bars represent the standard deviations across single cells. c UMAP visualization of the reconstructed single-cell structures from the snm-3C dataset, which contains two cell types: mESC and NMuMG. Tensor-FLAMINGO’s predictions capture the cell-type specific structural signatures and lead to clear separations between the two cell types, while other methods result in highly mixed cells of the two cell types. Source data are provided as a Source Data file.