Fig. 5: Random-walk is biased towards embedding high-order contacts.

a Embeddings mouse oocyte/zygote data using scHiCluster, Higashi, and Va3DE with and without random-walk imputation. b Compare the tools in (a) with the mouse embryo dataset. c Summarize the changes of ARI for the analysis in (a, b). n = 5 runs for each method. Data are presented as mean values ± SEM. Two-sided t-test was used to compare raw contact results to random walk results. Oocyte/Zygote: scHiCluster t = 13.19 (p < .00001), Higashi t = 17.9 (p < 0.00001), Va3DE t = 1.53 (p = 0.166). Embryo: scHiCluster t = 31 (p < 0.00001), Higashi t = 18.85 (p < 0.00001), Va3DE t = 1.66 (p = 0.135). Significant p < 0.01 is indicated by * and p > 0.01 is indicated by NS. d Embeddings the mESC cell cycle dataset using scHiCluster, InnerProduct, and fastHiCRep with and without random-walk imputation. e, f Three methods (scHiCluster, InnerProduct, fastHiCRep) are evaluated on the human PFC and mouse hippocampus datasets with and without random-walk. Neuron populations are highlighted with dash circles. g ARI for the neuron cell populations from human PFC and mouse hippocampus datasets with and without random-walk. n = 5 runs for each method. Data are presented as mean values ± SEM. Two-sided t-test was used to compare raw contact results to random walk results. Human PFC Neurons: scHiCluster t = 15.64 (p < 0.00001), InnerProduct t = 9.13 (p = 0.000017), fastHiCRep t = 21.21 (p < 0.00001). Mouse Hippocampus Neurons: scHiCluster t = 7.95 (p = 0.000045), InnerProduct t = 11.51 (p < .00001), fastHiCRep t = 13.71 (p < .00001). Significant p < 0.01 is indicated by * and p > 0.01 is indicated by NS. h Pseudo-bulk compartment analysis of two non-neuron and three neuron cell types in the mouse hippocampus dataset. Note that non-neurons (MGC, ODC) have distinct compartment structure, but neurons (DG, CA1, CA3) are all highly similar. Source data are provided as a Source data file.