Fig. 3: Performance evaluation on cell-lineage-specific GRN construction.

The performance of GRN construction on different benchmark scRNA-seq datasets, measured in terms of AUPRC (a) and EPR (b), respectively. Five baselines were considered for comparison, including SCINET⁴², GRNBoost2²⁹, DeepSEM²⁸, NetREX³⁰, CellOracle⁴³ and Random_NicheNet (i.e., randomly selecting edges from the prior gene interaction network). Bars and error bars signify mean ± s.d. over n = 20 independent experiments. c The comparison between the CEFCON constructed GRN and the randomized GRN derived from the prior gene interaction network, in terms of R² values and clustering coefficients. The R² is the coefficient of determination for the linear regression model to measure how close the data points are with respect to the fitted linear line. Each dot represents one of the seven datasets. Box plots show the medians (central lines) and interquartile ranges, and the whiskers represent 1.5 × interquartile ranges. Statistical significance was calculated using the one-sided Wilcoxon signed-rank test. d Degree distribution and major topological properties of the GRNs constructed by CEFCON on the hESC dataset. The x-axis represents the network degree (denoted as k) and y-axis represents the frequency of the network degree k (denoted as P(k)). Both k and P(k) are log-transformed. The major network topological properties, including the number of genes, the number of edges, the slope and the R² of the degree distribution, and the average clustering coefficient, are listed in the upper right corner. The inset in the bottom left corner provides the degree distribution (also with the slope, R² and average clustering coefficient) of the randomized GRN derived from the prior gene interaction network. Source data are provided as a Source Data file.