Extended Data Fig. 8: Characterization of the single-cell gene expression prediction model based on CHARM multi-omics data.
From: Gene regulatory landscape dissected by single-cell four-omics sequencing
a, Distribution of Pearson’s correlation coefficients between predicted and observed gene expression on the test dataset, compared to a shuffled control. b, Comparison of Spearman correlation (ρ) of gene expression prediction between CHARM and SCARlink, computed on the 969 genes that passed method-specific QC in both pipelines. c-e, Box plots showing model performance according to gene expression level (c), gene length (d), and expression variability (e). f, Precision-recall curve evaluating E–P linkage predictions using Shapley values, compared with a correlation-based method, the ABC model and ENCODE rE2G on a gold standard set of 180 validated E–P pairs. Area under the Precision-recall curve (AUPRC) values are indicated. g, Bar plot showing the number of significant CRE-gene linkages identified in each cell type. h, Venn diagram showing the number of predicted enhancer–gene pairs identified using different combinations of modalities. i, Histogram of the genomic distance between CREs and their linked genes. The dashed line marks 200 kb. j, Histogram showing the number of TSS bypassed by distal intergenic enhancers to reach their target gene. The dashed line marks the threshold for enhancers that bypass more than one TSS. k-m, Box plots showing the maximum Shapley values across cell types for chromatin accessibility (k), H3K27ac (l) and chromatin interaction strength (m), stratified by genomic context (within gene vs. intergenic) and binned by linear distance to the TSS. In panels c-e and k-m, boxes indicate the median and 25th-75th percentiles, with whiskers extending to 1.5× the interquartile range.
