Fig. 2: CTC-Tracer is able to distinguish CTCs from blood cells correctly.

a CTC-Tracer is able to distinguish CTCs from blood cells as a binary classification task. b Receiver operating characteristic (ROC) curves for the task of PBMCs identification (mean ± SD, n = 5 independent experiments; two-side Mann–Whitney U-test was used, no adjustment method for multiple comparisons was used). c ROC curves for the task of blood cell identification (mean ± SD, n = 5 independent experiments; two-side Mann–Whitney U-test was used, no adjustment method for multiple comparisons was used). d The t-distributed stochastic neighbor embedding (t-SNE) 2D embedding of cells (372 CTCs and 400 PBMCs) after domain adaptation. The processed PBMCs from the source samples (dots colored in light red) and target samples (triangles colored in dark red) are evenly mixed and can be clearly distinguished from cancer cells (colored in blue). e The t-SNE 2D embedding of cells (372 CTCs and 800 blood cells) after domain adaptation. The processed blood cells from the source samples (dots colored in light red) and target samples (triangles colored in dark red) are evenly mixed and can be clearly distinguished from cancer cells (colored in blue). f Functional enrichment analysis results on the top 30 differentially expressed genes (One-tailed hypergeometric test was used for statistics test, Benjamini–Hochberg adjustment method was used for multiple comparisons). g Expression of the top 5 upregulated and downregulated genes. h Stable NKAIN4 overexpression of pro-metastasis esophageal cancer cells KYSE150 were seeded into a 6-well plate with confluent monolayers and scarred; repair was monitored microscopically at 0 h and 12 h (mean ± SEM, n = 3 biologically independent samples for control group, n = 3 biologically independent samples for experiment group; **p (0.0058) < 0.01, unpaired two-sided t-test was used, no adjustment method for multiple comparisons was used).