Table 2.

Summary of evaluation metrics for each method in test data of each dataset.^a

[ACC, F1, AUC]	Buettner	Kolodziejczyk	Pollen	Usoskin	Zeisel	Cortex
SIMLR	[0.978, 0.978, 0.990]	[0.999, 0.999, 1.000]	[0.905, 0.847, 0.923]	[0.925, 0.887, 0.933]	[0.929, 0.806, 0.940]	[0.915, 0.895, 0.970]
CALLR	[0.314, 0.289,—]	[0.961, 0.960,—]	[0.784, 0.770,—]b	[0.946, 0.940,—]	[0.938, 0.934,—]	[0.943, 0.942,—]
scSemiGAN	[0.512, 0.501, 0.690]	[0.994, 0.994, 0.997]	[0.932, 0.935, 0.993]	[0.959, 0.958, 0.986]	[0.896, 0.873, 0.970]	[0.950, 0.949, 0.989]
scSemiAE	[0.605, 0.512,—]	[0.976, 0.976,—]	[0.822, 0.811,—]	[0.729, 0.714,—]	[0.912, 0.900,—]	[0.940, 0.940,—]
GCN	[0.849, 0.847, 0.900]	[0.997, 0.997, 1.000]	[0.915, 0.909, 0.984]	[0.929, 0.929, 0.971]	[0.901, 0.896, 0.980]	[0.608, 0.583, 0.840]
GAT	[0.791, 0.784, 0.840]	[0.976, 0.976, 0.988]	[0.856, 0.833, 0.968]	[0.844, 0.841, 0.900]	[0.908, 0.903, 0.970]	[0.940, 0.940, 0.986]
scSemiGCN	[0.977, 0.977, 0.983]	[1.000, 1.000, 1.000]	[0.983, 0.980, 1.000]	[0.949, 0.948, 0.977]	[0.928, 0.925, 0.970]	[0.953, 0.953, 0.984]

^aThe best are indicated in blue font.

^bWe used 10% of annotated cells in training instead of 5% such that there are at least two labeled samples for each cell type.