Fig. 2

The clustering results and corresponding performance metric values of scGAL, Dhaka, RobustClone, Seurat, scBGEDA, scTAG and AE on the SA501 dataset. The AE model is used as a baseline model that only analyzes single-cell copy number data. The 2-d plots are generated by using t-SNE to project the raw data (recovered genotype matrix is used for RobustClone) or low-dimensional latent representations generated by deep learning methods into a 2D space. A and G show the results of scGAL and AE, respectively, where both scGAL and AE identify 3 subpopulations, whereas scGAL is able to better distinguish the subclones. B-F Clustering results of other methods. All of the existing methods misestimate the number of cell subpopulations except for scBGEDA and scTAG (the ground truth number of clusters is given as part of the input of scBGEDA and scTAG). It should be noted that the genotype matrix recovered by RobustClone shows many cells have the same genotype, resulting in sporadic points on the 2-d plot. H, I ARI and NMI scores of all methods