Table 2. Average performance of different scoring metrics for clustering four cell types not used in the NN training.

Feature	Homo	Comp	Vmes	ARI	AMI	FM	Average
Original	0.785	0.875	0.82	0.73	0.775	0.841	0.805
pca 2	0.833	0.883	0.854	0.786	0.821	0.873	0.842
tsne 2	0.179	0.169	0.172	0.113	0.127	0.391	0.192
ica 2	0.833	0.882	0.853	0.785	0.82	0.873	0.841
nmf 2	0.659	0.735	0.69	0.592	0.638	0.75	0.677
pca 5	0.754	0.863	0.798	0.685	0.741	0.814	0.776
tsne 5	0.034	0.035	0.034	-0.009	-0.006	0.332	0.07
ica 5	0.662	0.839	0.733	0.615	0.649	0.787	0.714
nmf 5	0.707	0.848	0.76	0.693	0.689	0.824	0.753
pca 10	0.759	0.873	0.805	0.695	0.747	0.823	0.784
tsne 10	0.049	0.043	0.045	0.0	0.004	0.299	0.073
ica 10	0.452	0.733	0.545	0.416	0.431	0.706	0.547
nmf 10	0.717	0.812	0.751	0.676	0.688	0.812	0.743
pca 50	0.692	0.867	0.759	0.637	0.68	0.799	0.739
nmf 50	0.481	0.625	0.532	0.427	0.457	0.681	0.534
pca 100	0.742	0.879	0.795	0.695	0.731	0.829	0.779
nmf 100	0.397	0.635	0.472	0.358	0.37	0.672	0.484
pcaReduce	0.768	0.89	0.821	0.747	0.754	0.849	0.805
SIMLR_20	0.793	0.806	0.799	0.718	0.77	0.82	0.784
SIMLR_30	0.806	0.82	0.811	0.747	0.779	0.838	0.8
SIMLR_40	0.834	0.831	0.83	0.747	0.797	0.835	0.812
SNN-Cliq	0.751	0.905	0.802	0.716	0.726	0.843	0.79
sincera_hc	0.807	0.929	0.856	0.797	0.794	0.884	0.845
Dense 1 layer 100	0.905	0.897	0.9	0.872	0.885	0.915	0.896
Dense 1 layer 796	0.895	0.887	0.89	0.856	0.874	0.904	0.884
Dense 2 layer 796/100	0.892	0.882	0.886	0.854	0.869	0.903	0.881
PPITF 1 layer 696 + 100	0.897	0.896	0.896	0.866	0.882	0.911	0.891
PPITF 2 layer 696 + 100/100	0.906	0.902	0.903	0.874	0.89	0.917	0.899
Dense 1 layer 100 pre-train	0.883	0.871	0.877	0.838	0.858	0.892	0.87
Dense 1 layer 796 pre-train	0.887	0.872	0.879	0.843	0.86	0.896	0.873
Dense 2 layer 796/100 pre-train	0.884	0.875	0.879	0.838	0.862	0.894	0.872
PPITF 1 layer 696 + 100 pre-train	0.878	0.873	0.875	0.83	0.858	0.889	0.867
PPITF 2 layer 696 + 100/100 pre-train	0.9	0.894	0.897	0.87	0.881	0.914	0.893

Results are averaged over 20 clustering experiments (using different random initializations). AMI: adjusted mutual information; ARI: adjusted random index; Comp: completeness; FM: Fowlkes–Mallows; Homo: homogeneity; Vmes: v-measure. Bold represents the highest value for each column.