Combining gene ontology with deep neural networks to enhance the clustering of single cell RNA-Seq data

BMC Bioinformatics

Table 4 Average ARI score of 10 experiments compared with other methods on mus musculus dataset

Number of clusters	2	4	6	8	avg1	10	12	avg2
NN(dense)	0.9562	0.8231	0.6909	0.6832	0.7884	/	/	/
NN(ppi/tf)	0.9288	0.7983	0.7077	0.6630	0.7745	/	/	/
GONN	0.977	0.9199	0.7934	0.7599	0.8626	/	/	/
PCA2	0.9583	0.606	0.49	0.4489	0.6258	0.4184	0.3819	0.5505
ICA2	0.8296	0.5798	0.4786	0.4656	0.5584	0.4293	0.4026	0.5209
t-SNE2	0.4072	0.5223	0.5413	0.596	0.5167	0.5758	0.5725	0.5359
PCA10	0.9583	0.6373	0.5926	0.6073	0.6989	0.5761	0.5604	0.6553
ICA10	0.0535	0.4445	0.4119	0.5502	0.365	0.5231	0.5053	0.4148
PCA100	0.8707	0.7749	0.5792	0.5634	0.6971	0.5428	0.6031	0.6557
ICA100	0.281	0.075	0.0098	0.0307	0.0834	0.0324	0.0694	0.0726
ZIFA	-0.0143	0.2115	0.4275	0.5847	0.3024	0.6151	0.6212	0.4076
pcaReduce	0.6476	0.5604	0.5358	0.4777	0.5553	0.4399	0.3888	0.5084
DAE	0.9758	0.8435	0.718	0.698	0.8088	0.6226	0.584	0.7403
GOAE	0.967	0.8614	0.7875	0.7381	0.8385	0.6401	0.6085	0.7671

The number after other unsupervised methods means n components. i.e.PCA2 means using PCA method which we set 2 components. For DAE model, we set epoch number as 200 and learning rate as 1e-3. For GOAE model, we set epoch number as 100 and learning rate as 1e-3. The parameters in other NN-based models are shown in Table 2. Avg1 is the average ARI score of the formal four cluster results, while avg2 ARI score is the average of all 2,4,6,8,10 and 12 cluster results. The highest values are shown in boldface

ISSN: 1471-2105