Skip to main content

Table 3 Details of hyperparameters for DR-A based on the experimental results in Table 2. We carried out the experiments using the Rosenberg-156 k, Zheng-73 k, Zheng-68 k, Macosko-44 k, and Zeisel-3 k datasets. The DR-A algorithm was investigated with (a) 2 latent dimensions (K = 2), (b) 10 latent dimensions (K = 10), and (c) 20 latent dimensions (K = 20). G denotes a generative model and D denotes a discriminative model

From: A deep adversarial variational autoencoder model for dimensionality reduction in single-cell RNA sequencing analysis

DatasetBatch sizeHidden layerHidden unitLearning rate
(a) K = 2
 Rosenberg-156 k1284G: 1024/512/512/256
D: 32/16/16/8
7 × 10−5
 Zheng-73 k1283G: 512/512/512
D: 32/32/32
6 × 10−5
 Zheng-68 k1284G: 256/256/256/256
D: 32/32/16/16
0.0001
 Macosko-44 k1283G: 256/128/64
D: 64/64/64
0.0001
 Zeisel-3 k1284G: 512/512/512/512
D: 32/32/32/32
8 × 10−4
(b) K = 10
 Rosenberg-156 k1284G: 512/256/128/64
D: 256/128/64/32
6 × 10−5
 Zheng-73 k1284G: 1024/512/512/256
D: 32/32/32/32
2 × 10−5
 Zheng-68 k1284G: 256/256/256/256
D: 32/32/16/16
7 × 10−5
 Macosko-44 k1284G: 512/256/256/128
D: 256/128/128/64
7 × 10−5
 Zeisel-3 k1281G: 512
D: 512
7 × 10−4
(c) K = 20
 Rosenberg-156 k1284G: 1024/1024/1024/1024
D: 64/64/64/64
6 × 10−5
 Zheng-73 k1284G: 1024/512/512/256
D: 64/32/32/16
1 × 10−5
 Zheng-68 k1281G: 256
D: 256
2 × 10−5
 Macosko-44 k1281G: 256
D: 256
7 × 10−5
 Zeisel-3 k1281G: 512
D: 512
7 × 10−4