Skip to main content

Table 3 Description of parameter sweep

From: Tuning parameters of dimensionality reduction methods for single-cell RNA-seq analysis

Method Parameters Values
scran Size factors normalization {True, False }
  ERCC counts normalization {True, False }
  Assay type {logcounts, counts }
  High variance genes { 100, 300, 500, 1000, 2000, 3000 }
  Dimension of latent space { 2, 8, 10, 16, 32, 50, 64, 128 }
Seurat Normalization method {LogNormalize, CLR }
  Criteria for high variance genes {vst, mvp, dist }
  High variance genes { 100, 300, 500, 1000, 2000, 3000 }
  Dimension of latent space { 2, 8, 10, 16, 32, 50, 64, 128 }
ZinbWave Gene covariates { True, False }
  Epsilon (regularizer) { 200, 500, 1000, 2000 }
  High variance genes {100, 300, 500, 1000, 2000, 3000 }
  Dimension of latent space {2, 8, 10, 16, 32, 50, 64, 128 }
DCA Dispersion and reconstruction {zinb-conddisp, zinb, nb-conddisp, nb }
  Batch normalization {True, False }
  Dimension of the latent space { 2, 8, 10, 16, 32, 50, 64, 128 }
  Number of training epochs { 20, 50, 100, 200, 300, 500, 1000 }
  Normalize counts {True, False }
  Scale variance {True, False }
  Log normalization {True, False }
  Dropout rate {0, 0.1 }
  Number of hidden neurons {64, 128, 256 }
  Random seed {0, 1, 2, 3, 4 }
scVI Number of hidden neurons { 64, 128, 256 }
  Number of training epochs {20, 50, 100, 200, 300, 500, 1000 }
  Learning rate { 1e-2, 1e-3, 1e-4 }
  Dropout rate { 0, 0.1 }
  Layers {1, 2 }
  Dimension of the latent space { 2, 8, 10, 16, 32, 50, 64, 128 }
  Dispersion {gene, gene-cell }
  Reconstruction loss { nb, zinb }
  Random seed {0, 1, 2, 3, 4 }
  1. For each method (first column), we vary a number of tuneable parameters (second column) systematically over a grid of values (third column). The bold value in the third column is the default value