Table 1 List of model variants and their corresponding features. Usage of the interpretability term only applies to siVAE and its variants. A linear decoder is composed of the same number of layers as the non-linear decoder unless specified as single, in which case the latent embedding layer is directly transformed to an output layer
From: siVAE: interpretable deep generative models for single-cell transcriptomes
Model | Interpretability Reg. | Decoder | Observation model |
|---|---|---|---|
siVAE | Yes | Non-linear | Gaussian |
siVAE (γ = 0) | No | Non-linear | Gaussian |
siVAE-linear | No | Linear | Gaussian |
siVAE-NB | Yes | Non-linear | Negative binomial |
VAE | NA | Non-linear | Gaussian |
scVI | NA | Non-linear | Negative binomial |
LDVAE | NA | Linear, single | Negative binomial |
scETM | NA | Linear, tri-factorization | Gaussian |
VAE (linear) | NA | Linear, single | Negative binomial |