Helper Module for Deep Learning.

Pathway Modules Variational Auto-Encoder (pmVAE).

[1] pmVAE: Learning Interpretable Single-Cell Representations with Pathway Modules, Gilles Gut, biorxiv 2021.

Code: https://github.com/ratschlab/pmvae

class pynet.models.vae.pmvae.PMVAE(membership_mask, latent_dim, hidden_layers, bias_last_layer=False, add_auxiliary_module=True, terms=None, activation=None)[source]¶

__init__(membership_mask, latent_dim, hidden_layers, bias_last_layer=False, add_auxiliary_module=True, terms=None, activation=None)[source]¶

pmVAE constructs a pathway-factorized latent space.

Parameters

membership_mask: bool array (pathways, genes)

a binary mask encoding which genes belong to wich pathways.

latent_dim: int

the dimension of each module latent space.

hidden_layers: list of int

the dimension of each module encoder/decoder hidden layer.

bias_last_layer: bool, default False

use a bias term on the final decoder output.

add_auxiliary_module: bool, default True

include a fully connected pathway module.

terms: list of str (pathways, ), default None

the pathway names.

activation: klass, default None

the activation function.

static build_base_masks(membership_mask, hidden_layers, latent_dim)[source]¶

Builds the masks used by the encoders/decoders.

Parameters

membership_mask: bool array (pathways, genes)

a binary mask encoding which genes belong to wich pathways.

latent_dim: int

the dimension of each module latent space.

hidden_layers: list of int

the dimension of each module encoder/decoder hidden layer.

Returns

base: list of array

pathway mask assigns genes to pathway modules, and separation masks keep modules separated. Encoder modifies the last separation mask to give mu/logvar, and the decoder reverses and transposes the masks.

static build_decoder(membership_mask, hidden_layers, latent_dim, activation, batch_norm=True, bias_last_layer=False)[source]¶

Build the decoder/merger modules.

static build_encoder(membership_mask, hidden_layers, latent_dim, activation, batch_norm=True)[source]¶

Build the encoder module.

static build_module_isolation_mask(nmodules, module_output_dim)[source]¶

Isolates a single module for gradient steps. Used for the local reconstruciton terms, drops all modules except one.

static build_pathway_mask(nfeats, membership_mask, hidden_layers)[source]¶

Connects genes to pathway modules. Repeats the membership mask for each module input node. See M in Methods 2.2.

static build_separation_mask(input_dim, out_put_dim, nmodules)[source]¶

Removes connections betweens pathway modules. Block diagonal matrix, see Sigma in Methods 2.2.

decode(z)[source]¶

Computes the generative distribution p(x | z).

Parameters

z: torch.Tensor (batch_size, latent_dim)

the stochastic latent state z.

Returns

p(x | z): @callable

the distribution p(x | z) with shape (batch_size, data_size).

encode(x)[source]¶

Computes the inference distribution q(z | x).

Parameters

x: torch.Tensor (batch_size, data_size)

the input data.

Returns

q(z | x): @callable

the distribution q(z | x) with shape (batch_size, latent_dim.

forward(x)[source]¶

The forward method.

get_masks_for_local_losses()[source]¶

Get module/pathway associated masks.

kernel_initializer()[source]¶

Init network weights.

latent_space_names(terms=None)[source]¶

Get latent space associated names.

reparametrize(mu, logvar)[source]¶

Implement the reparametrization trick.