fit_vae: Variational autoencoder model fitting

In dongminjung/VAExprs: Genearting Samples of Gene Expression Data with Variational Autoencoders

Description

A fundamental problem in biomedical research is the low number of observations available. Augmenting a few real observations with generated in silico samples could lead to more robust analysis. Here, the variational autoencoder (VAE) is used for the realistic generation of single-cell RNA-seq data. Also, the conditional variational autoencoder (CVAE) can be used if labels of samples are available. This function allows us to fit variational autoencoders with the standard Gaussian prior to expression data. It is assumed that there will likely be no clusters in the latent space representation of variational autoencoders.

Usage

fit_vae(object = NULL,
        x_train = NULL,
        x_val = NULL,
        y_train = NULL,
        y_val = NULL,
        encoder_layers,
        decoder_layers,
        latent_dim = 2,
        regularization = 1,
        epochs,
        batch_size,
        preprocessing = list(
            x_train = NULL,
            x_val = NULL,
            y_train = NULL,
            y_val = NULL,
            minmax = NULL,
            lenc = NULL),
        use_generator = FALSE,
        optimizer = "adam",
        validation_split = 0, ...)

Arguments

object	SummarizedExperiment object
x_train	expression data for train, where each row is a cell and each column is a gene
x_val	expression data for validation, where each row is a cell and each column is a gene
y_train	labels for train
y_val	labels for validation
encoder_layers	list of layers for encoder
decoder_layers	list of layers for decoder
latent_dim	dimension of latent vector (default: 2)
regularization	regularization parameter, which is nonnegative (default: 1)
epochs	number of epochs
batch_size	batch size
preprocessing	list of preprocessed results, they are set to NULL as default x_train : expression data for train x_val : expression data for validation y_train : labels for train y_val : labels for validation minmax : result of min-max normalization lenc : encoded labels
use_generator	use data generator if TRUE (default: FALSE)
optimizer	name of optimizer (default: adam)
validation_split	proportion of validation data, it is ignored when there is a validation set (default: 0)
...	additional parameters for the "fit" or "fit_generator"

Value

model	trained VAE model
encoder	trained encoder model
decoder	trained decoder model
preprocessing	preprocessed results

Author(s)

Dongmin Jung

References

Marouf, M., Machart, P., Bansal, V., Kilian, C., Magruder, D. S., Krebs, C. F., & Bonn, S. (2020). Realistic in silico generation and augmentation of single-cell RNA-seq data using generative adversarial networks. Nature communications, 11(1), 1-12.

See Also

SummarizedExperiment::assay, SummarizedExperiment::colData, scater::logNormCounts, gradDescent::minmaxScaling, keras::fit, keras::fit_generator, keras::compile, CatEncoders::LabelEncoder.fit, CatEncoders::transform, DeepPINCS::multiple_sampling_generator

Examples

### simulate differentially expressed genes
set.seed(1)
g <- 3
n <- 100
m <- 1000
mu <- 5
sigma <- 5
mat <- matrix(rnorm(n*m*g, mu, sigma), m, n*g)
rownames(mat) <- paste0("gene", seq_len(m))
colnames(mat) <- paste0("cell", seq_len(n*g))
group <- factor(sapply(seq_len(g), function(x) { 
    rep(paste0("group", x), n)
}))
names(group) <- colnames(mat)
mu_upreg <- 6
sigma_upreg <- 10
deg <- 100
for (i in seq_len(g)) {
    mat[(deg*(i-1) + 1):(deg*i), group == paste0("group", i)] <- 
        mat[1:deg, group==paste0("group", i)] + rnorm(deg, mu_upreg, sigma_upreg)
}
# positive expression only
mat[mat < 0] <- 0
x_train <- as.matrix(t(mat)) 


### model
batch_size <- 32
original_dim <- 1000
intermediate_dim <- 512
epochs <- 2
# VAE
vae_result <- fit_vae(x_train = x_train,
                    encoder_layers = list(layer_input(shape = c(original_dim)),
                                        layer_dense(units = intermediate_dim,
                                                    activation = "relu")),
                    decoder_layers = list(layer_dense(units = intermediate_dim,
                                                    activation = "relu"),
                                        layer_dense(units = original_dim,
                                                    activation = "sigmoid")),
                    epochs = epochs, batch_size = batch_size,
                    validation_split = 0.5,
                    use_generator = FALSE,
                    callbacks = keras::callback_early_stopping(
                        monitor = "val_loss",
                        patience = 10,
                        restore_best_weights = TRUE))


### from preprocessing
vae_result_preprocessing <- fit_vae(preprocessing = vae_result$preprocessing,
                                    encoder_layers = list(layer_input(shape = c(original_dim)),
                                                        layer_dense(units = intermediate_dim,
                                                                    activation = "relu")),
                                    decoder_layers = list(layer_dense(units = intermediate_dim,
                                                                    activation = "relu"),
                                                        layer_dense(units = original_dim,
                                                                    activation = "sigmoid")),
                                    epochs = epochs, batch_size = batch_size,
                                    validation_split = 0.5,
                                    use_generator = FALSE,
                                    callbacks = keras::callback_early_stopping(
                                        monitor = "val_loss",
                                        patience = 10,
                                        restore_best_weights = TRUE))

dongminjung/VAExprs documentation built on Dec. 20, 2021, 12:13 a.m.