DB-AAE (Adversarial Autoencoder with dynamic batching )
Abstract
By revealing the transcriptome of individual cells, single-cell RNA sequencing (scRNA-seq) offers insights into cellular heterogeneity and disease mechanisms. However, challenges such as low capture rates and dropout events can introduce noise in data analysis. Here, we propose a generative framework, the dynamic batching adversarial autoencoder (DB-AAE), which excels at denoising in scRNA-seq datasets. DB-AAE directly captures optimal features from input data and enhances feature preservation, including cell type-specific gene expression patterns. Comprehensive evaluation on simulated and real datasets demonstrates that DB-AAE outperforms other methods in denoising accuracy and biological signal preservation. It also improves the accuracy of other algorithms in establishing pseudo-time inference. This study highlights DB-AAE's effectiveness and potential as a valuable tool for enhancing the quality and reliability of downstream analyses in scRNA-seq research.
Dataset
Spacial
Gdrive : https://drive.google.com/drive/folders/1DwRpIEStBTQ7XfAel7Qq1uJeCRZyUj9U?usp=sharing
Temporal
Gdrive : https://drive.google.com/drive/folders/1yciRZ-b8r27JcEsE_g0N99tPDMzSfTpe?usp=sharing
Codes
Enviroment
- Python 3.8
- scanpy 1.9.1
- seaborn 0.12.2
- pandas 1.5.3
- matplotlib 3.6.3
- numpy 1.23.5
- tensorflow 2.11.0
- keras-tuner 1.4.5
- keras 2.15.0
Installation
pip
For a traditional Python installation, use
$ pip install "$package name"
Usage
Data processing
- create_sce.R
- collect_data.R
simulating scRNA Dataset
- generate_sim_data.R
Data visualization
- figure5_A.py
- figure5_B.py
- sfigure1.R
Dynamic batching for DB-AAE (db_aae_tuning_example.py)
class AEHyperModel(kt.HyperModel):
def build(self, hp):
# Build the encoder / decoder
encoder = build_encoder(n1,n2,n3,n4,activation)
decoder = build_decoder(n1,n2,n3,n4,activation)
# Build and compile the discriminator
optimizer = RMSprop(learning_rate=0.00002)
discriminator = build_discriminator(n1,n2,n3,n4,activation)
discriminator.compile(loss='binary_crossentropy',optimizer=optimizer,metrics=['accuracy'])
autoencoder_input = Input(shape=(n1,))
reconstructed_input = Input(shape=(n1,))
encoded_repr = encoder(autoencoder_input)
reconstructed = decoder(encoded_repr)
# For the adversarial_autoencoder model we will only train the generator
discriminator.trainable = False
# The discriminator determines validity of the encoding
validity = discriminator(encoded_repr)
# The adversarial_autoencoder model (stacked generator and discriminator)
adversarial_autoencoder = Model(autoencoder_input, [reconstructed, validity])
adversarial_autoencoder.compile(optimizer=optimizer,loss=['mse', 'binary_crossentropy'],loss_weights=[0.999, 0.001])
return adversarial_autoencoder
def fit(self, hp, adversarial_autoencoder, *args, **kwargs):
batch_size=hp.Int('batch_size', min_value = 16, max_value = 128, step = 8)
encoder = build_encoder(n1,n2,n3,n4,activation)
decoder = build_decoder(n1,n2,n3,n4,activation)
discriminator = build_discriminator(n1,n2,n3,n4,activation)
optimizer = RMSprop(learning_rate=0.00002)
discriminator.compile(loss='binary_crossentropy',optimizer=optimizer,metrics=['accuracy'])
# For the adversarial_autoencoder model we will only train the generator
discriminator.trainable = False
valid = np.ones((batch_size, 1))
fake = np.zeros((batch_size, 1))
past = datetime.now()
for epoch in np.arange(1, n_epochs + 1):
for i in range(int(len(x) / batch_size)):
# ---------------------
# Train Discriminator
# ---------------------
# Select a random batch of input matrix
batch = x[i*batch_size:i*batch_size+batch_size]
latent_fake = encoder.predict(batch)
latent_real = np.random.normal(size=(batch_size, latent_dim))
# Train the discriminator
d_loss_real = discriminator.train_on_batch(latent_real, valid)
d_loss_fake = discriminator.train_on_batch(latent_fake, fake)
d_loss = 0.5 * np.add(d_loss_real, d_loss_fake)
# ---------------------
# Train Generator
# ---------------------
# Train the generator
g_loss = adversarial_autoencoder.train_on_batch(batch, [batch, valid])
return 100*d_loss[1]
