Trains one OmicsTweezer neural network with supervised cell-type proportion prediction on simulated pseudobulks and a Wasserstein-style domain adaptation term using real bulk samples.
Usage
omics_train(
train_x,
train_y,
test_x,
dims,
drops,
epochs = 128L,
batch_size = 128L,
device = c("auto", "cpu", "cuda"),
cuda_index = NULL,
learning_rate = 1e-04,
loss_weight = 1,
seed = 2021L,
verbose = TRUE
)Arguments
- train_x
Numeric matrix with samples in rows and genes in columns.
- train_y
Numeric matrix with samples in rows and cell types in columns.
- test_x
Numeric matrix with target-domain bulk samples in rows and genes in columns.
- dims
Integer vector of length 4 specifying hidden-layer dimensions.
- drops
Numeric vector of length 4 specifying dropout probabilities.
- epochs
Integer scalar. Number of training epochs.
- batch_size
Integer scalar. Minibatch size.
- device
Character scalar. One of `"auto"`, `"cpu"`, or `"cuda"`. If `"auto"`, CUDA is used when available, otherwise CPU is used.
- cuda_index
Optional integer scalar giving the CUDA device index to use when `device = "cuda"`. CUDA indices are zero-based.
- learning_rate
Numeric scalar. Learning rate passed to Adam.
- loss_weight
Numeric scalar. Weight applied to the Wasserstein-style domain adaptation loss.
- seed
Integer scalar random seed.
- verbose
Logical scalar. If `TRUE`, prints an epoch-level progress bar.
Value
A named list with components:
- encoder
Trained torch encoder module.
- predictor
Trained torch predictor module.
- history
Data frame with epoch-level total, MSE, OT, and weighted OT losses.
Details
Each batch contains source-domain simulated pseudobulks with known proportions and target-domain real bulk profiles without labels. The training objective combines supervised MSE loss with a domain adaptation term computed from the encoded source and target embeddings.
References
TAPE project PyTorch implementation: https://github.com/poseidonchan/TAPE/blob/main/TAPE/model.py