Trains the DISSECT expression model and returns estimated expression profiles for each sample-cell-type combination.
Usage
dissect_expr(
bulk = NULL,
fractions,
sim_data,
normalize_simulated = "cpm",
normalize_test = "cpm",
n_steps_expr = 5000,
expr_scaling = "p99",
latent_dim = 128,
batch_size = 128,
lr = 0.001,
beta_vae = 0.01,
lambda_cons = 0.1,
seed = 42,
device = c("auto", "cpu", "cuda"),
cuda_index = NULL
)Arguments
- bulk
Numeric matrix with genes in rows and samples in columns.
- fractions
A data frame or matrix of estimated cell-type fractions with samples in rows and cell types in columns, typically obtained from [dissect_prop()].
- sim_data
A list returned by [dissect_simulate()] with `save_expr = TRUE`.
- normalize_simulated
Character scalar or `NULL`. Currently `"cpm"` or `NULL`.
- normalize_test
Character scalar or `NULL`. Currently `"cpm"` or `NULL`.
- n_steps_expr
Integer scalar or `NULL`. Number of training steps for the expression model.
- expr_scaling
Character scalar. Scaling method used before model fitting, typically `"p99"` or `"max"`.
- latent_dim
Integer scalar. Latent dimension for the conditional VAE.
- batch_size
Integer scalar. Batch size.
- lr
Numeric scalar. Learning rate.
- beta_vae
Numeric scalar. Weight of the KL divergence term.
- lambda_cons
Numeric scalar. Weight of the consistency loss term.
- seed
Integer scalar or `NULL`. Optional random seed.
- device
Character scalar. One of `"auto"`, `"cpu"`, or `"cuda"`.
- cuda_index
Integer scalar or `NULL`. Optional CUDA device index.
Value
A named list with components:
- expression_layered
Named list of matrices, one per cell type, with samples in rows and genes in columns.
- expression_combined
Data frame containing estimated expression for all sample-cell-type combinations, along with `cell_type` and `sample` columns.
- scaled_counts
Scaled expression matrix corresponding to the final combined estimates.
- celltypes
Character vector of cell-type names.
- genes
Character vector of genes used in the model.
Details
This implementation keeps the DISSECT expression methodology the same, but avoids materializing the full expanded `(sample x celltype)` training design matrices in memory. Instead, minibatches are assembled on the fly.
dissect_expr implements the DISSECT cell-type-specific expression estimation workflow described by Khatri, Machart, and Bonn (2024) in torch in R. It follows the DISSECT expression-estimation methodology based on simulated per-cell-type layers, estimated sample fractions, and a conditional variational autoencoder. The training logic is unchanged, but expanded input matrices are built batch-wise rather than all at once.
Progress bars and elapsed time are reported during training.
References
Khatri, R., Machart, P., & Bonn, S. (2024). DISSECT: deep semi-supervised consistency regularization for accurate cell type fraction and gene expression estimation. Genome Biology, 25(1), 112.
Original DISSECT software repository: https://github.com/imsb-uke/DISSECT