Trains the DISSECT fraction model and returns estimated cell-type proportions and pre-activation scores for each sample.
Usage
dissect_prop(
bulk = NULL,
reference = NULL,
sim_data = NULL,
processed = NULL,
test_dataset_type = "bulk",
duplicated = "first",
normalize_simulated = "cpm",
normalize_test = "cpm",
var_cutoff = 0.1,
test_in_mix = 1,
n_hidden_layers = 4,
hidden_units = c(512, 256, 128, 64),
hidden_activation = "relu6",
output_activation = "softmax",
loss = "kldivergence",
n_steps = 5000,
lr = 1e-05,
batch_size = 64,
dropout = NULL,
alpha_range = c(0.1, 0.9),
normalization_per_batch = "log1p-MinMax",
models = c(1, 2, 3, 4, 5),
mix = "srm",
device = c("auto", "cpu", "cuda"),
cuda_index = NULL
)Arguments
- bulk
Numeric matrix with genes in rows and samples in columns, or `NULL` if `processed` is supplied.
- reference
Numeric matrix with genes in rows and cell types in columns, or `NULL`.
- sim_data
A list returned by [dissect_simulate()], or `NULL`.
- processed
A list returned by [dissect_process()], or `NULL`.
- test_dataset_type
Character scalar. Either `"bulk"` or a supported alternative matching the original workflow.
- duplicated
Character scalar controlling duplicated gene handling.
- normalize_simulated
Character scalar or `NULL`. Currently `"cpm"` or `NULL`.
- normalize_test
Character scalar or `NULL`. Currently `"cpm"` or `NULL`.
- var_cutoff
Numeric scalar or `NULL`. Variance threshold for bulk preprocessing.
- test_in_mix
Integer scalar. Number of real samples to use in online mixing.
Integer scalar. Number of hidden layers in the fraction model.
Integer vector. Number of units in each hidden layer.
Character scalar. Hidden-layer activation, applied to every ensemble model. Note that the reference Python mutates its config in place inside the model loop, so only the first model receives `relu6` and the remaining four fall back to `relu`; this implementation follows the documented behaviour instead.
- output_activation
Character scalar. Output activation function.
- loss
Character scalar. One of `"kldivergence"`, `"l2"`, or `"l1"`.
- n_steps
Integer scalar. Number of training steps.
- lr
Numeric scalar. Learning rate.
- batch_size
Integer scalar. Batch size.
- dropout
Numeric vector or `NULL`. Dropout rates for hidden layers.
- alpha_range
Numeric vector of length two. Range for the mixing coefficient used in online mixtures.
- normalization_per_batch
Character scalar or `NULL`. Currently `"log1p-MinMax"` or `NULL`.
- models
Integer vector. Ensemble model identifiers. As in the original Python code, only the number of models is used.
- mix
Character scalar. Mixing strategy, either `"srm"` or `"rrm"`.
- 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:
- fractions
A data frame of estimated cell-type proportions with samples in rows and cell types in columns.
- scores
A data frame of pre-activation output scores with samples in rows and cell types in columns.
- processed
The processed input object used for training.
Details
dissect_prop implements the DISSECT cell-type fraction estimation strategy described by Khatri, Machart, and Bonn (2024) in torch in R. It reproduces the semi-supervised consistency-regularized training workflow for fraction estimation within an R interface.
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