Scaden

Scaden trains an ensemble of three multilayer perceptrons of increasing width (m256, m512, m1024) on simulated pseudo-bulks and averages their predictions. It is the simplest of the four ported methods and a reasonable first thing to run on a new dataset.

Chunks below are not evaluated during build. Sample counts and epochs are reduced for demonstration; use the function defaults for real analyses.

Data

library(quasar)
library(QuasarDeconData)

sc_ref <- load_cov_imm("train")
pbulk  <- load_cov_pbulk(1)

bulk  <- pbulk$bulk_expression_profiles   # genes x samples
truth <- pbulk$ground_truth_proportions   # samples x cell types

Simulate

scaden_sim <- scaden_sim_pb(
  sc_data      = sc_ref,
  celltype_col = "cell_type",
  n            = 500,      # cells per pseudo-bulk
  samplenum    = 5000,     # number of pseudo-bulks
  sparse       = TRUE,
  seed         = 42
)

sparse = TRUE zeroes a subset of cell types in a subset of samples, producing training mixtures in which some populations are absent. This matters when the target samples may not contain every cell type in the reference.

scaden_sim_pb() can also merge several cell types into a single unknown class via unknown_celltypes. The merge happens after simulation — each original type gets its own Dirichlet component and is sampled independently, then the proportion columns are summed. This mirrors the original Python behaviour and differs from relabelling cells before simulation.

Process

scaden_proc <- scaden_process(
  sim_data   = scaden_sim,
  bulk_data  = bulk,
  mode       = "scaden",
  var_cutoff = 0.1
)

Two preprocessing modes are available:

"scaden" applies an absolute variance cutoff to the bulk matrix only, then intersects with the training genes. Uses log2(x + 1) and per-sample min-max scaling.
"tape" applies a top-quantile cutoff to both matrices independently before intersecting. Uses log(x + 1) and either min-max or standardisation via scaler.

The log base is immaterial in practice — both scalers are invariant to multiplication by a positive constant — but each mode stays literal to its source.

Train

scaden_fit <- scaden(
  train_x    = scaden_proc$train_x,
  train_y    = scaden_proc$train_y,
  lr         = 1e-4,
  batch_size = 128,
  epochs     = 20,
  seed       = 123
)

All three architectures train sequentially, each with its own progress bar. Per-batch and per-epoch losses are returned for every model.

Predict

scaden_pred <- scaden_predict(scaden_fit, scaden_proc$test_x)

head(scaden_pred$average_output)

The individual ensemble members are returned as well (prediction_model256, prediction_model512, prediction_model1024), which is useful for gauging how much the three architectures disagree — wide disagreement on a sample is a signal worth following up.

Evaluate

metrics <- quasar_prop_metrics(scaden_pred$average_output, truth)

metrics$cell_type_rmse
metrics$pearson_celltype_cor

Reference

Menden, K., Marouf, M., Oller, S., Dalmia, A., Magruder, D. S., Kloiber, K., … & Bonn, S. (2020). Deep learning-based cell composition analysis from tissue expression profiles. Science Advances, 6(30), eaba2619.

If you use this implementation, cite both the QUASAR paper and Menden et al.

Sergej Ruff