DISSECT

DISSECT trains an ensemble with semi-supervised consistency regularization: at each step it forms mixtures of simulated and real samples, and penalises predictions on the mixture that disagree with the corresponding mixture of the individual predictions. Because that constraint uses the real samples without needing their labels, the model sees target-domain data during training.

It estimates cell-type fractions and, optionally, cell-type-specific expression.

Chunks below are not evaluated during build. Step counts 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

Simulate

dissect_sim <- dissect_simulate(
  sc_data      = sc_ref,
  celltype_col = "cell_type",
  batch_col    = "donor_id",
  type         = "bulk",
  n_samples    = 5000,
  save_expr    = TRUE,
  seed         = 42
)

save_expr = TRUE stores per-cell-type expression contributions for every simulated sample. It increases memory use considerably but is a hard requirement for the expression model — dissect_expr() errors without it. If you only want proportions, set it to FALSE.

batch_col performs simulation separately per batch and concatenates the results, so batch structure in the reference is preserved rather than mixed away.

Setting type = "st" produces spot-level profiles instead: only a handful of cells per spot, allocated by rounding rather than flooring. downsample then thins the resulting counts without replacement.

Process

dissect_proc <- dissect_process(
  bulk        = bulk,             # genes x samples
  sim_data    = dissect_sim,
  var_cutoff  = 0.1,
  test_in_mix = 1
)

test_in_mix controls how many real samples enter the online mixing step. For microarray data, pass test_dataset_type = "microarray" — the values are back-transformed out of log space before normalisation.

Estimate fractions

dissect_prop_res <- dissect_prop(
  processed = dissect_proc,
  n_steps   = 5000,
  models    = 1:5,                # ensemble size
  device    = "auto"
)

head(dissect_prop_res$fractions)
head(dissect_prop_res$scores)

The consistency term is introduced in stages: the first 2000 steps use the supervised loss alone, after which the consistency weight switches on. scores holds the pre-activation outputs, before the softmax.

Estimate expression

dissect_expr_res <- dissect_expr(
  bulk         = bulk,
  fractions    = dissect_prop_res$fractions,
  sim_data     = dissect_sim,
  n_steps_expr = 5000,
  device       = "auto"
)

names(dissect_expr_res$expression_layered)
dim(dissect_expr_res$expression_layered[[1]])

Expression estimation is a separate stage that takes fractions as input, via a conditional variational autoencoder. If you want to evaluate the expression model in isolation, supply ground-truth proportions here instead of estimated ones, so that expression differences are not confounded by errors in the fractions.

expression_layered is one samples × genes matrix per cell type; expression_combined is the same content as a long data frame with cell_type and sample columns.

Wrapper

dissect_res <- dissect(
  sc_data      = sc_ref,
  bulk         = bulk,
  celltype_col = "cell_type",
  batch_col    = "donor_id",
  save_expr    = TRUE,
  device       = "auto"
)

head(dissect_res$fractions)

The wrapper runs simulation, processing, fraction estimation, and — when save_expr = TRUE — expression estimation. If the expression stage fails it warns and returns the proportion results rather than aborting.

Evaluate

quasar_prop_metrics(dissect_prop_res$fractions, truth)$cell_type_rmse

Reference

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.

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

Sergej Ruff

2026-08-27