Skip to contents

Get started

Sergej Ruff

2026-08-27

Source: vignettes/quasar-deconvolution.Rmd
quasar-deconvolution.Rmd

Installation 

# install.packages("pak")
pak::pak("SergejRuff/QUASAR")

The example datasets used throughout the documentation live in a companion data package:

pak::pak("SergejRuff/QuasarDeconData")

All deep-learning methods in this package are built on torch. If this is a fresh installation, run the following once:

torch::install_torch()

Every training and prediction function accepts device = "auto", "cpu", or "cuda". "auto" selects CUDA when a compatible torch build is available and falls back to CPU otherwise, so the same code runs on a laptop and on a GPU node without modification.

library(quasar)
library(QuasarDeconData)

What this package does

Bulk transcriptomics measures gene expression from mixtures of cell types, so a difference between two samples can reflect a change in cellular composition, a change within individual cell types, or both. Cell-type deconvolution estimates those proportions from bulk data and makes the two explanations separable.

The package provides three things:

  • QUASAR, an uncertainty-aware deconvolution method that returns confidence and prediction intervals alongside point estimates.
  • Native R implementations of Scaden, TAPE, DISSECT, and OmicsTweezer, written in torch for R with GPU support and direct Seurat / SingleCellExperiment compatibility.
  • Supporting utilities for pseudo-bulk simulation, normalisation, evaluation metrics, and .h5ad interoperability.

Where to go next

If you want to Go to
Estimate proportions with uncertainty intervals QUASAR
Compare deconvolution methods, or use one in R instead of Python Deep learning methods
Load the COVID-19, PBMC, or benchmark datasets Data
Simulate pseudo-bulks, score predictions, read or write .h5ad Utilities
Look up a specific function Reference

If you are new to the package, the Data page is the shortest route to a working example: it shows how to load a single-cell reference and a matching bulk matrix, which is the input every method expects.

The shape of a deconvolution workflow

All four ported methods follow the same four stages, which is worth internalising before reading any individual method page:

  1. Simulate pseudo-bulk training data from an annotated single-cell reference, with known cell-type proportions.
  2. Process the simulated and real bulk matrices onto a shared gene set, with variance filtering, log transformation, and per-sample scaling.
  3. Train a model to predict proportions from expression.
  4. Predict proportions for the real bulk samples.

Each method exposes those four stages as separate functions, plus a wrapper that runs all of them in one call. Once you have seen one method, the others read the same way — the differences are in the model and in what happens during training, not in the surrounding workflow.

Citation

If you use QUASAR, cite the QUASAR paper. If you use one of the R implementations of Scaden, TAPE, DISSECT, or OmicsTweezer, please cite both the QUASAR paper — for the R implementation — and the original publication of the method. Full references are given on each method’s page.

citation("quasar")