Run DEGAS Analysis for Single-Cell and Bulk RNA-seq Data Integration

This function performs DEGAS to integrate single-cell and bulk RNA-seq data, identifying phenotype-associated cells using a bootstrap aggregated multi-task learning approach.

Usage

DoDEGAS(
  select_fraction = 0.05,
  min_thresh = 0.4,
  matched_bulk,
  sc_data,
  phenotype = NULL,
  sc_data.pheno_colname = NULL,
  label_type = "DEGAS",
  phenotype_class = c("binary", "continuous", "survival"),
  tmp_dir = "tmp",
  env_params = list(),
  degas_params = list(),
  normality_test_method = c("jarque-bera", "d'agostino", "kolmogorov-smirnov"),
  verbose = TRUE,
  ...
)

Arguments

select_fraction

The top percentage of selected cells will be considered as Positive cells, without considering how much larger the possible correlation coefficient of the observation group is compared to that of the control group. Only usedl when phenotype_class is "binary" or "survival". (default: 0.05)

min_thresh

DEGAS will calculate the possible correlation coefficients for each cell related to the phenotype. When the coefficient of the observation group is at least min_thresh larger than that of the control group, it can be considered related to the phenotype and will be marked as Positive. The priority of min_thresh is higher than that of select_fraction. (default: 0.4)

matched_bulk

Bulk RNA-seq data as matrix or data.frame (rows=genes, columns=samples)

sc_data

Single-cell data as Seurat object containing RNA assay

phenotype

Bulk-level phenotype data. For classification: binary matrix with one-hot encoding. For survival: matrix with two columns (time and event status). Can be NULL, matrix, data.frame, or vector.

sc_data.pheno_colname

Column name for single-cell phenotype in metadata (if available)

label_type

Label type for DEGAS results (default: "DEGAS")

phenotype_class

Type of phenotype: "binary" (classification), "continuous", or "survival"

tmp_dir

Temporary directory for intermediate files (default: "tmp")

env_params

List of environment parameters for Python setup including:

env.name: environment name (default: "r-reticulate-degas")
env.type: environment type "conda", "environment", or "venv" (default: "environment")
env.method: environment setup method "system", "conda" (default: "system")
env.file: path to environment file (default: system.file("conda/DEGAS_environment.yml", package = "SigBridgeR"))
env.python_version: Python version (default: "3.9.15")
env.packages: named vector of Python packages and versions (default: c("tensorflow" = "2.4.1", "protobuf" = "3.20" ,"numpy" = "any"))
env.recreate: whether to recreate environment (default: FALSE)
env.use_conda_forge: whether to use conda-forge channel (conda only, default: TRUE)
env.verbose: verbose output (default: FALSE)

degas_params

List of DEGAS algorithm parameters including:

DEGAS.model_type: model type ("BlankClass", "ClassBlank", "ClassClass", "ClassCox", "BlankCox")
DEGAS.architecture: "Standard" (feed forward) or "DenseNet" (dense net)
DEGAS.ff_depth: number of layers in model (>=1, default: 3)
DEGAS.pyloc: path to Python executable (default: NULL, automatic detection)
DEGAS.bag_depth: bootstrap aggregation depth (>=1, default: 5)
DEGAS.train_steps: training steps (default: 2000)
DEGAS.scbatch_sz: single-cell batch size (default: 200)
DEGAS.patbatch_sz: patient batch size (default: 50)
DEGAS.hidden_feats: hidden features (default: 50)
DEGAS.do_prc: dropout percentage (default: 0.5)
DEGAS.lambda1: regularization parameter 1 (default: 3.0)
DEGAS.lambda2: regularization parameter 2 (default: 3.0)
DEGAS.lambda3: regularization parameter 3 (default: 3.0)
DEGAS.seed: random seed (default: 2)

normality_test_method

Method for normality testing: "jarque-bera", "d'agostino", or "kolmogorov-smirnov"

verbose

Logical, whether to print messages.

...

for future compatibility

Value

A list containing:

scRNA_data: Seurat object with DEGAS labels added to metadata
model: The model trained using the input data, andit can be used for cell classification prediction.
DEGAS_prediction: Data table with DEGAS predictions containing:
- Predicted label probabilities for each cell
- Cell labels ("Positive"/"Other") based on selection criteria
- Difference scores for binary phenotypes
- Cell identifiers

Details

The function performs the following steps:

Validates input data and parameters
Sets up Python environment with required dependencies
Trains bootstrap aggregated DEGAS model using runCCMTLBag
Generates cell-level predictions using predClassBag
Applies statistical testing to identify phenotype-associated cells
Labels cells as "Positive" or "Other" based on selection criteria

Model type is automatically determined:

BlankClass: only bulk phenotype specified (scLab = NULL)
ClassBlank: only single-cell phenotype specified (patLab = NULL)
ClassClass: both single-cell and bulk phenotypes specified
ClassCox: single-cell phenotype + bulk survival data
BlankCox: only bulk survival data specified

References

Johnson TS, Yu CY, Huang Z, Xu S, Wang T, Dong C, et al. Diagnostic Evidence GAuge of Single cells (DEGAS): a flexible deep transfer learning framework for prioritizing cells in relation to disease. Genome Med. 2022 Feb 1;14(1):11.

Examples

if (FALSE) { # \dontrun{
# Binary classification example
result <- DoDEGAS(
  select_fraction = 0.05, # `select_fraction` only used in binary and survival phenotyping
  matched_bulk = bulk_matrix,
  sc_data = seurat_obj,
  phenotype = bulk_phenotype,
  phenotype_class = "binary"
)

# Survival analysis example
result <- DoDEGAS(
  select_fraction = 0.05, # `select_fraction` only used in binary and survival phenotyping
  matched_bulk = bulk_matrix,
  sc_data = seurat_obj,
  phenotype = survival_data,
  phenotype_class = "survival"
)
} # }