Activity inference accelerators and batch mode • SecAct

After installing the SecAct R package, you can complete all analyses in the tutorial sections. Note that the activity inference functions (SecAct.activity.inference and its *.ST and *.scRNAseq variants) rely on ridge regression and permutation in R language, and may run relatively slowly. The default ridge + permutation kernel is pure R with no compiled dependencies — SecAct installs and runs anywhere R runs. For large datasets, two optional accelerator packages provide drop-in speed-ups:

Installation

Once either package is installed, no additional configuration is required—the activity inference functions will automatically use the available accelerator.

RidgeFast — multi-threaded CPU (GSL + OpenMP)
RidgeCuda — NVIDIA GPU (CUDA)

Example

With either installed, SecAct.activity.inference(..., backend = "auto") (the default) picks GPU > CPU-fast > pure-R automatically. At rng_method = "mt19937" (default) and ncores = 1, output is bit-identical across all three backends.

# Memory-bounded inference on 100k samples
res <- SecAct.activity.inference(
  inputProfile = Y_large,
  is.differential = TRUE,
  batch_size = 5000
)

For memory-constrained workflows, set batch_size to process samples in column-batches:

# Stream results to HDF5 when even the result matrices don't fit in RAM
SecAct.activity.inference(
  inputProfile = Y_large,
  is.differential = TRUE,
  batch_size = 5000,
  output_h5 = "results.h5",
  is.group.sig = FALSE
)