Parameter reference

The pipeline parameters behind the scfoundry commands. The common ones have a command-line option; every other one is forwarded as --name value (see forwarding).

Note

Method modules declare their own defaults, and Nextflow scopes them per module — so --model genuinely resolves to scGPT/scGPT_human under --method scgpt and to UCE/33l_8ep_1024t_1280.torch under --method uce, even though both modules are loaded. Passing a value on the command line overrides every module at once, which is harmless because only one method runs per invocation.

Global

Defined in the workspace nextflow.config and available to every task.

Parameter	Option	Default	Description
gpu_id	--gpu	null	GPU index (or list) exposed to the container. null exposes all devices.
model_weights_dir	--weights-dir	<workspace>/data/model_weights	Host directory bind-mounted at /data/model_weights.
cache_dir	--cache-dir	<workspace>/cache	Container image cache and Nextflow caches.
host_side_home_dir	—	<workspace>/cache/.home	Stand-in $HOME for containers.
bin_dir	—	<pipeline>/bin	Pipeline code, bind-mounted at /code. Leave it alone.
workspace_dir	--workspace	$SCFOUNDRY_WORKSPACE	The workspace; the three directories above derive from it.

download

Parameter	Option	Default	Description
method	--method	required	One of the 16 methods with a download task. pca and the integration baselines are not accepted.
model	--model	per method	Checkpoint variant, where a method offers more than one (Novae).

check

Parameter	Option	Default	Description
data	--data	required	Input .h5ad, directory, or quoted glob.
method	--method	none	Method the file is meant for; adds its requirements (ensembl_id, batch labels, spatial coordinates).
label_key	--label-key	cell_type	Label column to look for.
batch_key	--batch-key	batch_id	Batch column to look for.
role	--role	none	reference makes missing labels a problem rather than a warning.
results_dir	--outdir	<workspace>/results	Reports land in <root>/check/.

embed

These also apply to the embedding step of transfer.

Parameter	Option	Default	Description
method	--method	required	Method to run. Also selects the container image.
data	--data	required	Input .h5ad, one file per run.
model	--model	per method	Checkpoint path relative to /data/model_weights.
batch_size	--batch-size	per method	Inference batch size. Not declared by every method.
batch_key	--batch-key	batch_id	Batch column. Integration methods only.
emb_results_dir	--outdir	<workspace>/results	Output root; embeddings land in <root>/embeddings/<method>/.

Method-specific parameters — --bin_num, --nlayers, --how, --max_len, --scfoundation_pool_type, --top_k, --scimilarity_use_gpu, --scvi_*, --scprint_*, --pca_*, --integration_*, --seurat_* — are listed with their defaults in method-specific parameters.

transfer

Parameter	Option	Default	Description
method	--method	required	Model providing the embeddings.
reference	--reference	none	Labelled reference set. Required to fit.
query	--query	none	Query set to label. Required to predict.
fitted	--fitted	none	Fitted model directory. Required when reference is absent.
classifier	--classifier	logreg	logreg, prototype, knn or mlp.
label_key	--label-key	cell_type	obs column on the reference holding labels.
knn_k	--knn-k	15	Neighbours for the knn classifier.
transfer_results_dir	--outdir	<workspace>/results	Output root; models and predictions land in <root>/transfer/.

All embed parameters apply as well.

finetune

Parameter	Option	Default	Description
method	--method	required	Model to fine-tune.
reference	--reference	none	Labelled training set. Required to fine-tune.
query	--query	none	Query set to predict. Required to predict.
fitted	--fitted	none	Fine-tuned model directory. Required when reference is absent.
finetune_label_key	--label-key	cell_type	obs column holding training labels.
finetune_epoch	--epochs	per method	Training epochs. Defaults follow each model’s published recipe.
finetune_batch_size	--batch-size	per method	Training batch size. Not declared by CellPLM.
finetune_eval_size	—	per method	Validation fraction, 0.2 for every method. Ignored by CELLama, whose backbone training holds out a fixed 1,000 generated sentences and whose post-hoc classifier always uses 0.2.
predict_batch_size	—	per method	Inference batch size. Not declared by CellPLM or CELLama.
model	--model	per method	Pretrained checkpoint to start from.
finetune_results_dir	--outdir	<workspace>/results	Output root; models and predictions land in <root>/finetune/.
cellama_finetuned_top_k	—	30	CELLama only.
scfoundation_lr, scfoundation_scheduler, scfoundation_grad_clip	—	1e-4, "warmup_cosine", 0.0	scFoundation only.

Per-method values for the four training defaults are tabulated on Fine-tune.

benchmark

Parameter	Option	Default	Description
embedding	--embedding	required	Embedding .h5ad, directory, or quoted glob.
method	--method	directory name	Method label written into the tables.
label_key	--label-key	cell_type	Cell-type column. The Leiden target cluster count follows its cardinality.
batch_key	--batch-key	batch_id	Batch column. Only read when metrics includes batch.
metrics	--metrics	bio	bio, batch or all. --batch-key on the command line implies all.
clustering	--clustering	leiden	leiden or kmeans.
batch_max_cells	--batch-max-cells	0	Stratified subsample for the batch metrics; 0 = all cells.
results_dir	--outdir	<workspace>/results	Output root; tables land in <root>/benchmark/.

geometry

Parameter	Option	Default	Description
embedding	--embedding	required	Embedding .h5ad, directory, or quoted glob.
data	--data	required	The raw-count input, or a directory matched by file name.
method	--method	directory name	Method label written into the tables.
label_key	--label-key	cell_type	Cell-type column.
batch_key	--batch-key	batch_id	Batch column; a missing column means one batch.
max_cells	--max-cells	20000	Subsample larger datasets to this many cells.
seed	--seed	0	Seed for subsampling and pair sampling.
results_dir	--outdir	<workspace>/results	Output root; tables land in <root>/geometry/ and <root>/<method>_geometry.csv.

Nextflow options worth knowing

Passed after --, verbatim to nextflow run:

Option	Why it matters here
-process.maxForks 1	Cap concurrent tasks within a run. transfer embeds reference and query as two parallel tasks; this serialises them.
-with-report, -with-trace	Execution reports, useful for sizing memory and runtime before a large run.
-work-dir <dir>	Move scratch off the workspace filesystem. Useful on clusters with fast node-local disk.

-c, -profile and -resume have their own options (--config, --profile, --resume).

See also

• Command reference — the commands and their run options.

• Method reference — per-method defaults and method-specific flags.

• nextflow.config reference — the configuration file itself.