Supported methods

The canonical list of what --method accepts, what each value resolves to, and which tasks support it. scfoundry list methods prints the same information from the registry that ships with the package.

Task support matrix

Method	Category	GPU	download	embed	transfer	finetune
c2s	zero-shot	yes	✓	✓	✓	—
cellama	zero-shot	yes	✓	✓	✓	✓
cellfm	zero-shot	yes	✓	✓	✓	✓
cellplm	zero-shot	yes	✓	✓	✓	✓
geneformer	zero-shot	yes	✓	✓	✓	✓
genept	zero-shot	no	✓	✓	✓	—
langcell	zero-shot	yes	✓	✓	✓	✓
scbert	zero-shot	yes	✓	✓	✓	✓
sccello	zero-shot	yes	✓	✓	✓	✓ (linear probing)
scfoundation	zero-shot	yes	✓	✓	✓	✓
scgpt	zero-shot	yes	✓	✓	✓	✓
scimilarity	zero-shot	yes	✓	✓	✓	—
scprint	zero-shot	yes	✓	✓	✓	—
uce	zero-shot	yes	✓	✓	✓	—
novae	zero-shot	yes	✓	✓	—	—
pca	reference	no	—	✓	—	—
scvi	reference	yes	✓	✓	✓	—
scgpt_integrated	integration	yes	—	✓	—	—
scvi_denovo	integration	yes	—	✓	—	—
harmony	integration	no	—	✓	—	—
seurat_cca	integration	no	—	✓	—	—
seurat_rpca	integration	no	—	✓	—	—

Three categories:

• zero-shot — pretrained foundation models with frozen weights. What the benchmark ranks.

• reference — pca, the classical yardstick, and scvi, zero-shot projection onto the pretrained CELLxGENE Census scVI: a pretrained model rather than a baseline, kept apart from the transformer scFMs the paper counts.

• integration — methods trained on the input using its batch labels. Accepted by embed only, and ranked separately in every result table.

finetune lists only methods whose official adaptation updates weights; the models whose authors adapt them with a classifier on frozen embeddings (genept, scimilarity, scprint, uce) are covered by transfer --classifier mlp. scCello’s official recipe is linear probing on the frozen encoder and is kept under finetune for that reason. Cell2Sentence fine-tuning is not yet supported, and novae — a spatial model — only embeds.

Models and containers

Method	Name	Category	Container image	Default checkpoint (--model)	Upstream version
c2s	Cell2Sentence	zero-shot	housy17/c2s:latest	C2S/C2S-Pythia-410m-cell-type-prediction	v1.2.0
cellama	CELLama	zero-shot	housy17/cellama:latest	CELLama/all-MiniLM-L6-v2	v0.1.0
cellfm	CellFM	zero-shot	housy17/cellfm:latest	CellFM/CellFM_80M_weight.ckpt	commit 5054a2a (MindSpore-parity fixes)
cellplm	CellPLM	zero-shot	housy17/cellplm:latest	CellPLM/20231027_85M.best.ckpt	v0.1.0
geneformer	Geneformer	zero-shot	housy17/geneformer:latest	Geneformer/Geneformer-V2-316M	V2-316M
genept	GenePT-w	zero-shot	housy17/genept:latest	GenePT/ada-002 (selector)	commit 3602699
langcell	LangCell	zero-shot	housy17/langcell:latest	LangCell/ckpt (fixed)	commit 69e41ef
scbert	scBERT	zero-shot	housy17/scbert:latest	scBERT/panglao_pretrain.pth	v1.0.0
sccello	scCello	zero-shot	housy17/sccello:latest	scCello/scCello-zeroshot	commit 767585b
scfoundation	scFoundation	zero-shot	housy17/scfoundation:latest	scFoundation/models.ckpt	commit 397631c
scgpt	scGPT	zero-shot	housy17/scgpt:0.2.4	scGPT/scGPT_human	v0.2.4
scimilarity	SCimilarity	zero-shot	housy17/scimilarity:latest	SCimilarity/model_v1.1	v0.4.1 (model v1.1)
scprint	scPRINT	zero-shot	housy17/scprint:latest	scPRINT/medium-v1.5.ckpt	2.3.8 (checkpoint medium-v1.5)
uce	UCE	zero-shot	housy17/uce:latest	UCE/33l_8ep_1024t_1280.torch	commit 8227a65 (33 layers)
novae	Novae (spatial)	zero-shot	housy17/novae:1.0.0	Novae/novae-human-0	novae-human-0
pca	PCA	reference	housy17/scllms:latest	none	scanpy HVG(2000) → scale → PCA(50)
scvi	scVI (Census)	reference	scverse/scvi-tools:py3.11-cu12-runtime-stable	scVI/Census2024-07-01-HomoSapiens (--scvi_model)	Census 2024-07-01 Homo sapiens
scgpt_integrated	scGPT (integrated)	integration	housy17/scgpt:0.2.4	scGPT/scGPT_human	scGPT_human fine-tuned with DSBN + DAB + GEPC + ECS
scvi_denovo	scVI (de novo)	integration	scverse/scvi-tools:py3.11-cu12-runtime-stable	none (trained on the input)	scvi-tools
harmony	Harmony	integration	satijalab/seurat:5.5.0	none	Seurat v5 IntegrateLayers(HarmonyIntegration)
seurat_cca	Seurat CCA	integration	satijalab/seurat:5.5.0	none	Seurat v5 IntegrateLayers(CCAIntegration)
seurat_rpca	Seurat RPCA	integration	satijalab/seurat:5.5.0	none	Seurat v5 IntegrateLayers(RPCAIntegration)

Every image is pulled automatically on first use and cached under cache/.shared/nxf_singularity/ in the workspace. Upstream versions are the revision each container was built from — cite the upstream project, not this table, when you report results. Links are on the citation page.

Three entries do not follow the --model pattern:

• genept — --model is a selector, not a path. Only GenePT/ada-002 (1536-d) and GenePT/model-3-large (3072-d) are recognised.

• langcell — the checkpoint path is fixed to LangCell/ckpt; there is no --model.

• scvi — the pretrained Census model is selected with --scvi_model.

Method-specific parameters

Parameters that exist for one method only, forwarded as --name value. Shared parameters are in the parameter reference.

Method	Parameter	Default	Effect
scbert	--bin_num	5	Number of expression bins used to discretise counts.
uce	--nlayers	33	Transformer layers to use from the checkpoint.
scprint	--how	"random expr"	Gene-selection strategy for building the input sequence.
scprint	--max_len	4000	Maximum genes per cell.
scprint	--scprint_save_every	10000	Cells per intermediate checkpoint during inference.
scfoundation	--scfoundation_pool_type	"all"	Pooling over token embeddings. "all" or "max".
scfoundation	--scfoundation_lr, --scfoundation_scheduler, --scfoundation_grad_clip	1e-4, "warmup_cosine", 0.0	Fine-tuning learning rate, schedule and gradient clipping (0.0 disables).
cellama	--top_k	30	Highest-expressed genes used to build each cell’s sentence.
cellama	--cellama_finetuned_top_k	30	Same, when re-embedding with the fine-tuned backbone.
scimilarity	--scimilarity_use_gpu	true	Whether to run inference on the GPU.
scvi	--scvi_model	scVI/Census2024-07-01-HomoSapiens	Pretrained Census model.
scvi	--scvi_use_gpu	true	Whether to run inference on the GPU.
scvi_denovo	--scvi_n_latent, --scvi_max_epochs, --integration_n_hvg	30, null, 2000	Latent dimensions; training epochs (null lets scvi-tools choose); highly variable genes (0 = all genes).
scgpt_integrated	--integration_epoch, --integration_batch_size, --integration_n_hvg	15, 64, 1200	Fine-tuning epochs, batch size and highly variable genes.
harmony, seurat_cca, seurat_rpca	--seurat_n_hvg, --seurat_n_pcs	2000, 30	Highly variable genes and principal components computed before integration.
pca	--pca_n_top_genes, --pca_n_comps	2000, 50	Highly variable genes retained; principal components, and therefore output dimensions.

Default batch sizes

Inference batch size for embed (and the embedding step of transfer). The variation is real — models differ by orders of magnitude in memory per cell.

--batch-size	Methods	Note
2048	scimilarity	Lightweight encoder over a fixed gene panel.
1024	scvi, cellplm	Small models.
64	scgpt, cellfm, geneformer, scbert, sccello, scprint, langcell	The transformer default.
16	uce	Long sequences and a large model.
8	c2s	A 410M-parameter autoregressive language model.
not used	cellama, scfoundation, genept, novae, pca, integration methods	These declare no batch_size; passing --batch-size has no effect.

Fine-tuning uses separate parameters with their own per-method defaults — see Fine-tune.

The PCA reference

Worth documenting precisely, because it is the yardstick every foundation model is measured against:

normalize_total(target_sum=1e4)
log1p
highly_variable_genes(n_top_genes=2000, flavor="seurat")
subset to HVGs
scale(max_value=10)
pca(n_comps=50, svd_solver="arpack", random_state=0)

This is the standard Scanpy workflow with default settings, deliberately unoptimised. It needs no checkpoint and runs on the CPU in seconds.

See also

• Parameter reference — every parameter, grouped by task.

• Output reference — where each task writes.

• Adding a method — wrapping a model that is not listed here.