PRISM

Phenotype-Resolved Inference in Single-cell Mixed models via latent disease states and contextualized differential expression

Not every cell from a disease donor is in a disease state. PRISM is a negative-binomial mixed model for multi-donor single-cell RNA-seq. For each cell of a disease donor it infers the posterior probability q that the cell is in the disease state (cells from control donors are fixed to the healthy state). It then estimates, per gene, a constant disease effect alpha and its modulation by cell-level context z, theta. Parameters are fit by EM.

Install

pip install prism-de

The import name is prism. PRISM needs PyTorch >= 2.2. For a GPU, first install the torch build that matches your CUDA driver (pytorch.org). Fits default to device="cuda", so pass device="cpu" if you have no GPU.

Quick start

from prism import PRISMConfig, PRISMTrainer, generate_prism_data

data, truth = generate_prism_data(n_subjects=50, n_genes=60, n_cells_per_subject=200,
                                  rho=0.6, seed=42, device="cpu")

cfg = PRISMConfig(n_genes=data.n_genes, n_covars=data.n_covars, n_context=data.n_context,
                  ot_project_q=True, auto_q_prior=False,  # PRISM-OT, the configuration used in the paper
                  max_em_iter=30, wandb_enabled=False, device="cpu")
res = PRISMTrainer(cfg).fit(data)

print("general DE (FDR < 0.05):", int((res.q_values_de < 0.05).sum()))
print("context DE (FDR < 0.05):", int((res.q_values_context.min(dim=1).values < 0.05).sum()))

The fit takes about 4 minutes on a 4-core CPU. wandb_enabled=False turns off Weights & Biases logging, which is on by default whenever wandb is installed.

Your data

PrismData.from_anndata reads raw counts from adata.X and per-cell columns from adata.obs. Include an intercept column among the covariates, and standardise continuous covariates and context variables.

from prism import PrismData

adata.obs["intercept"] = 1.0
data = PrismData.from_anndata(adata, covar_cols=["intercept", "age", "sex"], context_cols=["ctx"],
                              condition_col="disease", subject_col="donor", device="cuda")

Then fit it with the configuration above, using device="cuda".

Outputs

res.de_summary(gene_names=...) returns the per-gene estimates and tests as a pandas DataFrame.

Citation

PRISM: Phenotype-Resolved Inference in Single-Cell Mixed Models via Latent Disease States and Contextualized Differential Expression. NeurIPS 2026. See the documentation for the BibTeX entry.

License

MIT