Downstream analysis

Once the sample embedding has been computed, every downstream module runs off the same object. The sample embedding is a single key, adata.uns['X_DR_sample'] (a samples × PCs DataFrame), written in place by compute_sample_embedding into the cell-level AnnData. The functions are identical across RNA, ATAC, and multi-omics — only the input AnnData changes.

Figures on these pages use the RNA COVID run

All example figures shown across the downstream tutorials come from the scRNA-seq COVID pipeline output. The calls are the same for ATAC and multi-omics; point the functions at the matching AnnData and the downstream behavior is identical.

Before you start

The cell-level AnnData written by preprocessing (adata_preprocessed.h5ad) carries the sample embedding in uns['X_DR_sample'] after compute_sample_embedding runs. Most downstream modules take a one-row-per-sample AnnData; build it on the fly from that key with build_sample_adata.

import anndata as ad
from sampledisco.sample_embedding import compute_sample_embedding
from sampledisco.sample_embedding.sample_embedding import build_sample_adata

adata_cell = ad.read_h5ad("sampledisco_demo_output/rna/preprocess/adata_preprocessed.h5ad")

# Compute the sample embedding in place (skip if uns['X_DR_sample'] already present)
adata_cell  = compute_sample_embedding(adata_cell, output_dir="sampledisco_demo_output/rna")

# Materialize the sample-level AnnData (samples × PCs) for downstream consumers
pseudo_adata = build_sample_adata(adata_cell, sample_col="sample")

The config-driven wrapper does all of this for you

Running the full pipeline with sampledisco -m complex --config <config.yaml> preprocesses, computes the sample embedding, and runs every downstream module below in one call — you rarely need to invoke these functions by hand.

Tutorials

Each function has its own page with the call, parameters, outputs, and representative figures.

Page	Function	What it does
Autotune	run_autotune	Selects the RMD-weight α (composition vs displacement block weighting) via the *_autotune_enable wrapper flags.
Sample distance	sample_distance	Pairwise distance matrices on the sample embedding with multiple metrics.
Dimension association	run_dimension_association_analysis	Per-PC variance-explained decomposition against every metadata variable — confounder / leading-covariate check.
Trajectory — CCA	CCA_Call, cca_pvalue_test	Supervised pseudotime and its permutation significance.
Trajectory — TSCAN	TSCAN	Unsupervised pseudotime via GMM + MST.
Trajectory DGE	run_trajectory_gam_differential_gene_analysis	GAM-based differential expression along pseudotime.
Sample clustering	cluster	K-means on the sample embedding.
Proportion test	proportion_test	Cell-type composition test across sample groups.
RAISIN cluster DGE	raisinfit, run_pairwise_tests	Hierarchical GLM for cluster-level differential expression.
General visualization	visualization	Cell-type dendrogram, proportion PCA, expression UMAP.

Recommended reading order: sample distance → dimension association → trajectory (CCA or TSCAN) → trajectory DGE → sample clustering → proportion test / RAISIN → general visualization.