SampleDisco

A cross-omics, cross-condition sample embedding tool for single-cell data.

SampleDisco takes a cell-level embedding (from any standard scRNA / scATAC / multi-omics integration method) and lifts it to a sample-level embedding that captures both cell-type composition and the per-cell-type state of each sample. Every downstream analysis — sample-to-sample distance, clustering, trajectory inference, phenotype association — then runs on that single shared sample embedding, regardless of modality.

Paper draft: /users/hjiang/GenoDistance/SampleDisco_Draft-11.pdf

What the method does

For each modality (RNA, ATAC, or integrated multi-omics) the pipeline produces two cell-level views:

It then assembles four blocks per sample (or per sample × modality for multi-omics):

1. A1 — one-hot cell-type composition
2. A2 — soft k-means composition at K_med (≈120)
3. A3 — soft k-means composition at K_fine (≈300)
4. RMD — leave-one-out cell-type-resolved displacement on Z_rmd

The four blocks are inverse-variance weighted, Frobenius-stacked, PCA-reduced to 10 dimensions, and Harmony-corrected at sample level. The result is stored as adata.uns['X_DR_sample'] and feeds every downstream module.

Repository layout

code/
├── SampleDisc.py              # CLI entry point (simple or complex mode)
├── config/                    # 9 YAML configs covering covid / blood / eye / heart / ENCODE / tabula / long_covid / unpaired / default
├── wrapper/                   # Orchestration
│   ├── wrapper.py             # Master wrapper; gates RNA + ATAC + multiomics + shared downstream
│   ├── rna_wrapper.py
│   ├── atac_wrapper.py
│   └── multiomics_wrapper.py
├── preparation/               # Preprocessing
│   ├── rna_preprocess_{cpu,gpu}.py   # QC → HVG → PCA → dual Harmony → Z_comp + Z_rmd
│   ├── atac_preprocess_{cpu,gpu}.py  # QC → TF-IDF → HVF → LSI → dual Harmony → Z_comp + Z_rmd
│   ├── cell_type_{cpu,gpu}.py        # Leiden clustering on Z_comp (RNA or ATAC)
│   ├── ATAC_cell_type{,_gpu}.py      # ATAC-specific cell typing variants
│   ├── multi_omics_glue.py           # scGLUE integration (cross-modality VAE + guidance graph)
│   ├── multi_omics_batch_correction.py # Harmony post-pass on X_glue → Z_comp
│   ├── multi_omics_merge.py          # post-GLUE merge + per-modality preprocess/slimming
│   └── multi_omics_cell_type_{cpu,gpu}.py  # RNA-Leiden + k-NN label transfer to ATAC
├── sample_embedding/          # Core method
│   ├── blocks.py              # composition, RMD, weighting, Frobenius stack, final PCA + Harmony
│   ├── sample_embedding.py    # CPU pipeline
│   └── sample_embedding_gpu.py # GPU pipeline (cuML + cupy)
├── parameter_selection/
│   └── autotune.py            # Bayesian GP sweep over RMD α; adaptive proxy ensemble
├── sample_distance/           # Pairwise sample distances (DR / EMD / chi-square / JS)
├── sample_clustering/         # Hierarchical (HRA / HRC / NN / UPGMA / consensus), K-means, proportion test, RAISIN
├── sample_trajectory/         # CCA (supervised) and TSCAN (unsupervised) + GAM-based trajectory DGE
├── sample_association/        # Per-PC variance explained vs sample-level covariates (permutation FDR)
├── visualization/             # Embedding plots, dendrograms, DGE volcanos, modality-aware multi-omics scatters
├── utils/                     # Shared helpers: seed, safe h5ad I/O, limma, TF-IDF, batch regress, Grouping
├── gene_activity/             # ATAC peak → gene activity inference + RNA-ATAC validation
└── claude/                    # Active one-off run scripts (rerun launchers, monitored SE, parameter sweeps)

Usage

Complex mode (recommended) — YAML-driven

python SampleDisc.py -m complex --config config/config.yaml

The YAML drives every flag and parameter for all three pipelines:

• Pipeline gates (top-level): run_rna_pipeline, run_atac_pipeline, run_multiomics_pipeline
• Per-modality phase gates (Phase 1): *_preprocessing, *_cell_type_cluster, *_derive_sample_embedding
• Per-modality downstream gates (Phase 2): *_sample_distance_calculation, *_trajectory_analysis, *_trajectory_dge, *_sample_cluster, *_proportion_test, *_cluster_dge, *_visualize_data, *_dimension_association_analysis
• Multi-omics-specific: multiomics_run_glue_*, multiomics_treat_sample_as_batch, multiomics_run_glue_twice_for_sample_removal

The 9 ready-to-use configs in config/ are point-in-time snapshots for the datasets used in the paper; copy one and adjust paths / column names for your own data.

Simple mode — one positional file, defaults everywhere

python SampleDisc.py -m simple -c <count_data.h5ad> -o <output_dir>

Inputs

A standard scanpy AnnData file with at minimum:

• .X — count matrix (genes for RNA, peaks for ATAC)
• .obs['sample'] — sample column (required)
• Optional: .obs['batch'], .obs['cell_type'], sample-level metadata file (CSV) to merge

For multi-omics, the pipeline takes two separate h5ads (RNA + ATAC) and integrates them via scGLUE; samples may be paired (1:1 cell correspondence) or unpaired.

Outputs (under output_dir)

<output_dir>/
├── rna/
│   ├── preprocess/adata_preprocessed.h5ad
│   ├── sample_embedding/sample_embedding.csv
│   ├── Sample_distance/{cosine,correlation}/*
│   ├── CCA/  or  TSCAN/                       # whichever trajectory mode
│   ├── trajectoryDEG/
│   ├── sample_cluster/{kmeans_*,proportion_test/}
│   ├── sample_association/variance_explained_sample.csv + figures/
│   └── visualization/*.png
├── atac/   (parallel structure)
├── multiomics/
│   ├── integration/glue/{rna-pp,atac-pp,guidance.graphml.gz}
│   ├── preprocess/adata_sample.h5ad           # post-GLUE merged adata with Z_comp + Z_rmd
│   ├── sample_embedding/sample_embedding.csv
│   └── (same downstream subdirs as single-omics)
└── sys_log/main_process_status.json           # which stages completed

Installation

SampleDisco is one package. The CPU install is pip-only; GPU acceleration is activated simply by installing the GPU libraries separately — the same package detects and uses them at runtime. There is no separate "GPU build" of SampleDisco.

1. Core install (CPU)

pip install sampledisco          # once published — or `pip install -e .` from a clone

2. System prerequisite — bedtools

scGLUE (the multi-omics integrator) calls the bedtools binary, which pip cannot provide:

conda install -c bioconda bedtools

3. GPU acceleration (optional, install yourself)

The GPU functions (RAPIDS-accelerated normalization, Harmony, k-means / PCA, Leiden, scGLUE training) turn on only when the RAPIDS stack is present in your environment. RAPIDS is CUDA-driver-specific and conda-only, so you install it separately, matching your driver (the pins below target a CUDA-12.5 driver such as the cluster's GPU nodes):

conda install -c rapidsai -c conda-forge -c nvidia \
    cuml=24.12 cudf=24.12 cugraph=24.12 rmm=24.12 cuvs=24.12 cupy=13 cuda-version=12.5
pip install rapids-singlecell==0.13.1 --no-deps

Then set use_gpu: true in your config. You do not reinstall SampleDisco — once those packages are importable the GPU paths activate automatically; if they are missing or the driver is too old, SampleDisco falls back to CPU equivalents (harmonypy / linear regression, scikit-learn k-means, PyTorch CPU).

One-command environments (recommended)

For a fully reproducible environment (including bedtools), use the provided conda files instead of the manual steps above — see INSTALL.md for the driver/version notes:

conda env create -f environment-cpu.yml      # CPU
conda env create -f environment-gpu.yml      # GPU (RAPIDS 24.12)