Environment Setup

To get started, first create the environment using conda or mamba. This will install snakemake, snakedeploy and a version of scanpy that is compatible with the workflow. Note that anndata is pinned to an older version to ensure compatibility with the selected scanpy version.

mamba create -c conda-forge -c bioconda --name snakemake_base snakemake snakedeploy scanpy=1.10.2 anndata=0.10.2

Activate the environment.

mamba activate snakemake_base

Use the workflow (clone repository)

git clone https://github.com/Campello-Lab/Clustering-benchmarking-for-scRNAseq
cd Clustering-benchmarking-for-scRNAseq

Configure Workflow:

Configuration file config/config.yaml contains configuration that will enable reproduction of the benchmarking performed in the paper. Edit config/config.yaml to adjust settings. You can comment out datasets to run the workflow on a smaller collection.

Running the workflow

Option A: Run Entire Workflow (Including Data Download)

snakemake --cores all --sdm conda

Output Structure

data/raw_data/

Original dataset files downloaded from the sources provided by the dataset authors.

Examples include:

• matrix files
• metadata tables
• raw count matrices

---

data/processed_adata/

Datasets converted into standardized AnnData (.h5ad) format.

These files contain:

• raw count matrices
• cell metadata
• gene metadata

No normalization or feature selection has been applied at this stage.

---

results/benchmark_analysis/processed_normalized_adata/

Preprocessed and normalized AnnData objects used for benchmarking.

This folder contain files after application of:

• preprocessing steps
• selection of 1,000 highly variable genes (HVGs)

---

After preprocessing is completed, the workflow reevaluates the DAG based on the selected algorithms and hyperparameter combinations specified in config.yaml.

The resulting benchmarking outputs are stored in:

---

results/benchmark_analysis/pairwise_similarity/

Pairwise similarity matrices stored in:

adata.obsp["distances_full"]

---

results/benchmark_analysis/graphs/

k-nearest neighbor (kNN) graphs used for clustering.

Stored in:

adata.obsp["distances"]
adata.obsp["connectivities"]

where:

• distances contains graph distances
• connectivities contains edge weights

---

results/benchmark_analysis/clustering/

Clustering assignments for all evaluated hyperparameter combinations.

Cluster labels are stored in:

adata.obs

with keys formatted as:

<quality_function>_<resolution>

Example:

adata.obs["CPMVertexPartition_0.002"]

---

results/benchmark_analysis/evaluation/

Evaluation results for all clustering partitions across multiple clustering evaluation measures.

---

Option B: Skipping Data Download

To avoid rerunning raw data download and conversion steps, download the prepared AnnData (.h5ad) files from Zenodo.

---

These files already contain raw count matrices and metadata in standardized AnnData format, but have not been normalized, filtered, or otherwise preprocessed for clustering.

1. Download and extract processed_adata.zip from Zenodo.
2. Place it into your project working directory under:

data/processed_adata/

📫 Contact

For questions or support, please open an issue or contact: alszm@imada.sdu.dk.