ORACLE scRNA-seq matrix-to-report workflow

ORACLE is a Snakemake workflow for reproducible downstream analysis of single-cell RNA-seq count matrices. It performs per-sample quality control and doublet detection, merges samples, normalizes counts, builds a low-dimensional representation, clusters cells at a configured biological resolution, ranks cluster markers, supports guarded pseudobulk differential expression, applies optional cell-type annotation, and produces a self-contained HTML report plus a run manifest.

This is deliberately a count-matrix-to-report workflow. Read alignment, UMI quantification, cell calling, demultiplexing, and ambient-RNA correction are upstream responsibilities. Optional FastQC, FastQ Screen, and MultiQC jobs form a parallel QC branch; they do not create or filter the expression matrix.

Start here

The complete beginner-friendly guide is TUTORIAL.md. It explains the biology and statistics as well as the commands, including experimental design, input validation, interpretation, troubleshooting, and limitations.

For a local run from the repository root:

conda env create -f envs/scrna.yaml
conda activate scrna

# Replace the example sample sheet and review every setting first.
$EDITOR config/samples.tsv
$EDITOR config/config.yaml

# Inspect the planned jobs before doing work.
snakemake --use-conda --cores 8 --dry-run

# Run and safely resume incomplete jobs if necessary.
snakemake --use-conda --cores 8 --rerun-incomplete --printshellcmds

When the workflow completes, open results/report.html and inspect results/run_manifest.json. A successful command is the start of review, not proof that the biological interpretation is correct.

Input contract

Each row of config/samples.tsv represents one independent sample or capture. Every expression input must be one of:

• an .h5ad file whose X matrix contains nonnegative integer raw counts;
• a filtered 10x .h5 count matrix; or
• a filtered 10x MEX directory.

All samples in one run must use the same species and the identical named gene universe from one reference annotation; gene order may differ. The workflow stops rather than turning absent features into false zero counts. Normalized, log-transformed, scaled, or integrated values are not valid inputs. Unfiltered droplet matrices must first undergo cell calling outside this workflow.

For condition-level inference, sample must identify the biological replicate. Cells are observations within a replicate; they are not independent replicates.

Workflow

validated raw count matrices
        |
        +-- optional raw-FASTQ QC branch (independent; no quantification)
        |
        v
per-sample cell QC -> per-sample doublet detection
        |
        v
merge -> normalize -> HVGs/PCA -> optional batch correction
        |
        v
neighbors/UMAP -> Leiden diagnostics + configured resolution
        |
        v
cluster markers -> guarded pseudobulk DE -> optional annotation
        |
        +-- optional PAGA / scVelo
        |
        v
HTML report + tables + H5AD objects + run manifest

Scientific guardrails

• QC thresholds are starting points, not universal biological truths. Review every sample's distributions and retention.
• Correct only genuine technical batches. If batch and condition are confounded, integration cannot recover the missing design.
• UMAP is a visualization; its distances, islands, and apparent paths are not statistical evidence.
• The configured Leiden resolution is the reported clustering. Silhouette values across candidate resolutions are diagnostics, not an automatic definition of cell types.
• Cluster-marker tests rank features for annotation. Their cell-level p-values are not condition-level inference.
• Condition DE uses raw-count pseudobulks at the biological-sample level. Completed runs record per-cluster outcomes in pseudobulk_status.csv; if no requested comparison is eligible, the rule fails and prints the status table to its durable log.
• Annotation is opt-in. An incompatible marker dictionary fails fast at the coverage check; weak cluster scores, score margins, or classifier confidence remain Unknown. A manual map must cover every cluster and may explicitly assign Unknown.
• Cell-type composition tables are descriptive. Replicate-aware differential-abundance testing is outside the workflow.
• PAGA and RNA velocity are optional hypothesis-generating analyses with strict biological and identifier requirements.

Main outputs at the default paths

| Output | Meaning | |---|---| | results/report.html | Self-contained visual summary for review | | results/run_config.yaml | Effective configuration after Snakemake overrides | | results/run_manifest.json | Configuration/input fingerprints, Conda-spec hashes, detected report-environment packages, Git SHA, seed, and warnings | | results/<sample>/00_qc/ | Per-sample QC object, statistics, and figures | | results/<sample>/01_doublets/ | Per-sample singlet object and doublet diagnostic | | results/merged/06_cluster/ | Clustered object and resolution diagnostics | | results/merged/07_markers/ | Cluster markers, pseudobulk results, and DE status table | | results/merged/08_annotate/ | Annotated object, annotation summary, and composition table | | results/merged/09_trajectory/ | Optional PAGA and velocity artifacts | | results/multiqc/multiqc_report.html | Optional raw-FASTQ QC summary or explicit no-FASTQ notice | | logs/ | One log per rule; inspect these first when a job fails or is skipped |

Reference documentation

• Complete tutorial
• Single-cell best-practices reference

License

MIT — see LICENSE.