bulk-transcriptomics

Bulk Transcriptomics

Method selection and execution guidance for bulk RNA-seq and microarray differential expression analysis.

Method Selection Decision Tree

Choose the DE method based on input data type, sample size, and experimental design:

Input data?
├── Raw integer counts (RNA-seq)
│   ├── Simple design (2 conditions, no interaction terms)
│   │   ├── n >= 3 per group, n <= 50 per group → PyDESeq2 (Python, default)
│   │   ├── n = 2 per group → edgeR QLF via rpy2 (better small-sample performance)
│   │   ├── No biological replication in any group (1 vs 1) → NO inferential DE.
│   │   │     Report descriptive log2 fold changes only, and state why
│   │   │     (see Anti-Patterns)
│   │   └── n > 50 per group → limma-voom via rpy2 (faster, scales well)
│   ├── Complex design (interaction terms, >2 factors, nested)
│   │   ├── Standard factorial/interaction → DESeq2 via rpy2 (full formula support)
│   │   └── Large n or many covariates → limma-voom via rpy2
│   ├── Longitudinal / repeated measures
│   │   └── dream (variancePartition) via rpy2 (mixed-effects voom)
│   └── Batch effects present
│       ├── Known batches → sva ComBat_seq on raw counts, then DE as above
│       └── Unknown confounders → svaseq to estimate surrogate variables, include in model
├── Pre-normalized data (TPM, FPKM, RPKM, log-CPM, microarray intensities)
│   └── limma via rpy2 (do NOT use DESeq2/edgeR — they require raw counts)
└── Raw microarray CEL / intensity files
    └── Out of scope. Reading them resolves a platform design package per array
        design, which cannot be staged for every design and cannot be installed.
        Report that and ask for the normalized expression matrix — every array
        platform publishes one, and it enters the branch above.

Workflow Phases

1. Data ingestion: Load count matrix + sample metadata. Verify counts are raw integers.
2. QC: Library size distribution, gene detection rates, PCA for outlier detection.
3. Filtering: Remove low-count genes (e.g., keep genes with >= 10 counts in >= n samples where n is the smallest group size).
4. Batch assessment: PCA colored by batch — if batch clusters dominate, apply correction.
5. Normalization: Handled internally by each method (DESeq2 median-of-ratios, edgeR TMM, limma-voom). Do NOT pre-normalize.
6. DE testing: Apply the method from the decision tree. Extract results with log2FC, p-value, adjusted p-value.
7. Downstream: Pathway enrichment via decoupler, volcano/MA plots.