`de-dgca` — Differential Expression & Differential Gene Correlation¶

contrastive no embedding no hotspot

The canonical perturbed-vs-control comparisons. This module produces no embedding and therefore does not run Hotspot — it is the only module whose structure differs from the other seven.

The two analyses¶

Differential expression¶

Which genes change in mean abundance between perturbed and control cells. The standard first question, and the standard limitation: it sees only shifts in mean.

Differential gene correlation¶

Which gene pairs change in correlation between the two conditions.

This is what DE cannot see. A perturbation can leave every gene's mean expression untouched while decoupling a regulatory relationship entirely — two genes tightly correlated in controls and independent after perturbation. The means are unchanged; the regulatory structure is not. DGCA detects exactly this.

flowchart TB
    A[(h5ad)] --> B["rule de_analysis<br/>run_de.py"]
    B --> C[de_results.csv]
    C --> D["rule dgca_analysis"]
    A --> D
    D --> E["prepare_dgca_inputs.py"]
    E --> F["run_dgca.R"]
    F --> G[dgca_avg_dcor.csv]

DGCA depends on the DE results, so the two rules run in sequence rather than in parallel. The DGCA step is implemented in R (run_dgca.R); its dependencies come from the module's environment.yaml.

Parameters¶

# de-dgca/config.yaml
de_pval_threshold: 0.05

dgca_cor_method: "pearson"
dgca_alpha: 0.05
dgca_min_genes: 10
dgca_use_de_genes_only: false
dgca_avg_method: "median"
dgca_n_perm: 100
dgca_filter_central: 0.01
dgca_filter_dispersion: 0.01

DE¶

Parameter	Type	Default	Description
`de_pval_threshold`	float	`0.05`	Significance cutoff for differential expression

DGCA¶

Parameter	Type	Default	Description
`dgca_cor_method`	str	`"pearson"`	Correlation method
`dgca_alpha`	float	`0.05`	Significance cutoff for differential correlation
`dgca_min_genes`	int	`10`	Minimum genes required to run
`dgca_use_de_genes_only`	bool	`false`	Restrict to DE genes before correlating
`dgca_avg_method`	str	`"median"`	Aggregation for average differential correlation
`dgca_n_perm`	int	`100`	Permutations for the null distribution
`dgca_filter_central`	float	`0.01`	Filter on central tendency
`dgca_filter_dispersion`	float	`0.01`	Filter on dispersion

dgca_n_perm and dgca_use_de_genes_only drive the cost

Differential correlation is computed over gene pairs, so cost is quadratic in genes and linear in dgca_n_perm. The default of 100 permutations bounds achievable p-values at roughly 0.01 — fine for ranking, too coarse for stringent significance claims. Raising it to 1,000 improves resolution at ten times the cost.

Setting dgca_use_de_genes_only: true restricts the pair space to DE genes and is by far the most effective way to make this module tractable on a large screen.

Rules¶

Rule	Produces
`de_analysis`	Differential expression results
`dgca_analysis`	Average differential correlation

Neither is a Hotspot rule — run_hotspot has no effect on this module.

Outputs¶

In results/<target>/:

File	Contents
`de_results.csv`	Per-gene differential expression statistics
`de_metadata.txt`	DE run parameters
`dgca_avg_dcor.csv`	Per-gene average differential correlation
`dgca_metadata.txt`	DGCA run parameters

The full pairwise DGCA matrix is not retained

Differential correlation is computed over gene pairs, so a complete pairwise result is quadratic in genes and was producing very large files. Nothing downstream of program creation consumes it, so only the per-gene average differential correlation, dgca_avg_dcor.csv, is written.

Mode behavior¶

Unlike the embedding modules, de-dgca passes the full target list through to its scripts in pooled mode via a --targets argument, so the pooled comparison is explicit about which perturbations were combined.

de-dgca — Differential Expression & Differential Gene Correlation¶