Highly variable genes#

bullkpy.pp.highly_variable_genes(adata, *, layer='log1p_cpm', n_top_genes=2000, n_bins=20, min_mean=0.0, max_mean=inf, min_disp=-inf, key_added='highly_variable')[source]#

Select highly variable genes (bulk-friendly version of HVGs).

This computes mean and variance across samples on the chosen layer and identifies genes with high dispersion relative to genes of similar mean (mean-binned z-score of log-dispersion, Scanpy/Seurat spirit).

Stores results in adata.var:

  • mean, variance, dispersion, dispersion_norm

  • boolean flag adata.var[key_added] (default: ‘highly_variable’)

Identify highly variable genes (HVGs) in bulk RNA-seq data.

Description#

highly_variable_genes selects genes whose expression varies strongly across samples, relative to genes of similar mean expression.

This function follows the Scanpy / Seurat spirit, but is adapted for bulk RNA-seq: variation is computed across samples, not cells, and works directly on normalized expression layers.

The method: 1. Computes per-gene mean and variance across samples 2. Calculates dispersion (variance / mean) 3. Log-transforms dispersion for numerical stability 4. Bins genes by mean expression 5. Computes a z-score of dispersion within each mean bin 6. Selects the top-ranked genes by normalized dispersion

Results are stored in adata.var.

Parameters#

adata
AnnData object with samples in .obs and genes in .var.

layer.

  • str | None

  • Layer used to compute variability (default: "log1p_cpm").
    If None, uses adata.X.

n_top_genes

  • int

  • Number of highly variable genes to select (default: 2000).

n_bins

  • int

  • Number of mean-expression bins used for dispersion normalization.

min_mean*

  • float

  • Minimum mean expression for a gene to be considered.

max_mean

  • float

  • Maximum mean expression for a gene to be considered.

min_disp

  • float

  • Minimum dispersion threshold.

key_added

  • str

  • Column name in adata.var where the HVG boolean flag is stored.

Returns#

Results are stored in adata.var.

The following columns are added to adata.var:

Column –> Description.
means –> Mean expression across samples.
variances –> Expression variance across samples.
dispersions –> Variance / mean.
dispersions_norm –> Mean-binned z-score of log-dispersion.
<key_added> –> Boolean flag indicating highly variable genes.

Example usage#

#Basic HVG selection
bk.pp.highly_variable_genes(adata)

# Using a different layer
bk.pp.highly_variable_genes(
    adata,
    layer="log2_tpm",
    n_top_genes=3000,
)

# Applying expression filters
bk.pp.highly_variable_genes(
    adata,
    min_mean=0.1,
    max_mean=5.0,
    min_disp=0.5,
)

# Subset AnnData to HVGs
adata = adata[:, adata.var["highly_variable"]].copy()

Notes#

•	Designed for bulk RNA-seq, not single-cell data
•	Works with dense or sparse matrices
•	Uses population variance (ddof=0)
•	Mean binning is performed on log1p(mean) to improve stability
•	Genes failing mean/dispersion thresholds are excluded from ranking