Preprocessing API Reference¶

This module provides comprehensive data preprocessing functions for spatial transcriptomics data analysis in SPEX.

Functions¶

preprocess¶

preprocess(adata, scale_max=10, size_factor=None, do_QC=False)

Comprehensive preprocessing pipeline for AnnData objects.

Performs quality control, normalization, feature selection, and scaling in a single function call.

Parameters:

Parameter	Type	Default	Description
`adata`	AnnData	-	AnnData object to preprocess
`scale_max`	float	10	Maximum value for scaling (clips larger values)
`size_factor`	float	None	Size factor for normalization. If None, uses median library size
`do_QC`	bool	False	Whether to perform quality control filtering

Returns:

Type	Description
AnnData	Preprocessed AnnData object

Notes:

The preprocessing pipeline includes the following steps:

Quality Control (if do_QC=True):
Filter cells with low gene counts using MAD threshold
Filter genes with low cell counts
Remove cells with low total counts
Normalization:
Calculate size factors (median library size if not provided)
Normalize total counts per cell
Apply log1p transformation
Feature Selection:
Identify highly variable genes
Account for batch effects if present
Scaling:
Z-transform data
Clip values to scale_max

Examples:

import spex as sp
import scanpy as sc

# Load data
adata = sc.read_h5ad("data.h5ad")

# Basic preprocessing
adata = sp.preprocess(adata)

# Preprocessing with quality control
adata = sp.preprocess(
    adata,
    scale_max=10,
    do_QC=True
)

# Custom size factor
adata = sp.preprocess(
    adata,
    size_factor=10000,
    do_QC=True
)

Storage:

Results are stored in the AnnData object: - adata.raw: Original data backup - adata.uns['prepro']: Preprocessing parameters - adata.var.highly_variable: Highly variable gene flags - adata.layers['counts']: Normalized count data

MAD_threshold¶

MAD_threshold(variable, ndevs=2.5)

Calculate threshold using Median Absolute Deviation (MAD).

This function calculates a robust threshold value using the MAD method, which is less sensitive to outliers than standard deviation-based methods.

Parameters:

Parameter	Type	Default	Description
`variable`	array-like	-	Input data array
`ndevs`	float	2.5	Number of MAD deviations from the median

Returns:

Type	Description
float	Threshold value calculated as median - ndevs * MAD

Notes:

MAD is calculated as median(|x - median(x)|)
The threshold is computed as median - ndevs * MAD
This method is robust to outliers and non-normal distributions
Commonly used for quality control filtering in single-cell data

Examples:

import spex as sp
import numpy as np

# Calculate threshold for total counts
total_counts = adata.obs['total_counts']
threshold = sp.MAD_threshold(total_counts, ndevs=2.5)

# Filter cells below threshold
filtered_adata = adata[adata.obs['total_counts'] > threshold, :]

# Use different threshold for genes
n_genes = adata.obs['n_genes_by_counts']
gene_threshold = sp.MAD_threshold(n_genes, ndevs=1.5)

should_batch_correct¶

should_batch_correct(adata)

Check if batch correction should be performed.

This function checks if batch correction is needed by looking for batch information in the AnnData object.

Parameters:

Parameter	Type	Description
`adata`	AnnData	AnnData object to check for batch information

Returns:

Type	Description
bool	True if batch correction should be performed, False otherwise

Notes:

Checks for 'batch_key' in adata.uns
Returns True if batch_key exists and is not None
Used internally by preprocessing functions to determine batch correction strategy

Examples:

import spex as sp

# Check if batch correction is needed
needs_batch_correction = sp.should_batch_correct(adata)

if needs_batch_correction:
    print(f"Batch correction will be performed using: {adata.uns['batch_key']}")
else:
    print("No batch correction needed")

reduce_dimensionality¶

reduce_dimensionality(adata, prefilter=False, method='pca', mdist=0.5, n_neighbors=None, latent_dim=None)

Reduce dimensionality of AnnData object.

This function performs dimensionality reduction using various methods including PCA, UMAP, t-SNE, and scVI. It also constructs neighborhood graphs for downstream analysis.

Parameters:

Parameter	Type	Default	Description
`adata`	AnnData	-	AnnData object to reduce dimensionality
`prefilter`	bool	False	Whether to prefilter highly variable genes
`method`	str	'pca'	Dimensionality reduction method. Options: 'pca', 'umap', 'tsne', 'scvi', 'diff_map'
`mdist`	float	0.5	Minimum distance for UMAP
`n_neighbors`	int	None	Number of neighbors for UMAP/t-SNE. If None, uses sqrt(n_cells)
`latent_dim`	int	None	Number of latent dimensions. If None, uses automatic estimation

Returns:

Type	Description
AnnData	Updated AnnData object with dimensionality reduction results

Notes:

Supported Methods:

PCA: Principal Component Analysis (fastest, most common)
UMAP: Uniform Manifold Approximation and Projection (good for visualization)
t-SNE: t-Distributed Stochastic Neighbor Embedding (good for visualization)
scVI: Single-Cell Variational Inference (good for batch correction)
diff_map: Diffusion maps (good for trajectory analysis)

Automatic Parameter Estimation:

n_neighbors: If None, uses sqrt(n_cells)
latent_dim: If None, uses elbow method on singular values

Batch Correction:

If batch information is present, applies Harmony integration
Results stored in adata.obsm['X_pca_harmony']

Examples:

import spex as sp

# Basic PCA reduction
adata = sp.reduce_dimensionality(adata, method='pca')

# UMAP with custom parameters
adata = sp.reduce_dimensionality(
    adata,
    method='umap',
    n_neighbors=30,
    latent_dim=50,
    mdist=0.3
)

# scVI for batch correction
adata = sp.reduce_dimensionality(
    adata,
    method='scvi',
    latent_dim=20
)

Storage:

Results are stored in the AnnData object: - adata.obsm['X_pca']: PCA coordinates - adata.obsm['X_umap']: UMAP coordinates - adata.obsm['X_tsne']: t-SNE coordinates (if computed) - adata.obsm['X_scvi']: scVI latent space (if computed) - adata.obsp['connectivities']: Neighborhood graph - adata.obsm['distances']: Distance matrix

load_anndata¶

load_anndata(path=None, files=None)

Load AnnData objects from file(s).

This function loads AnnData objects from single or multiple files. When loading multiple files, they are combined into a single AnnData object.

Parameters:

Parameter	Type	Default	Description
`path`	str	None	Path to a single AnnData file (.h5ad)
`files`	list	None	List of file paths to load and combine

Returns:

Type	Description
dict	Dictionary containing the loaded AnnData object(s): 'adata': AnnData object

Notes:

If both path and files are provided, files takes precedence
When combining multiple files, a 'filename' column is added to adata.obs
Files are combined using scanpy.concat with outer join
Supports .h5ad format files

Examples:

import spex as sp

# Load single file
result = sp.load_anndata(path="data.h5ad")
adata = result['adata']

# Load multiple files
files = ["sample1.h5ad", "sample2.h5ad", "sample3.h5ad"]
result = sp.load_anndata(files=files)
combined_adata = result['adata']

# Check combined data
print(f"Combined shape: {combined_adata.shape}")
print(f"Files included: {combined_adata.obs['filename'].unique()}")

Best Practices¶

Quality Control¶

Always perform QC before analysis:

adata = sp.preprocess(adata, do_QC=True)

Check data quality metrics:

import scanpy as sc
sc.pp.calculate_qc_metrics(adata, inplace=True)
print(f"Cells: {adata.n_obs}, Genes: {adata.n_vars}")

Use appropriate thresholds:

# Conservative thresholds for high-quality data
threshold_counts = sp.MAD_threshold(adata.obs['total_counts'], ndevs=2.0)
threshold_genes = sp.MAD_threshold(adata.obs['n_genes_by_counts'], ndevs=1.5)

Dimensionality Reduction¶

Choose appropriate method:
PCA: For fast preprocessing and clustering
UMAP: For visualization and exploration
scVI: For batch correction and integration

Estimate parameters automatically:

# Let the function estimate optimal parameters
adata = sp.reduce_dimensionality(adata, method='pca')

Check results:

# Verify dimensionality reduction worked
print(f"PCA components: {adata.obsm['X_pca'].shape}")
print(f"UMAP coordinates: {adata.obsm['X_umap'].shape}")

Batch Correction¶

Set up batch information:
```
adata.uns['batch_key'] = 'batch'
```

Use batch-aware preprocessing:

# Preprocessing will automatically detect and handle batches
adata = sp.preprocess(adata, do_QC=True)

Verify batch correction:

# Check if batch correction was applied
if 'X_pca_harmony' in adata.obsm:
    print("Batch correction applied successfully")

Troubleshooting¶

Common Issues¶

Memory errors during preprocessing:
Reduce scale_max parameter
Use data subsampling
Increase available RAM
Slow dimensionality reduction:
Use prefilter=True to reduce gene number
Choose faster methods (PCA over UMAP)
Reduce latent_dim parameter
Batch correction not working:
Ensure adata.uns['batch_key'] is set
Check that batch column exists in adata.obs
Verify batch column contains valid values
File loading errors:
Check file paths are correct
Ensure files are in .h5ad format
Verify file permissions

Performance Optimization¶

Use appropriate data types:

# Convert to float32 for memory efficiency
adata.X = adata.X.astype('float32')

Subsample large datasets:

# Random subsampling
import scanpy as sc
sc.pp.subsample(adata, n_obs=10000, random_state=42)

Save intermediate results:

# Save after preprocessing
adata.write("preprocessed_data.h5ad")

Integration with Other Modules¶

With Clustering¶

# Complete preprocessing for clustering
adata = sp.preprocess(adata, do_QC=True)
adata = sp.reduce_dimensionality(adata, method='pca')
adata = sp.cluster(adata, method='leiden')

With Spatial Analysis¶

# Preprocessing for spatial analysis
adata = sp.preprocess(adata, do_QC=True)
adata = sp.reduce_dimensionality(adata, method='pca')
# Spatial analysis functions can now use preprocessed data

With Segmentation Results¶

# Load and preprocess feature extraction results
adata = sp.load_anndata(path="segmentation_features.h5ad")['adata']
adata = sp.preprocess(adata, do_QC=True)