Comparison with other tools

This page explains when to use crispyx, when to use another tool, and what the practical differences are across common workflows.

crispyx vs Scanpy

Summary: crispyx produces identical results to Scanpy for t-test and Wilcoxon differential expression (Pearson r > 0.9999), but uses 2–43× less time and 2–6× less memory by streaming from disk instead of loading the full count matrix into RAM.

When to use crispyx instead of Scanpy

Your dataset has more than ~200,000 cells, or your count matrix exceeds the available RAM on your machine or HPC job.
You are running multiple DE methods on the same dataset and need reproducible, low-memory pipelines.
You want to keep data on disk between steps and avoid repeated loading.

When to use Scanpy instead of crispyx

Your dataset fits comfortably in RAM and you need a broad ecosystem of single-cell analysis tools beyond QC and DE (trajectory inference, RNA velocity, spatial transcriptomics, etc.).
You need tight integration with third-party Scanpy extensions that require in-memory AnnData.

Example: Wilcoxon DE on a large screen

# crispyx: streams from disk, minimal memory
import crispyx as cx
adata = cx.read_h5ad_ondisk("large_screen.h5ad")
adata = cx.pp.qc_summary(adata, perturbation_column="perturbation")
adata = cx.tl.rank_genes_groups(adata, perturbation_column="perturbation",
                                method="wilcoxon")

# Scanpy: loads entire matrix into RAM
import scanpy as sc
adata = sc.read_h5ad("large_screen.h5ad")  # may OOM on large datasets
sc.tl.rank_genes_groups(adata, groupby="perturbation", method="wilcoxon")

The crispyx result for logfoldchanges, scores, and p-values matches Scanpy with Pearson r > 0.9999 across all tested datasets.

Performance comparison

Dataset	crispyx Wilcoxon	Scanpy Wilcoxon
Adamson (65K cells)	123 s, 3.2 GB peak	1305 s, 3.3 GB peak
Feng-gwsf (321K)	10761 s, 17.2 GB peak	timeout (>9 hours)
Feng-gwsnf (1.97M)	completes	timeout / OOM

crispyx vs Pertpy / PyDESeq2

Summary: crispyx’s NB-GLM is approximately 2× faster than Pertpy/PyDESeq2 and uses substantially less memory. On genome-wide datasets, Pertpy/PyDESeq2 typically exceeds 120 GB peak memory or fails outright; crispyx completes within 64 GB. Pearson r > 0.97 vs PyDESeq2 for log-fold-change estimates.

When to use crispyx NB-GLM instead of Pertpy/PyDESeq2

Your dataset has many perturbation groups (hundreds to thousands) or many cells per group, making PyDESeq2’s per-group fitting memory-intensive.
You need NB-GLM DE on a genome-wide screen (>10,000 perturbations).
You want a streaming pipeline that integrates QC, pseudobulk, and DE in a single on-disk workflow.

When to use Pertpy/PyDESeq2 instead of crispyx

You need the full PyDESeq2 feature set (custom design matrices, contrasts, Cook’s distance filtering, etc.).
Your dataset is small enough that memory is not a constraint.
You need Pertpy’s other perturbation analysis methods (augur, mixscape, etc.).

Example: NB-GLM DE with crispyx

import crispyx as cx

adata = cx.read_h5ad_ondisk("screen.h5ad")
adata = cx.pp.qc_summary(adata, perturbation_column="perturbation")
adata = cx.tl.rank_genes_groups(
    adata,
    perturbation_column="perturbation",
    method="nb_glm",
    memory_limit_gb=32,
)

Performance comparison (NB-GLM)

Dataset	crispyx NB-GLM	Pertpy/PyDESeq2
Adamson (65K cells)	2472 s, 3.5 GB	5318 s, 33.8 GB
Feng-gwsf (321K)	completes	memory limit (>124 GB)
Feng-gwsnf (1.97M)	completes	memory limit / fails

crispyx vs edgeR

Summary: edgeR (via rpy2) times out or errors on most large CRISPR screen datasets tested in the benchmark. crispyx completes all 12 tested datasets. edgeR remains the standard for bulk RNA-seq and small pseudobulk experiments; for large perturbation screens with many groups, crispyx is more practical.

When to use edgeR instead of crispyx

You are working with bulk RNA-seq data or have a small number of perturbation groups (< 50) where edgeR’s quasi-likelihood F-test is well-calibrated.
You need an R-native workflow integrated with DESeq2, limma, or other Bioconductor packages.

Limitations of crispyx

crispyx does not replace the full Scanpy or Pertpy ecosystem. It focuses on QC, normalization, pseudobulk, and differential expression for large CRISPR screens. Trajectory inference, RNA velocity, and spatial transcriptomics are outside its scope.
The NB-GLM uses a simplified model (per-gene intercept and perturbation effect) without support for complex design matrices or batch covariates beyond what is in the current API.
On small datasets that fit in RAM, crispyx offers no meaningful advantage over Scanpy or Pertpy. Use Scanpy or Pertpy for datasets under ~50,000 cells if memory is not a concern.