CRISPRWorks

Open-source software for pooled CRISPR screen analysis.

CRISPRWorks connects guide counting, gene-effect inference and review through ordinary count tables and explicit sample metadata. CRISPRWorks Count is powered by DotMatch. CRISPRWorks Fit is a separate experimental source package for gene-level inference.

Components and status

Name	Task	Current status
CRISPRWorks Count	Turn FASTQ reads and a known guide library into counts and assignment QC	Available through DotMatch
CRISPRWorks Fit	Fit MAGeCK2 MLE effects, learn guide efficacy jointly across screens, and calibrate effects against control genes	Experimental alpha available from source
CRISPRWorks Review	Inspect QC, comparisons and screen results	Planned as a broader tool; DotMatch already provides count and assignment-QC reports

The Count documentation describes released DotMatch functionality. Fit is an experimental source package with a native MLE accelerator, a JACKS-derived joint model, whole-gene control calibration and reproducible offline demos. It has no PyPI release or independent lab evaluation yet. Read From counts to gene effects for input formats, model choice, outputs and interpretation. The broader Review tool is planned. The desktop Workbench is developed separately in dotmatch-community; its presence does not imply a released Review product.

Use Count today

Install the published package and use its existing command:

python3 -m pip install dotmatch==0.7.0
dotmatch --version
dotmatch demo --out-dir first-run/

Continue with the CRISPR guide-counting tutorial. Count writes MAGeCK-compatible counts for downstream analysis. It does not perform gene-level inference.

The DotMatch engine also supports barcode demultiplexing and other known-target short-DNA assays. Those capabilities remain documented under DotMatch.

Count benchmark spotlight: faster one-mismatch workflows

The benchmarked source checkout’s dotmatch guide-counter count is 1.6–5.5× faster with 11.3–14.4× lower peak memory than guide-counter 0.1.3, with identical full count matrices. The comparison uses controlled 100k/1M-read FASTQs, the public Yusa library, one CPU thread and five paired repeats. Guide-counter is faster in the tested exact-mode million-read cases. These measurements concern guide counting; gene-effect accuracy and full experimental-screen performance require separate validation. See both modes, graphs and reproduction commands.

Try Fit from source

From a repository checkout:

python3 -m pip install ./packages/crisprworks-fit
crisprworks-fit demo --out-dir fit-demo/
crisprworks-fit demo --model joint --out-dir fit-joint-demo/

Fit currently requires Python 3.10+, a C++ compiler and make for its MAGeCK2 dependency. The Fit README also covers the tested container build, count-table analysis and batch workflow. The default demo runs MLE. The joint demo creates counts, a sample map and 32 declared synthetic control genes, then fits two conditions and calibrates them with the default global BY adjustment. Both demos exercise software behavior; they do not validate biological hits. Use a new directory for the joint demo.

Choose a Fit workflow

Task	Command	Input and interpretation
Fit an explicit experimental design	`crisprworks-fit mle`	MAGeCK-compatible counts and a labeled design matrix; natural-log beta coefficients
Share guide-efficacy estimates across screens	`crisprworks-fit joint`	The same guide library, counts and a `Sample/Condition/Control` map; log2 effects and relative guide efficacies
Calibrate frozen fitted effects	`crisprworks-fit calibrate`	Full-precision MLE or joint JSON plus independently justified control genes; empirical p-values and BH/BY adjusted values

Count’s FASTQ sample sheet and Fit’s sample metadata serve different purposes. Count records which reads belong to a sample. Fit records which samples are baselines and biological replicates. Preserve sample and guide IDs at the handoff; do not treat sequencing lanes as independent biological replicates. Joint inference requires at least two replicates per treatment condition and 64 guides for variance smoothing. The calibrator defaults to 20 control genes per exact guide-count stratum, global adjustment and BY. It can return no hits when control counts limit p-value resolution.

What has been evaluated

The native MLE HAP1 measurement covered 17,445 complete four-guide gene labels. Three paired runs per engine measured 332.45 seconds for MAGeCK2 reference, 131.06 seconds for the NumPy accelerator and 38.68 seconds for the native accelerator: 8.59× versus reference on that cohort and runner. Printed summaries were identical. This is numerical and runtime evidence, not an improvement in biological hit accuracy.

The scientific comparison includes BAGEL2 on two HAP1 screens and current Chronos/JACKS on three cell lines in two libraries. Joint improved reference-essential ranking over Fit MLE in all six library/cell comparisons and exceeded Chronos in four. The panel is a curated, control-enriched subset; several paired intervals overlap zero. Those results do not establish overall state of the art.

Whole-gene calibration retains correlated guide behavior that pooled-guide resampling can lose. Its known-truth score simulation reports both false discoveries and power. Finite tails and multiple-testing adjustment require a representative null; they do not establish biological FDR control. Copy-number correction, complete independent screens and biological perturbation evidence remain outside the current validation.

Names and compatibility

Use CRISPRWorks for the family, and CRISPRWorks Count, CRISPRWorks Fit and CRISPRWorks Review for its components. Spell the family name as one word, with uppercase CRISPR and an uppercase W.

For the existing engine, use “CRISPRWorks Count, powered by DotMatch” on introductions and product pages. Use DotMatch for package names, commands, Python imports, methods and citations.

The branding change preserves the existing identifiers:

Identifier	Value
Python package and import	`dotmatch`
Command	`dotmatch`
GitHub repository	`dnncha/dotmatch`
Container	`ghcr.io/dnncha/dotmatch`
Documentation host	`dotmatch.readthedocs.io`
Release citation	Existing DotMatch release metadata and DOI

Users can keep their current installations and workflows. Record the actual DotMatch version and assignment policy in methods; the family name does not replace the versioned software citation.

Fit development criteria

The MLE accelerator preserves a pinned MAGeCK2 workflow; the joint engine and gene-control calibrator have separate model and inference checks. Before a broader production release or expanded scientific claim:

Profile a pinned upstream version on reproducible workloads.
Compare inference using identical frozen count matrices, design matrices, settings and permutation inputs. Record numerical tolerances and differences in effect estimates, standard errors, p-values and adjusted p-values.
Measure runtime on a documented environment. Measure peak memory before making memory claims; report workload-specific limits.
Obtain independent lab workflow evaluation and investigate discrepancies before claiming production readiness.

Evaluate guide-counting changes separately from inference changes. DotMatch count comparisons can help establish the former; they do not validate gene-level statistics. Upstream collaboration and compatibility with existing pipelines are development priorities.

Rollout

Introduce the family through the README and documentation first. Retain the published package names, links and citations while the additional components are evaluated. Any future command aliases or distribution changes need their own compatibility plan.

CRISPRWorks is the selected family name. Domain, package-registry and trademark availability have not been established; this document does not claim ownership of those names or namespaces.