CRISPR-centipede

Sensor-based per-variant effect scoring for base-editing tiling screens — the gRNA-only, frequentist companion to CRISPR-millipede. Where CRISPR-millipede models installed alleles from direct target amplicon-sequencing with a Bayesian linear model, CRISPR-centipede regresses a per-guide phenotype on per-guide editing values using ridge regression, and reports a per-variant effect-score table shaped like a millipede sigma table.

CRISPR-centipede is blind to how the editing was measured — it works the same whether the editing comes from a guide's self-edited protospacer, an attached sensor/surrogate, or a predictive model. Per guide it needs only:

1. editing information — a per-variant editing value (a guides × variants matrix);
2. counts in both populations — an enriched and a baseline read count;
3. coordinates — an optional coordinate per variant (only for position-aware estimators and plots).

Converting any particular assay (e.g. CRISPR-Correct observed mutations) into an editing matrix is done upstream, outside this package.

Install

pip install crispr-centipede           # once published
# or, from source:
cd CRISPR-centipede && pip install -e .

Python >=3.8,<3.12; depends only on numpy / scipy / pandas / scikit-learn / matplotlib.

Pipeline

import pandas as pd
import crispr_centipede as cc

# 1) Build a dataset: per-replicate editing matrix (guides x variants) + per-guide counts.
rep0 = cc.CentipedeReplicate(
    editing=editing_df0,        # DataFrame, index=guide id, columns=variant ids, values in [0,1]
    enriched=high_counts0,      # Series, index=guide id
    baseline=low_counts0,       # Series, index=guide id
)
ds = cc.CentipedeDataset([rep0, rep1, ...], coordinates=None)   # coordinates optional (parsed from ids if absent)

# 2) Model: ridge (default) with Wald / hat-matrix-df / permutation inference.
spec  = cc.CentipedeModelSpecification(estimator="ridge", inference=("wald", "hatdf", "perm"))
group = cc.CentipedeModelExperimentalGroup(ds, spec)

group.summary          # per-variant: PIP, Coefficient, Coefficient StdDev, wald_p, p_hatdf, p_perm, ...
group.metadata         # {alpha, cv_r2, n_guides, edf}
group.save("my_screen")  # -> my_screen_summary.csv + my_screen_result.pkl

# 3) Visualize (generic per-variant plots).
cc.visualization.plot_coefficient_board(group.summary)

Variant ids are opaque labels. If they follow the "{position}{ref}>{alt}" convention (e.g. "24A>G"), centipede can parse a coordinate from them; otherwise pass coordinates={variant: position, ...}. Coordinates are only needed by the position-aware estimators (sliding/segmented/fused) and plots — plain ridge/OLS don't use them.

Have your data as one combined frame (variant-frequency columns + count columns)? Use cc.CentipedeDataset.from_design_matrix(df, enriched_col=..., baseline_col=...).

Editing as counts, per-population edited counts, and a coverage filter

Instead of a [0,1] frequency you can pass raw counts and let the package compute the frequency (editing = presort_edited / presort_total). You may also supply the per-population edited counts to get per-variant editing rates, and apply a global guide-coverage filter (a guide failing a threshold on the pooled, summed-across-replicate count is dropped from every replicate):

ds = cc.CentipedeDataset.from_arrays(
    presort_edited=[edited_df0, ...],     # list[DataFrame]: guides x variants edited-read counts
    presort_total=[total_s0, ...],        # list[Series]:    per-guide total reads (the denominator)
    enriched=[high0, ...], baseline=[low0, ...],          # phenotype counts (Series per replicate)
    enriched_edited=[high_edited0, ...],  # optional: edited counts in the enriched population
    baseline_edited=[low_edited0, ...],   # optional: edited counts in the baseline population
    coordinates=variant_coords,
    min_enriched=10, min_baseline=10,     # optional global coverage filter (also min_presort_total)
)

ds.variant_enrichment()        # per-variant editing-weighted mean response (model-free "plain enrichment")
ds.population_editing_rate("enriched")   # per-variant pooled edited/total in the enriched population
ds.editing_rate_lfc()          # per-variant log2(enriched rate / baseline rate)
ds.subset_guides(guide_ids)    # restrict every replicate to a guide subset

Output (group.summary)

Indexed by variant id; columns mirror a CRISPR-millipede sigma table plus ridge-native fields:

Alternative estimators (non-default)

estimator="ols" (unbiased, valid t-test), or the position-aware variants "fused" / "sliding" / "segmented" via CentipedeModelSpecification(estimator=..., extra={...}).

Count normalization (optional)

count_normalization rescales the enriched (High) counts by a single pooled factor before scoring, to remove the enriched-vs-baseline library-size imbalance — the guide-count analog of CRISPR-millipede's WT-allele normalization:

• None (default) — no rescaling.
• "total" — factor Σ baseline / Σ enriched over the fitted guides (pins whole-library log2(H'/L') ≈ 0).
• "null_anchor" — factor computed over null_anchor_guide_ids (your negative-control guides), pinning their pooled log2(H'/L') ≈ 0. The controls define the scale only; they are not added as design rows.

# score relative to a negative-control null
cc.ridge_scores(ds, count_normalization="null_anchor", null_anchor_guide_ids=neg_control_ids,
                editing_threshold=None, coedit_threshold=None, degree=None)

Under the mean-centered Normal-Ridge/OLS fit this is ~a constant shift of the response, so β and hit calls change only at second order — its role is to anchor the reported response/β scale to a defined null, not to reshape the fit.

What this package is not

It does not include the Bayesian allele model (that's CRISPR-millipede), the parsing of any particular assay's raw output (do that upstream), nor the genome-coordinate β-logo track plotting (that lives with the analysis project and calls this package for the per-variant scores).