CRISPR-Correct

CRISPR-Correct (Python package crispr-ambiguous-mapping) was developed by the Pinello Lab as an easy-to-use Python package for performing guide-RNA mapping from raw FASTQs against a guide-library DataFrame. Specifically, CRISPR-Correct handles imperfect mapping — either from self-editing by SpRY-based base-editors or sequencing errors — by mapping observed protospacer sequences to the guide RNA with the closest Hamming distance. It cannot handle indels in the protospacer (for that, consider the Pinello Lab tool CRISPR BEAN). If you aren't expecting self-editing or sequencing error, this tool still works but CRISPR SURF may be faster.

CRISPR-Correct also handles guide-RNA sensor / surrogate constructs, UMIs, and single-cell sample barcodes (see Figure 1). Mapping is performed on each protospacer / surrogate / barcode permutation and the editing outcomes at the protospacer and surrogate are characterized. You provide regex strings or positional specifications for extracting each component from the read.

Figure 1. Schematic of a guide-RNA sensor construct. Typically, the expressed guide RNA edits both the endogenous target site and the surrogate target site. A short barcode distinguishes between similar protospacer / surrogate sequences in the guide-RNA library. Paired-end sequencing is performed to capture the protospacer (R1) and the surrogate + barcode (R2).

If you are mapping many large samples that would take too long on a personal computer, CRISPR-Correct can also run on the Broad Institute's Terra Platform. The workflow file is at the Terra Firecloud repository.

Migrating from 0.0.x? See ../USAGE.md §6 for the CountInput rename + post-processing kwarg rename table, and ../CHANGELOG.md for the accumulated 0.0.236 → Unreleased changes (new API surface, CLI, parquet save/load, memory + performance deltas).

Installation

pip install crispr-ambiguous-mapping

Development install (for contributing):

git clone https://github.com/pinellolab/CRISPR-Correct.git
cd CRISPR-Correct/crispr-ambiguous-mapping
pip install -e .

Python constraint: >=3.8, <3.12.

System requirements

Any OS that can run a supported Python version. Multi-core CPU recommended (the inference step parallelizes via multiprocessing.Pool); sufficient RAM for in-memory Counter of unique observed sequences; SSD recommended for large FASTQ I/O. pandarallel, biopython, pysam, anndata are pulled in as dependencies.

---

Prepare inputs

The three inputs are:

1. Demultiplexed R1 (and R2) FASTQ(s). See the tip below if your reads arrive with in-read indices that need extra demultiplexing.
2. Guide library TSV with columns protospacer (required), surrogate (optional), barcode (optional). Lengths may differ per column but must be uniform within a column.
3. Parsing specifications that tell the tool where each component lives within each read (or the FASTQ header).

Tip: if indices need to be parsed out of the read before demultiplexing (i.e. in-read indices), use UMItools and BBMap demuxbyname. A Terra Firecloud method runs both.

Guide library TSV (example)

protospacer            surrogate                         barcode
TGTCGTGAGGTAGCTACGAC   CAGCAATGTCGTGAGGTAGCTACGACTTGTCA  GCTC
AGTCGTAGCTACCTCACGAC   ATGACAAGTCGTAGCTACCTCACGACATTGCT  GTTG
CCTAGTGGTTATTCGATGTC   AGGTTACCTAGTGGTTATTCGATGTCTCAGAA  CGAA
...

Header regexes (example)

Suppose UMI-tools put the sample barcode + UMI into the FASTQ header:

@lh00134:140:225VLGLT3:7:1101:1028:1080_ANGC_GGCA 1:N:0:GAAATAAG+ACGTCCTG

• Guide barcode regex: r"_([^_ ]+)[\s+]" captures GGCA
• UMI regex (8 bp UMI variant): r":([^+:]{8})(.{2})\+" captures GAAATAAG
• UMI regex (6 bp UMI variant): r":([^+:]{6})(.{2})\+"

Hamming thresholds

For a 20 bp protospacer, 32 bp surrogate, and 4 bp barcode, these defaults work well across published CRISPR-Correct runs:

PROTOSPACER_HAMMING_THRESHOLD = 7
SURROGATE_HAMMING_THRESHOLD   = 10
BARCODE_HAMMING_THRESHOLD     = 2

These are guidelines for a canonical base-editor editing window. As long as numbers are not too low (which would discard true edits) or too high (which would bring in random reads), they don't need fine tuning.

---

Running the guide mapping

Primary entry point

from crispr_ambiguous_mapping.mapping import get_whitelist_reporter_counts_from_fastq

The function accepts a guide-library DataFrame, one or more R1 FASTQ paths, optional R2 FASTQ paths, a per-component parsing spec (regex or left/right flanks or fixed start/end positions), location flags (R1 body / R2 body / FASTQ header), reverse-complement flags, per-component Hamming thresholds, and parallelism controls:

def get_whitelist_reporter_counts_from_fastq(
    whitelist_guide_reporter_df,
    fastq_r1_fns,                      # list of R1 FASTQ paths
    fastq_r2_fns=None,                 # optional list of R2 FASTQ paths

    # Parsing -- per component, ONE of: regex, left+right flank, start+end position
    protospacer_pattern_regex=None,
    surrogate_pattern_regex=None,
    guide_barcode_pattern_regex=None,
    guide_umi_pattern_regex=None,
    sample_barcode_pattern_regex=None,

    protospacer_left_flank=None,       protospacer_right_flank=None,
    protospacer_start_position=None,   protospacer_end_position=None,
    protospacer_length=None,

    surrogate_left_flank=None,         surrogate_right_flank=None,
    surrogate_start_position=None,     surrogate_end_position=None,
    surrogate_length=None,

    guide_barcode_left_flank=None,     guide_barcode_right_flank=None,
    guide_barcode_start_position=None, guide_barcode_end_position=None,
    guide_barcode_length=None,

    guide_umi_left_flank=None,         guide_umi_right_flank=None,
    guide_umi_start_position=None,     guide_umi_end_position=None,
    guide_umi_length=None,

    sample_barcode_left_flank=None,    sample_barcode_right_flank=None,
    sample_barcode_start_position=None, sample_barcode_end_position=None,
    sample_barcode_length=None,

    # Location flags -- where does each component live?
    is_protospacer_r1=None,    is_surrogate_r1=None,
    is_guide_barcode_r1=None,  is_guide_umi_r1=None,   is_sample_barcode_r1=None,

    is_protospacer_header=None,   is_surrogate_header=None,
    is_guide_barcode_header=None, is_guide_umi_header=None, is_sample_barcode_header=None,

    # Reverse complement (usually True for any component read off R2)
    revcomp_protospacer=None,    revcomp_surrogate=None,
    revcomp_guide_barcode=None,  revcomp_guide_umi=None,   revcomp_sample_barcode=None,

    # Hamming thresholds (pass-through to the internal Hamming matcher)
    surrogate_hamming_threshold_strict=None,
    guide_barcode_hamming_threshold_strict=None,
    protospacer_hamming_threshold_strict=None,

    # Output control -- NEW
    retain_inference_results=False,    # see "Slim vs full output" below

    # Internal
    store_intermediates=False,
    cores=1,
) -> WhitelistReporterCountsResult

Parameter naming conventions:

• fastq_r1_fns / fastq_r2_fns — always lists, even with one file. This is the current multi-sample-support API.
• guide_barcode_*, guide_umi_*, sample_barcode_* — prefix distinguishes a guide's library-encoded barcode/UMI from a per-read (single-cell) sample barcode.
• is_<comp>_r1=True extracts from R1 body; is_<comp>_header=True extracts from the R1 FASTQ header. If both are False / None and fastq_r2_fns is provided, the component is extracted from R2.

Slim vs full output

By default the returned object holds the count Series + QC summary + input config — a lean object that pickles in ~150 KB on typical runs.

If you need allele/mutation post-processing (get_matchset_alleleseries, get_mutation_profile, helper_get_observed_values_given_whitelist_value), pass retain_inference_results=True. The result then also carries the full per-observation inference dict (observed_guide_reporter_umi_counts_inferred), which can be several GB on 45 M-read runs.

If you call a post-processing function on a slim result you get a clear error:

ValueError: get_matchset_alleleseries requires `observed_guide_reporter_umi_counts_inferred`
but the result object was built with `retain_inference_results=False` (the default).
Re-run the mapping call with `retain_inference_results=True` to enable
allele / mutation post-processing.

Example -- full triplet with UMI in header

import crispr_ambiguous_mapping as cam
import pandas as pd

whitelist_guide_reporter_df = pd.read_table("guide_library.tsv")

result = cam.mapping.get_whitelist_reporter_counts_from_fastq(
    whitelist_guide_reporter_df=whitelist_guide_reporter_df,
    fastq_r1_fns=["sample_R1.fastq.gz"],
    fastq_r2_fns=["sample_R2.fastq.gz"],

    # Protospacer: first 20 bp of R1
    protospacer_start_position=0, protospacer_length=20,
    is_protospacer_r1=True, revcomp_protospacer=False,

    # Surrogate: first 32 bp of R2 (reverse-complemented)
    surrogate_start_position=0, surrogate_length=32,
    is_surrogate_r1=False, revcomp_surrogate=True,

    # Guide barcode: parsed from R1 header via regex
    guide_barcode_pattern_regex=r"_([^_ ]+)[\s+]",
    is_guide_barcode_header=True, revcomp_guide_barcode=True,

    # UMI: parsed from R1 header via regex
    guide_umi_pattern_regex=r":([^+:]{8})(.{2})\+",
    is_guide_umi_header=True, revcomp_guide_umi=False,

    # Thresholds
    protospacer_hamming_threshold_strict=7,
    surrogate_hamming_threshold_strict=10,
    guide_barcode_hamming_threshold_strict=2,

    cores=8,

    # If you want to run allele / mutation post-processing on this result,
    # uncomment the next line. Default is False (slim result).
    # retain_inference_results=True,
)

Accessing counts

The main output lives at:

result.all_match_set_whitelist_reporter_counter_series_results

which has one attribute per match tier — protospacer_match, protospacer_match_surrogate_match, protospacer_match_barcode_match, protospacer_match_surrogate_match_barcode_match, and two mismatch tiers — each with nine pd.Series counters (3 ambiguity strategies × 3 UMI modes).

# Example: UMI-collapsed, ambiguous-accepted full-triplet counts
series = (
    result.all_match_set_whitelist_reporter_counter_series_results
          .protospacer_match_surrogate_match_barcode_match
          .ambiguous_accepted_umi_collapsed_counterseries
)
series.head()

QC summary:

result.quality_control_result.protospacer_match.num_non_error_umi_noncollapsed_counts
result.quality_control_result.protospacer_match.num_total_umi_noncollapsed_counts

Post-processing (allele + mutation profiles)

When retain_inference_results=True, you can further analyze per-allele observations:

match_set = cam.processing.get_matchset_alleleseries(
    result.observed_guide_reporter_umi_counts_inferred,
    "protospacer_match_surrogate_match_barcode_match",
    contains_surrogate=result.count_input.contains_surrogate,
    contains_barcode=result.count_input.contains_guide_barcode,
    contains_umi=result.count_input.contains_guide_umi,
)

mutations = cam.processing.get_mutation_profile(
    match_set,
    whitelist_reporter_df=result.count_input.whitelist_guide_reporter_df,
    contains_surrogate=result.count_input.contains_surrogate,
    contains_barcode=result.count_input.contains_guide_barcode,
)

linked = cam.processing.tally_linked_mutation_count_per_sequence(
    mutations,
    contains_surrogate=result.count_input.contains_surrogate,
    contains_barcode=result.count_input.contains_guide_barcode,
    count_attribute_name="ambiguous_accepted_umi_noncollapsed_mutations",
)

cam.visualization.plot_mutation_count_histogram(
    linked.protospacer_total_mutation_counter,
    filename="protospacer_mut_hist.png",
)

Saving + reloading

Standard Python pickling works. With the slim default (retain_inference_results=False), pickle sizes are much smaller:

import pickle

with open("result.pkl", "wb") as f:
    pickle.dump(result, f, protocol=pickle.HIGHEST_PROTOCOL)

Intermediate save utilities (e.g. cam.utility.save_or_load_pickle) remain available.

---

Command-line interface (v0.1.0)

The package registers a crispr-correct console script with two subcommands. All flags mirror the Python ParsingConfig fields 1:1 (field names with _ → -) — no config file, no YAML.

# Map
crispr-correct map \
    --r1 R1.fq.gz \
    --r2 R2.fq.gz \
    --library library.tsv \
    --out result.pickle \
    --protospacer-start-position 0 --protospacer-length 20 \
    --is-protospacer-r1 --no-revcomp-protospacer \
    --protospacer-hamming-threshold-strict 7 \
    --surrogate-start-position 0 --surrogate-length 32 \
    --no-is-surrogate-r1 --revcomp-surrogate \
    --surrogate-hamming-threshold-strict 10 \
    --guide-barcode-pattern-regex '_([^_ ]+)[\s+]' \
    --is-guide-barcode-header --revcomp-guide-barcode \
    --guide-barcode-hamming-threshold-strict 2 \
    --retain-inference-results \
    --cores 4

# Emit one tier's count Series as TSV
crispr-correct count \
    --in result.pickle \
    --out counts.tsv \
    --tier protospacer_match_surrogate_match_barcode_match \
    --strategy ambiguous_accepted_umi_noncollapsed_counterseries

Repeat --r1 / --r2 for multi-file input. Boolean flags use --flag/--no-flag convention. Run crispr-correct map --help for the full flag list.

Save/load — parquet + JSON (cross-language durable)

# Convert a pickle into a parquet directory (portable across Python/R/Julia)
crispr-correct save --in result.pickle --out-dir result/
# Inspect in pandas:
#   pd.read_parquet("result/counts_protospacer_match_surrogate_match_barcode_match.parquet")

# Reconstruct a pickle from the directory
crispr-correct load --in-dir result/ --out result.pickle

Save/load via Python:

import crispr_ambiguous_mapping as cam
cam.save(result, "result/")      # writes parquet + manifest.json + qc.json + count_input.json
result2 = cam.load("result/")    # reconstructs the result (without the inference dict)

The per-observation inference dict (observed_guide_reporter_umi_counts_inferred) is not round-tripped through parquet — pickle it if you need it.

Post-processing — alleles subcommand

crispr-correct alleles \
    --in result.pickle \
    --tier protospacer_match_surrogate_match_barcode_match \
    --ambiguity accepted --umi-strategy noncollapsed \
    --out alleles.parquet

Requires the source pickle to have been produced with --retain-inference-results; otherwise emits a clear error pointing at the flag.

---

Testing

This repository ships with a test suite under ../tests/ (one directory up from the package). Three layers of coverage:

• tests/test_sccrispr_cell_barcode.py — pytest regression against a real scCRISPR FASTQ pair with an auto-baselined golden pickle. Runs in ~75 s.
• tests/simulation/ — ground-truth simulated data where every read's source guide, edits, UMI, and recombination state are known; compares all 135 combinations of (parsing mode × tier × ambiguity × UMI-strategy) to expected counts. Runs in ~90 s.
• tests/benchmarks/ — real HbF sample data with bench_results.jsonl tracking wall time and peak RSS across optimization iterations.

Run everything:

cd CRISPR-Correct-Folder
ENV=/data/pinello/SHARED_SOFTWARE/envs/bfb12_envs/bb_crisprcorrect_LOCAL   # or your editable-install env
"$ENV/bin/pytest" tests/ -v
"$ENV/bin/jupyter" nbconvert --to notebook --execute --inplace tests/simulation/test_local.ipynb

---

Supporting documentation

One level up from the package source, in CRISPR-Correct-Folder/:

• ARCHITECTURE.md — module map, pipeline stages, data models, Hamming / ambiguity algorithm.
• USAGE.md — recipes (protospacer-only / full-triplet+UMI / single-cell with sample_barcode), parameter cheat sheet, common pitfalls.
• BRANCHES.md — per-branch status (master, multi-sample-support, perf/counter-series-fix, perf/roadmap-phase1, …) and merge path.
• EXAMPLES.md — pointers to the in-tree production projects that use the package.
• IMPROVEMENTS.md — roadmap of tactical + architectural improvements across correctness, memory, performance, usability, DX.