Stomata

GPU-accelerated exhaustive CRISPR off-target search.

Version: 0.10.0 · License: MIT · Requires: CUDA 11.8+, C++17, CMake 3.20+

What it does

Given a guide RNA (or a batch of them) and a reference genome, Stomata enumerates every position in the genome within a user-specified edit distance. Output is a TSV with per-hit coordinates, strand, CIGAR, PAM, CFD activity score, and seed/distal mismatch breakdown — or an aggregate per-spacer summary.

• Exhaustive. Every position is evaluated.
• PAM-agnostic search. Stomata finds all sequence matches first, then annotates PAM. You can re-filter without re-searching and support any IUPAC PAM (SpCas9 NGG, Cas12a TTTV, engineered variants).
• Indel-aware. Levenshtein search implemented through Myers' bit-parallel algorithm.
• GPU-resident genome. The .st index is memory-mapped, allowing instant search post-load.

Getting started

# 1. Build
conda create -n stomata python=3.10 -y && conda activate stomata
conda install -c conda-forge cmake compilers gxx=11 spdlog catch2 -y
cmake -B build -DCMAKE_BUILD_TYPE=Release
cmake --build build -j$(nproc)

# 2. Index a genome (one-time; typically 30 seconds to 5 minutes, hardware-dependent, for hg38)
./build/src/stomata --index-genome hg38.fa   # produces hg38.fa.st

# 3. Search
./build/src/stomata \
  --genome hg38.fa.st \
  --pattern GAGTCCGAGCAGAAGAAGAA \
  --threshold 3 \
  --pam NGG \
  --output hits.tsv

Batch mode: replace --pattern with --spacer-file guides.txt (one sequence per line, optional name<TAB>seq format, # comments allowed).

Use --spacer-summary summary.tsv for per-spacer aggregates (hit counts by distance, CRISPick aCFD promiscuity score, BED overlap counts).

Documentation

• User Guide — installation, full CLI, output formats, troubleshooting
• --help on the binary prints the full flag reference

Things that will surprise you

• Default threshold is 3. Pass --threshold N explicitly if you care.
• Default PAM filter is none. Without --pam NGG (or similar) the output includes every sequence match regardless of PAM context. This is intentional but it means a first run produces more rows than Cas-OFFinder does. Use --max-hits / --max-total-hits to cap, or --summary for aggregates only.
• GPU is automatic if present. Pass --cpu-only to force CPU. Pattern length must be 1–64 bp for GPU; longer patterns fall back to multi-word Myers on CPU.
• Hamming mode searches by Hamming distance but aligns by edit distance. So --distance-mode hamming output may still contain indels in the alignment column — the search found the position via mismatch counting, but the reported CIGAR is the optimal edit-distance alignment at that position. If you need strict mismatch-only output, post-filter rows whose cigar contains I or D, or use --no-compute-mismatches to skip alignment entirely.
• U→T normalization is on by default. RNA spacers work out of the box. Disable with --no-treat-u-as-t if you want strict validation.

Issues: https://github.com/simpsondl/stomata/issues