seqseeker 0.1.0

Fast k-mer–based detection of any target sequence in paired single-cell FASTQ files

seqseeker

Fast k-mer–based detection of any target sequence in paired single-cell FASTQ files

seqseeker finds reads containing a user-specified nucleotide sequence (GFP, mCherry, Cre, any transgene) directly in raw gzipped FASTQ files from 10x Chromium experiments, attributes each hit to a cell barcode + UMI, and writes output tables ready for Seurat.

No custom genome build. No Cell Ranger re-run. Per-sample runtimes of 5–15 minutes vs. hours for alignment.

---

Table of contents

• Installation
• Quick start
• CLI reference
• Python API
• R interface
• Bash wrapper
• Built-in sequences
• Output files
• Loading into Seurat
• Why not just build a custom genome?
• How it works
• Tuning guide
• Citation

---

Installation

Requirements: Python ≥ 3.10, gcc or cc (for compiling the C core), zlib.

pip install seqseeker

On first use seqseeker automatically compiles its C backend. You can also pre-compile manually:

seqseeker compile

macOS

# Install Xcode Command Line Tools if you don't have gcc:
xcode-select --install
pip install seqseeker

Linux

sudo apt-get install build-essential zlib1g-dev   # Debian/Ubuntu
# or
sudo yum install gcc zlib-devel                    # RHEL/CentOS
pip install seqseeker

R package

# Requires the Python package to be installed first (above).
# Install from GitHub:
if (!requireNamespace("remotes")) install.packages("remotes")
remotes::install_github("ethanfenton/seqseeker", subdir = "R/seqseeker")

---

Quick start

Python / CLI

# Search for eGFP using a built-in sequence
seqseeker search \
  --sequence-name egfp \
  --input /data/novogene/01.RawData \
  --output results/gfp

# Any gene — pass a FASTA file
seqseeker search \
  --sequence-file my_transgene.fasta \
  --input /data/fastq/ \
  --output results/transgene

# Raw sequence string
seqseeker search \
  --sequence ATGGTGAGCAAGGGCGAGGAG... \
  --input /data/fastq/ \
  --output results/custom

# Only the 3' end (faster, specific)
seqseeker search \
  --sequence-name egfp \
  --region 3prime --region-len 200 \
  --input /data/fastq/ \
  --output results/gfp_3p

# Parallel processing
seqseeker search \
  --sequence-name egfp \
  --input /data/fastq/ \
  --workers 0            # 0 = auto-detect CPUs

Python API

import seqseeker

results = seqseeker.search(
    input_dirs=["/data/novogene/01.RawData",
                "/data/novogene2/01.RawData"],
    sequence_name="egfp",
    output_dir="results/gfp",
    workers=8,
)

print(f"Total hits:     {results['n_hits']:,}")
print(f"Unique barcodes:{results['n_unique_barcodes']:,}")

R

library(seqseeker)

results <- seqseeker(
  input         = c("/data/novogene/01.RawData",
                    "/data/novogene2/01.RawData"),
  sequence_name = "egfp",
  output        = "results/gfp",
  workers       = 8L
)

# Add to Seurat
seurat_obj <- seqseeker_add_to_seurat(seurat_obj, results)
FeaturePlot(seurat_obj, features = "seq_hits")

---

CLI reference

seqseeker <command> [options]

Commands:
  search            Search FASTQ files for a target sequence
  compile           Pre-compile the C search binary
  list-sequences    Show built-in sequences

seqseeker search options:

  Input/output:
    --input/-i DIR [DIR ...]    FASTQ directories (required)
    --output/-o DIR             Output directory [default: seqseeker_output]

  Target sequence (exactly one required):
    --sequence/-q ACGT...       Raw nucleotide string
    --sequence-file/-f FILE     FASTA file
    --sequence-name/-n NAME     Built-in name (egfp, mcherry, venus, tdtomato)

  Region selection:
    --region/-R full|5prime|3prime   Part of target to index [default: full]
    --region-len/-L N               Length for 5'/3' slice [default: 200]

  k-mer tuning:
    --kmer-size/-k K            k-mer length [default: 20, range: 8–31]
    --min-kmers/-m N            Min hits to call read positive [default: 1]

  10x read layout:
    --barcode-len/-b N          Barcode length in R1 [default: 16]
    --umi-len/-u N              UMI length in R1 [default: 12]

  Parallelism:
    --workers/-w N              Parallel processes [default: 1, 0=auto]

  Filtering:
    --sample-filter SAMPLE ...  Process only these sample IDs
    --report-interval N         Progress every N reads [default: 1000000]

  Advanced:
    --binary PATH               Path to pre-compiled kmer_search binary

---

Python API

seqseeker.search(
    input_dirs,            # str | list[str | Path]
    output_dir = "seqseeker_output",
    *,
    sequence      = None,  # raw nucleotide string
    sequence_file = None,  # FASTA file path
    sequence_name = None,  # built-in name
    region        = "full",     # "full" | "5prime" | "3prime"
    region_len    = 200,
    kmer_size     = 20,
    barcode_len   = 16,
    umi_len       = 12,
    min_kmers     = 1,
    workers       = 1,          # 0 = auto
    sample_filter = None,
    report_interval = 1_000_000,
    binary_path   = None,
) -> dict

Returns a dict with keys: output_dir, raw_csv, per_cell_csv, positions_csv, n_hits, n_unique_barcodes.

---

R interface

# Search
results <- seqseeker(input, output, sequence_name, ...)

# Load existing results
results <- seqseeker_load_results("results/gfp")

# Add to Seurat
seurat_obj <- seqseeker_add_to_seurat(seurat_obj, results,
                                      barcode_suffix = "-1",
                                      meta_col       = "gfp_reads")

# List built-ins
seqseeker_list_sequences()

---

Bash wrapper

The bash/seqseeker script is a thin wrapper that delegates to the Python package. After pip install seqseeker the seqseeker entry point is already on your PATH. The bash script is useful in cluster environments where you want an explicit shell command without activating a Python environment:

# Set a custom Python path
SEQSEEKER_PYTHON=/path/to/env/bin/python3 seqseeker search ...

---

Built-in sequences

Name           Length  Description
--------------------------------------------------------------
egfp              720 bp  Enhanced GFP (canonical, non-codon-optimised)
venus             720 bp  Venus YFP (T203Y/S65G/V68L/S72A mutations)
mcherry           711 bp  mCherry red fluorescent protein
tdtomato         1386 bp  tdTomato (tandem dimer Tomato)

Run seqseeker list-sequences to see the current list.

For a codon-optimised variant of eGFP (common in AAV constructs), use --sequence-file pointing to your construct FASTA rather than the built-in.

---

Output files

All outputs land in --output (default: seqseeker_output/):

File	Description
hits_raw.csv	One row per positive read; sample, lane, barcode, UMI, hit count, position, strand, read name, R2 sequence
hits_per_cell.csv	One row per unique barcode; n_reads, n_umi, samples seen
kmer_positions.csv	Histogram of first k-mer match position along the target (useful for diagnosing partial integrations)
_per_pair/	Intermediate per-FASTQ-pair CSVs (safe to delete after aggregation)

---

Loading into Seurat

# Method 1: via seqseeker R package
library(seqseeker)
results    <- seqseeker_load_results("results/gfp")
seurat_obj <- seqseeker_add_to_seurat(seurat_obj, results)

# Method 2: base R
hits         <- read.csv("results/gfp/hits_per_cell.csv")
hits$bc      <- paste0(hits$barcode, "-1")
seurat_obj$gfp_reads <- hits$n_reads[match(colnames(seurat_obj), hits$bc)]
seurat_obj$gfp_reads[is.na(seurat_obj$gfp_reads)] <- 0

---

Why not just build a custom genome?

See docs/why_kmer_search.md for the full argument. In short:

1. ~100× faster — no STAR index build, no full alignment pass.
2. Higher sensitivity — searches all raw reads, not just those that survive MAPQ filtering; catches partial integrations and junction reads.
3. No false negatives from multi-mapping — transgene k-mers that share similarity with endogenous loci are still detected if any portion is unique.
4. No genome required — works from the sequence string alone.
5. Direct barcode attribution — output is immediately Seurat-ready.

---

How it works

1. The target sequence (and its reverse complement) is encoded into a compact open-addressing hash table of 2-bit k-mers at startup (~1 ms).
2. R2 reads are scanned with a rolling window; any k-mer hit increments a counter. Once --min-kmers threshold is met the read is emitted.
3. The cell barcode and UMI are extracted from the corresponding R1 read (first barcode_len + next umi_len bases).
4. Results are aggregated per barcode across all FASTQ pairs.

The C core processes ~50 M reads/min/core on a modern server; total time for a 200 M read 10x library is typically <5 min on 4 cores.

---

Tuning guide

Goal	Recommended settings
Initial survey (maximum sensitivity)	k=20, min=1, region=full
Reduce noise after confirming signal	k=25, min=3
Fast specific detection of 3' end	region=3prime, region-len=200
Partial / truncated integration	k=15, min=1, region=full
Codon-optimised variant	--sequence-file construct.fasta

Inspect kmer_positions.csv to see where hits cluster along the target — a peak at position 600/720 (3' end of eGFP) is expected for normal expression; a flat distribution suggests unspecific noise.

---

Citation

If you use seqseeker in published work, please cite:

Fenton E (2026). seqseeker: fast k-mer detection of target sequences in single-cell FASTQ files. GitHub: https://github.com/ethanfenton/seqseeker

---

License

MIT © 2026 Ethan Fenton