DirectClean

Artifact-aware preprocessing and read rescue for Oxford Nanopore direct-cDNA sequencing.

DirectClean takes raw ONT direct-cDNA FASTQ and produces clean, 5′→3′ oriented reads for transcript quantification, isoform analysis, and fusion calling.

directclean -i raw_reads.fastq -r genome.fa -o results/ -t 8 -j gencode.v41.bed12

Why DirectClean?

Direct-cDNA libraries carry artifact classes that primer-based preprocessing does not resolve:

• Foldback inversions — the sequenced strand folds back on itself.
• Internal adapter concatemers — two molecules joined at a TSO/RTP junction.
• Homopolymer-mediated RT template switching — reverse transcriptase detaches at an A/T-rich region and re-primes on an unrelated transcript. These read like genuine gene fusions at the alignment level.
• Unclassified reads — reads that cannot be oriented from their termini, but whose internal sequence is still recoverable.

The usual response is to discard the whole read. In PCR-free protocols every read is a unique molecule, so that is expensive. DirectClean instead splits or trims at the artifact junction and keeps the usable sequence.

Pychopper handles orientation and terminal-primer-based rescue well. DirectClean keeps that logic and extends coverage to direct-cDNA-specific artifact classes:

Pipeline

flowchart LR
    A(["Raw FASTQ"]) --> S1["<b>Stage 1</b><br/>Foldback removal"]
    S1 --> S2["<b>Stage 2</b><br/>Orientation"]
    S2 -->|oriented| S4["<b>Stage 4</b><br/>Adapter rescue"]
    S2 -->|unknowns| S3["<b>Stage 3</b><br/>Unknowns rescue"]
    S3 --> AL{{"minimap2<br/>splice-aware"}}
    S4 --> AL
    AL --> S5["<b>Stage 5</b><br/>Homopolymer rescue"]
    S5 --> OUT(["Cleaned FASTQ"])

    classDef io fill:#e8e8e8,stroke:#5a5a5a,stroke-width:2px,color:#1a1a1a
    classDef base fill:#cfe2f3,stroke:#2e6da4,stroke-width:2px,color:#1a1a1a
    classDef unk fill:#fce5cd,stroke:#c47d1a,stroke-width:2px,color:#1a1a1a
    classDef adp fill:#d9ead3,stroke:#3d8b52,stroke-width:2px,color:#1a1a1a
    classDef hom fill:#e6d5f2,stroke:#7b4fa3,stroke-width:2px,color:#1a1a1a
    classDef algn fill:#fff2cc,stroke:#b8933a,stroke-width:2px,color:#1a1a1a

    class A,OUT io
    class S1,S2 base
    class S3 unk
    class S4 adp
    class S5 hom
    class AL algn

A read is only split when the junction is supported by sequence evidence; fragments below the minimum length are dropped. Stage 5 requires both A/T density ≥ 85% and a consecutive A/T run ≥ 5 bp within a 10 bp window, so non-A/T junctions, including real fusions, pass through untouched.

Performance

Read and base retention across four ONT direct-cDNA datasets:

On VCaP, where the additional reads come from:

Chimeric read rate on VCaP: 30.9% raw → 26.8% Pychopper → 16.9% DirectClean.

Installation

pip

pip install directclean

External tools (minimap2, samtools, breakinator, restrander) must be available on PATH.

Developer installation

git clone https://github.com/ylab-hi/DirectClean.git
cd DirectClean
mamba env create -f environment.yml
mamba activate directclean
pip install -e .

Verify with:

directclean --help

Usage

DirectClean runs a splice-aware genome alignment internally. Allow ~24 GiB peak RAM; 32 GiB is recommended.

directclean \
  -i raw_reads.fastq \
  -r genome.fa \
  -o results/ \
  -t 8 \
  -j gencode.v41.bed12

Detection thresholds and report options can be tuned as well, run directclean --help for the full list.

Output

Citation

To be updated on publication. Please also cite the integrated tools:

• Breakinator: Heinz JM, Meyerson M, Li H. Detecting foldback artifacts in long reads. BMC Genomics (2026).
• Restrander: Schuster J, Ritchie ME, Gouil Q. Restrander: rapid orientation and artefact removal for long-read cDNA data. NAR Genomics and Bioinformatics 5(4):lqad108 (2023).

Support

Bug reports and feature requests: open an issue.

License

MIT

Contact

Qingxiang Guo — qingxiang.guo@northwestern.edu · Rendong Yang Lab