chelae 0.2.1

A fast, accurate, multi-threaded toolkit for trimming and filtering short-read FASTQ data, written in Rust. Its name is the plural of chela, the pincer-like claws of crustaceans: a nod to what it does to FASTQ reads.

Contents

• Overview
• Examples
• Performance
• Installing
• About Fulcrum Genomics

Every option, the input and output rules, and more examples are in docs/usage.md.

Overview

chelae trims and filters short-read FASTQ in one multi-threaded pass, and identifies the adapters in a library you know nothing about.

• Accurate paired-end adapter trimming. chelae finds each pair's insert from where R1 and R2 overlap, then checks the implied adapter against known adapter sequences. That catches adapters too short for sequence matching alone to find, without mistaking adapter-like sequence inside the insert for adapter.
• Fast. SIMD kernels and a pipeline built for many cores trimmed about 1.7 M read pairs per second on 8 cores in our benchmark.
• Benchmarked. Against six other trimmers, chelae was the fastest, 1.25× faster than the runner-up and 2.6–5.7× faster than cutadapt, trim-galore-rs and fastp, and the most accurate on 8 of 11 simulated datasets. See Performance.
• One pass does it all: poly-G, adapter, read-structure hard-trimming with UMI extraction (which also trims the mate's UMI from reads that run through a short insert), poly-X and quality trimming, then length, N-base and quality filters.
• Adapter detection. chelae detect reports the adapters in a library and writes them as FASTA for chelae trim --adapter-fasta.
• Fits into pipelines. Split or interleaved paired-end files, stdin and stdout, gzip or BGZF input detected automatically, and a fastp-compatible JSON report for MultiQC.

Every option, the input and output rules, and more examples are in docs/usage.md.

Examples

Paired-end reads can be trimmed without naming any adapters. chelae finds adapters from where R1 and R2 overlap, and checks each one against every adapter in its built-in database (TruSeq, Nextera, small RNA, AVITI and MGI/DNBSEQ):

chelae trim \
    -i sample.r1.fq.gz sample.r2.fq.gz \
    -o trimmed.r1.fq.gz trimmed.r2.fq.gz

When you know the kit, name it - specify the right kit will make chelae a little faster and a little more accurate (since it can't match to the wrong kits) The following example also moves an 8 bp UMI from the start of R1 into the read name, skips the next 4 bases, and quality-trims 3' ends:

chelae trim \
    -i sample.r1.fq.gz sample.r2.fq.gz \
    -o trimmed.r1.fq.gz trimmed.r2.fq.gz \
    --kit truseq \
    --read-structures 8M4S+T +T \
    --quality-trim-3p 8:20

When a pair's insert is short enough that R2 reads through to the far end of the molecule, R2 ends in the reverse complement of R1's UMI and skip bases, so chelae trims those 12 bases from R2 as well. See Read-structures on paired-end reads.

-i and -o default to stdin and stdout, and interleaved paired-end input is recognized from its read names, so chelae can sit in a pipeline with no intermediate files:

samtools fastq sample.unmapped.bam \
    | chelae trim --kit truseq \
    | bwa mem -p ref.fa - \
    | samtools sort -o sample.bam

For a library of unknown provenance, chelae detect finds the adapters, and its FASTA feeds straight back into chelae trim:

chelae detect -i sample.r1.fq.gz sample.r2.fq.gz -o adapters.fa
chelae trim \
    -i sample.r1.fq.gz sample.r2.fq.gz \
    -o trimmed.r1.fq.gz trimmed.r2.fq.gz \
    --adapter-fasta adapters.fa

For single-end reads, chelae detect reports which built-in kit matches best:

chelae detect -i sample.fq.gz

Performance

chelae was benchmarked against six other FASTQ trimmers on simulated short-read libraries, for runtime and for adapter-trim accuracy against the simulator's ground truth. RESULTS.md has the method, the dataset definitions and every tool's numbers.

Runtime

Wall seconds for 50 M read pairs at 8 threads, median of 3 runs (WGS, cfDNA):

• Both libraries are simulated 2×150 paired-end reads with TruSeq adapters, trimmed with a realistic configuration: adapter, poly-G, sliding-window quality and length trimming, and an N filter.
• The host was an EC2 r8a.2xlarge (AMD EPYC 9R45, Zen 5, 8 cores). Inputs and outputs were on a RAM disk so the timings measure the tools rather than storage, and every tool with a compression option wrote gzip at level 1.
• The three runs of each tool agreed to within 6%.
• bbduk and trimmomatic weren't timed at this scale: on the paired-end accuracy datasets below, they took 8–16× and 4–27× as long as chelae (see broad toolset timings).

Accuracy

Eight paired-end and three single-end datasets of 1–4 M reads each simulate common library types: WGS at several insert sizes, cfDNA, exome with Nextera adapters, miRNA, and a high error rate (definitions). Accuracy is the RMSE of each read's trim point against the truth, in bases (lower is better).

ar = adapterremoval, tg-rs = trim-galore-rs, tmatic = trimmomatic. Each ID links to that dataset's full results.

Dataset 8 is an exome library with Nextera adapters and dataset 11 is miRNA; the rest are WGS or cfDNA with TruSeq adapters. On dataset 11 the top six tools differ by at most 5 of 4.2 M reads.

Versions: chelae 0.2.0, adapterremoval 3.0.2, bbduk 40.02, cutadapt 5.2, fastp 1.3.7, trim-galore-rs 2.3.0 and trimmomatic 0.41, the latest on bioconda as of 2026-09-25.

Installing

From bioconda

Using pixi, after adding the bioconda channel:

pixi add chelae

Or using your favorite conda client (conda, mamba, micromamba, …):

conda install -c bioconda chelae

With cargo

With Rust 1.89 or newer installed:

cargo install chelae

To build from source, see CONTRIBUTING.md.

About Fulcrum Genomics

Visit us at Fulcrum Genomics to learn more about how we can power your Bioinformatics with chelae and beyond.