[wdecoster]

Closes #37. Adds removal of poly(A) tails and poly(T) heads from RNA and cDNA reads. It is off by default: default output is unchanged, and the v0.14.1 digest test still passes.

Usage

--trim-poly-tails [<MIN_LENGTH>]

--trim-poly-tails removes a poly(T) head from the start and a poly(A) tail from the end of each read; U counts as T. Both ends are trimmed because cDNA reads come in both orientations. Runs shorter than MIN_LENGTH (default 10) are left alone. --trim-poly-tails 15 or --trim-poly-tails=15 sets another minimum. chopper has no positional arguments, so the optional value can't be confused with anything else: --trim-poly-tails -i x.fq uses the default.

How tails are found

The scoring is the same as cutadapt's --poly-a (poly_a_trim_index in qualtrim.pyx):

• Scan in from the read end: +1 for each A (or T at the start), −2 for anything else.
• The tail ends where the score peaks, provided at most 20% of it are other bases.

This tolerates sequencing errors inside the tail without running on into the transcript.

The one difference from cutadapt is the default minimum: 10 instead of 3. In simulation, a minimum of 3 trimmed about 3% of reads that have no tail, because they end in "AAA" by chance; 10 trims about 0.1% and loses nothing on real tails.

Anchored only: the tail must reach the end of the read, so adapters and primers have to be removed first, for example by the basecaller or dorado trim. A few leftover bases past a tail are removed with it, as long as the whole stays within the 20%.

How it fits

• It runs after --trim-approach and narrows only the outer ends of the kept segments: a polyT head off the first segment and a polyA tail off the last. Interior split points are not read ends and are left alone.
• Because it only narrows the (start, end) ranges, --update-mods, uBAM output and segment naming work without changes. A test checks that a modification call stays on its base.
• As with the trim approaches, filters see the untrimmed read and only --minlength is checked again afterwards.

Validation (simulated only)

ONT-like errors: half deletions, a quarter substitutions, a quarter insertions. 400 bp body, 10,000 reads per condition, run through the release binary:

error rate	tails removed	tail bases left / read	body bases lost / read	tailless reads trimmed
2%	100%	0.1	0.8	0.1%
5%	100%	0.3	0.8	0.1%
10%	100%	0.5–0.7	0.8	0%
15%	≥99.9%	0.9–1.5	0.7–0.8	0.1%

Body bases are lost when a body happens to start with T or end with A. Those bases can't be told apart from a tail, so no method avoids this.

Not yet validated on real reads: the data at hand isn't adapter-clipped.

Tests

make ci passes: 80 unit tests (8 new, including a seeded simulation at 2–15% errors) and 21 end-to-end tests (2 new).

🤖 Generated with Claude Code