length-bias

When length alone separates the classes — and the only example with a continuous label.

Human mRNA translation efficiency: 1000 transcripts from mRNABench, running from 425 to 17,497 nucleotides with a median of 3463. The label is a continuous value, not a class.

This example carries the README's first claim — your negatives are shorter than your positives, and a length classifier now beats your model — and two options nothing else exercises.

Run it

gb-qc evaluate-classes \
  --input examples/length-bias/data/train.csv \
  --label-column target \
  --regression \
  --end-position 500 \
  --out-folder qc-out

Two options here are doing real work:

• --regression splits the continuous target column at its median into high and low classes. That is why the report directory is class/sequence/high_vs_low rather than 0_vs_1.
• --end-position 500 stops the per-position checks at position 500. These are whole mRNAs; left unbounded the checks would run for thousands of positions and produce a report too large to open, for positions only the longest few transcripts reach. This is what that option is for.

What it produces

class/sequence/high_vs_low — open the report

Check	Flag	AU-ROC
Unique bases	Pass	—
Sequence Duplications within Labels	Pass	—
Duplicate Sequences between Labels	Pass	—
Sequence lengths	Warning	0.631
Per sequence GC content	Pass	0.582
Per sequence nucleotide content	Warning	0.617
Per sequence dinucleotide content	Warning	0.618
Per position nucleotide content	Pass	0.552
Per position reversed nucleotide content	Pass	0.548

What you should conclude

Length carries the label. Sequence lengths scores AU-ROC 0.631 — a classifier told nothing but how long a transcript is predicts which side of the median its translation efficiency falls on, 63% of the time.

For this dataset that is biology: transcript length genuinely relates to translation efficiency. But it is also the single easiest thing for a model to learn, and a sequence model scoring in the mid-0.60s here may have learned nothing but length. If you are benchmarking on this, that is the number to beat, and it is worth reporting alongside your model's.

The composition warnings are downstream of it. Base composition 0.617 and dinucleotide frequencies 0.618 are not separate findings: longer transcripts have systematically different composition (GC content varies with length in mRNAs), so a length difference shows up as a composition difference. The flagged features are T, TT, CA and GT — consistent with a length-linked shift rather than a specific motif.

Per-position is clean (0.552) even though composition is not. That fits: the difference is in overall composition, not in any particular position, which is exactly the pattern to expect from a length-driven bias.

It is a Warning, not a Fail, and 0.631 is as high as the length check goes anywhere in these examples — every other one passes it. That is worth knowing when you read the README leading with length bias: it leads with it because it is the easiest bias to introduce by accident and the cheapest to rule out, not because it is the most common.