variable-length¶

Why most positions in a report go unscored.

Tomato long non-coding RNAs: 7274 training sequences from OmniGenBench's lncrna_s_lycopersicum, running from 105 to 5964 nucleotides with a median of 331. Shipped whole, so the flags are the published dataset's.

Every other example has fixed-length sequences, where every sequence reaches every position and the per-position checks are simple. This one is what happens when they do not.

Run it¶

gb-qc evaluate-classes \
  --input examples/variable-length/data/train.csv \
  --out-folder qc-out

gb-qc evaluate-splits \
  --train-input examples/variable-length/data/train.csv \
  --test-input examples/variable-length/data/test.csv \
  --sequence-column sequence \
  --out-folder qc-out

What it produces¶

class/sequence/0_vs_1 — open the report

Check	Flag	AU-ROC
Unique bases	Fail	—
Sequence Duplications within Labels	Warning	—
Duplicate Sequences between Labels	Fail	—
Sequence lengths	Pass	0.538
Per sequence GC content	Pass	0.508
Per sequence nucleotide content	Pass	0.556
Per sequence dinucleotide content	Pass	0.558
Per position nucleotide content	Pass	0.559
Per position reversed nucleotide content	Pass	0.571

split/sequence/train_vs_test — open the report

Check	Flag	Leakage
Data Leakage	Warning	0.80% of queries, 0.12% of targets

What you should conclude¶

Most of this report is `Unknown`, and that is correct¶

Of the 26,360 per-position sub-checks in the class report, 20,870 are Unknown — 79% of them. Four different boundaries are in play, and seeing them in one report is the point of this example:

Position	What happens there
549	As far as 25% of each class still reaches. This is the default `--min-coverage` window: the furthest a position can be flagged, and the last position the figures draw
1358	As far as 250 sequences of each class reach. This is the floor `--min-coverage 0` leaves, and it cannot be switched off
2636	As far as 50 sequences of each class reach. The furthest any position is reported on — every one past 549 as `Unknown`
5964	The longest sequence in the dataset

Past 549, positions are reported Unknown. Not Pass — the comparison was not made.

The reason is not sample size. It is that a cohort far along the sequence is not a sample of the class; it is a sample of the class's longest sequences. If length correlates with composition — and in lncRNAs it does — then a difference found at position 3000 is a difference between long sequences, which is a different claim from a difference between classes. No amount of data fixes that. Declining to score is the only honest answer.

If you need to look further out, --min-coverage 0 drops the fraction rule and leaves only the 250-sequence floor, which extends flagging to position 1358. Read what comes back with the caveat above firmly in mind.

The vocabulary check fails, and for a sharper reason than it looks¶

Fail on Unique bases. The check does not flag non-ACGT characters as such — a dataset where both classes contain N passes. What it flags is an asymmetry, and that is what is here:

Class	Alphabet
`0`	A, C, G, T
`1`	A, C, G, N, T

Six sequences out of 3,607 carry an N that the other class never contains. Which means any sequence containing N is perfectly classifiable, by a rule a model will find long before it finds any biology. Six sequences will not move an AU-ROC, but the asymmetry is the tell: it almost always means the two classes came through different pipelines, and that is worth knowing about a dataset before you train on it.

It also matters mundanely — your tokeniser has to do something with N, and whatever it does silently is a modelling decision you did not make on purpose.

Sequences repeat, within and between classes¶

Fail on Duplicate Sequences between Labels and Warning within labels. Identical sequences carrying opposite labels are unlearnable by construction and cap achievable accuracy.

Composition is clean¶

Everything statistical passes: GC 0.508, composition 0.556, dinucleotides 0.558, per-position 0.559. So this is a dataset whose content looks fine and whose bookkeeping does not — duplicates and non-ACGT characters. A useful shape to recognise, because it is fixable without touching the biology.

Leakage¶

Warning at 0.80% of queries, 0.12% of targets. Small, real, worth a look at the alignment panel to see what the shared sequence actually is.