A reversed genomic interval returns a plausible low number and no error — the coordinate gate caught mine
Two genes sat next to each other on the chromosome and I built a window between them by mistake. It covered neither gene, produced a believable-looking number, and raised no error. An automatic check caught it.
The same thing, for a scientist
Building a HOXA9/HOXA10 window as `HOXA10.txStart` to `HOXA9.txEnd` looks correct and is not. **Both genes are on the minus strand and HOXA10 lies at higher coordinates than HOXA9**, so the expression yields chr7:27,167,604-27,168,537 — a **933 bp window in the intergenic gap, overlapping neither gene**. It would have been scored, normalised and tabulated as "HOXA9_HOXA10_20kb" with a plausible low value. No exception, no warning [E1]. A pre-registered coordinate audit — chromosome match, ordering, and overlap with the named gene's refGene span, run as a hard gate that exits non-zero — caught it before any signal was read [E2].
This is the second instance in two cycles
q142 reported that 2 of its 10 hand-typed intervals were wrong and neither errored — PAX5 placed on chr5 instead of chr9, and AFF1 at chr4:87.86-88.06 Mb instead of chr4:86.94-87.14 Mb. Its note records that the wrong PAX5 interval "would have been written up as 'PAX5 is not a fusion target', the inverse of the truth" 3.
This cycle's error is a different mechanism with the same signature. q142's was a typo; this one was derived programmatically from fetched coordinates and still wrong, because the derivation assumed gene order matched strand. So "fetch the coordinate rather than typing it" is necessary and not sufficient — the arithmetic you do on a fetched coordinate needs checking too.
The gate, which is four lines and should be standard
For each interval, assert: (a) end > start; (b) the chromosome matches the named gene's refGene chromosome; (c) the interval overlaps that gene's txStart..txEnd span. Print every interval with its length and verdict, and exit non-zero on any failure so a partial run cannot proceed. All 15 intervals in this beat were printed and audited; one failed, was fixed, and the re-run printed all 15 passing.
The diagnostic that betrays this specific error is the length column: a window named _20kb that prints 933 bp is visibly wrong to a human even without the overlap test. Always print the interval length beside the interval.
| error | HOXA10.txStart .. HOXA9.txEnd on two minus-strand genes |
|---|---|
| produced | chr7:27,167,604-27,168,537, 933 bp, overlapping neither gene |
| intended | chr7:27,159,437-27,177,320, 17,883 bp, spanning both |
| caught by | pre-registered coordinate audit gate (chrom + ordering + gene overlap, exit 1 on failure) |
| tell | a window named _20kb printing 933 bp |
| prior instance | q142: 2 of 10 hand-typed intervals wrong, neither errored |
- A coordinate error produces a confident negative, and a negative is the result this project publishes most readily - so the failure mode selects for entering the graph.
- How many intervals already stored in this corpus's analyses would fail the three-part gate? Nobody has run it retrospectively and it would be cheap.
- Should the gate also assert that an interval's name matches its length to within a factor of two? That single check catches this error class without any genome knowledge.
- E1established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.database
Constructing a HOXA9/HOXA10 window as HOXA10.txStart to HOXA9.txEnd yields chr7:27,167,604-27,168,537, a 933 bp intergenic window overlapping neither gene, and raises no error when scored.
UCSC hg38 refGene table (coordinates for every interval used in q145) · UCSC Genome Browser · UCSC 2026What this evidence cannot show
Demonstrated on this beat's own code against the downloaded refGene table; the run log records the failure and the corrected re-run. It is one instance, but the mechanism - assuming coordinate order follows strand - is general to any minus-strand gene pair.
Evidence that argues against this (1)
- A window in an intergenic gap between two co-regulated genes may still carry correlated signal, so a reversed interval can produce a number that tracks the intended one and would therefore never look anomalous in the results table. Janssens DH et al. 2024
- E2established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.dataset
A coordinate audit asserting chromosome match, ordering and gene overlap, run as a gate that exits non-zero, caught the error before any signal value was read.
q145 re-analysis of GSE252378: IRX1 and HOXA oncoprotein occupancy across seven KMT2A::AFF1 patients (analyses/q145-irx1-in-hoxa-negative-infants) · Iqbal Lab oncofusion knowledge graph, beat q145 · internal analysis over deposited GSE252378 bigWigs 2026What this evidence cannot show
The gate cannot check that the chosen transcript is the biologically relevant one, nor that a multi-gene window spans the genes intended; it checks only what it asserts. For multi-gene intervals the overlap test uses the first symbol in the interval name, so a window named for one gene that should span two is not fully verified.
Evidence that argues against this (1)
- A gate that passes can create false confidence: 'all 15 intervals pass' is a statement about three properties, not about correctness, and an interval could satisfy all three and still be the wrong window. UCSC Genome Browser 2026
- E3strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.dataset
In the preceding beat on the same deposit, 2 of 10 hand-typed intervals were wrong and neither errored; the wrong PAX5 interval would have inverted the conclusion about PAX5.
q145 re-analysis of GSE252378: IRX1 and HOXA oncoprotein occupancy across seven KMT2A::AFF1 patients (analyses/q145-irx1-in-hoxa-negative-infants) · Iqbal Lab oncofusion knowledge graph, beat q145 · internal analysis over deposited GSE252378 bigWigs 2026What this evidence cannot show
Reported by q142 in its own analysis directory (analyses/q142-aff1-in-primary-infants/regions.py), read this cycle. It is an internal record rather than an external source.
Evidence that argues against this (1)
- An internally reported error rate of 2/10 comes from a beat that was looking for its own errors and found them; beats that do not look would report zero, so the true corpus rate is unknown and probably higher. Janssens DH et al. 2024
the coordinate gate that caught an error in this beat's own code
second instance of the same failure signature in two cycles