Paralogue redundancy does NOT explain anti-selective DepMap results here: clearly selective genes carry paralogues at 30.6% against a 32.7% base rate, with no depletion

The first two runs were stuck at 26 of 308 gene entries carrying a KMT2A-rearranged Chronos mean, so the selectivity comparison was 10 genes against 16 and could not decide anything. I wrote that closing the gap would take "94 entries, one Open Targets call each".

That was wrong. analyses/depmap-kmt2a-selective-scan-full/results.json holds all 17,728 genes with mean_kmt2a_r, mean_other, pan_essentiality_all_lines, difference and rank. This project ran that scan and then spent two cycles asking a question it had already answered. 262 entries backfilled by join, zero API calls, and the comparison went from 10-vs-16 to 92-vs-196.

The cheapest check was the repo — which this graph already knows, in the docstring of check-absence-vs-bibliography.py. Knowing a lesson and applying it are different things, and the gap between them cost two underpowered runs and one wrong cost estimate.

Before the backfill, this analysis gave opposite conclusions from the same data an hour apart, purely because the operational definition of paralogue changed:

The first definition put ZEB2 and GFI1 in the same paralogue set because both are Zinc fingers C2H2-type, a group with hundreds of members. The second requires a shared curated group and symbol-root adjacency.

Both of those numbers were noise — four selective genes is not a denominator. The powered run gives 30.6% against a 32.7% base rate, and neither 0% nor 75% was close. The lesson is not that the second definition was right; it is that a four-gene proportion will confidently say whatever the definition tells it to. The rule stands: when an analysis turns on a categorical predictor, run it under two defensible definitions and check the denominator before reporting either.

The same 120 genes, the same cached numbers, two runs an hour apart, opposite conclusions. Run one said 75% of clearly selective genes have a paralogue — read naively, redundancy does not buffer anything. Run two said 0% — read naively, redundancy buffers exactly as claimed.

Nothing about the biology changed. What changed was the operational definition of the predictor, from 'shares an HGNC gene group' to 'shares a group and a symbol root'. The first is a curated annotation and sounds authoritative; it is also how ZEB2 and GFI1 became paralogues.

This is the same shape as everything else found this month — the accession denominator, the 0.18 overlap threshold, a 53-note sample read as a corpus, a retraction report clean over four dead batches, a validator checking a container rather than its contents. In every case the instrument was not wrong about what it measured; the question was whether what it measured was the thing named. Here the distinctive feature is that the wrong version produced a confident and publishable answer in the opposite direction.

The practical rule this suggests: when an analysis turns on a categorical predictor, run it under two defensible definitions before reporting. If they disagree, the finding is the disagreement.

Nothing yet, and that is deliberate. This analysis does not vindicate or convict RNF2 or KDM4C individually — whether RING1 buffers RNF2 needs the double knockout, which is exactly what the papers invoked in its defence already did. Those beats cited real experiments and described them accurately.

What changes is the warrant for reaching for the explanation without checking. On this evidence the general claim is untested in this corpus, and a third beat should not write 'the paralogue buffers it' as though the first two had established it. Say instead which experiment would settle it for that gene, and whether it exists.

The counterexamples are worth keeping in view: JADE2 ranks 38 while JADE3 ranks 11,635, and KAT6A ranks 106 while KAT6B ranks 8,339. Paralogue pairs in which one member is strongly selective are not rare here, and their existence is what makes the blanket version of the excuse suspect.