The SEC-family target sets, reconstructed as a full three-way Venn: 27.5% of the family's transcriptional output is invisible to AFF4 depletion — and the two 'distinct' sibling sets overlap 8-20x more than chance

Stated in the paper 1:

Solved (one line of arithmetic, no model, no data download):

AFF2-only        = 0.553 x 226 = 125
AFF3-only        = 0.617 x 423 = 261
AFF2&AFF4-only   = 226 - 125 - 76 =  25
AFF3&AFF4-only   = 423 - 261 - 76 =  86
all three        = 150 -  25 -  86 =  39
AFF2&AFF3-only   =  76 -  39       =  37
AFF4-only        = 1116 - 150      = 966

The check that makes this trustworthy rather than plausible: the two unique fractions are stated as percentages and both land within 0.03 of an integer (0.553 x 226 = 124.978; 0.617 x 423 = 260.991), which is what you see when a percentage was computed from an integer count and rounded to one decimal. And all three marginals reproduce exactly — 125+37+25+39 = 226, 261+37+86+39 = 423, 966+25+86+39 = 1,116. A wrong solution would miss at least one.

One coincidence to not misread: the count of genes responsive to a sibling but not to AFF4 is 423, which is also the AFF3 DEG total. Those are two different sets that happen to be the same size.

Union of the three DEG sets = 125 + 261 + 966 + 37 + 25 + 86 + 39 = 1,539 genes.

The argument this is built for: ENL and AF9 are subunits of all three complexes 2, and the YEATS pocket is not the scaffold-binding surface — the AHD is (mechanism-kmt2a-mllt1-engages-ptefb-through-the-enl-ahd-494-559). So a YEATS-pocket ligand cannot ask which scaffold the reader it is occupying is sitting on. 27.5% is the measured size of the extra territory that scaffold-blindness puts in scope, in the only system where the three sets have ever been compared.

It is a floor and a ceiling at the same time, for opposite reasons, and both belong in the same paragraph. A floor, because AFF2 and AFF3 mRNA are about 10-fold lower than AFF4 in HEK293T 2, so their knockdowns are the underpowered arms and their true sets are probably larger. A ceiling, because a knockdown-response set is not a YEATS-dependent set: removing a scaffold removes P-TEFb delivery too, and nothing here says any of those 423 genes needs the reader.

Rule 84 says a proportion is not a result until its null is written down. The paper's headline is that the family members control different subsets. Against what?

Hypergeometric expectation for |AFF2 ∩ AFF3| given |AFF2| = 226, |AFF3| = 423 and N expressed genes tested:

The paper does not state its tested-gene denominator, so this is given as a sensitivity sweep rather than one number — and the conclusion does not depend on the choice: at no plausible denominator is 76 anywhere near chance. Observed Jaccard is 0.133 against a null between 0.006 and 0.015.

So both statements are true and the graph should hold both. 55-62% of each sibling's response set is its own (the paper's point), and the part that is shared is an order of magnitude more shared than independence predicts (the part nobody stated). The second does not refute the first; it says the three complexes are not independent instruments, which matters when you propose to hit a subunit they all contain.

And the obvious deflation has to be stated, because it is the strongest objection to this whole section: two knockdowns in the same cell line, called at FDR<0.05 on n=2 versus n=4, will share their most detectable genes whatever the biology. Highly expressed genes are easiest to call differential in every arm, so any two DEG lists from one experiment are enriched against a uniform hypergeometric null. The 8-20x is therefore an upper bound on shared targeting and a lower bound on nothing. The clean version of this test is a degree-matched or expression-matched permutation over the actual tested set, which requires GSE34097 and is not run here.

The underlying data are public: GSE34097, whose !Series_pubmed_id is 22547686, public since 2 May 2012 3. Everything above is a re-derivation from six printed numbers and inherits every property of the original DEG calls; none of it is a reanalysis.

The analysis that would replace it, in order of cost:

1. Recover the three gene lists. The seven-cell partition above names counts, not genes. Which 39 genes respond to all three scaffolds is the single most interesting unasked question in this dataset, and it is a table nobody has printed.
2. Expression-matched permutation for the AFF2/AFF3 overlap, replacing the uniform null above.
3. Intersect the 423 non-AFF4 genes with an imprinted-gene list and with the XIST locus, because protein-sec-l3-complex records AFF3's documented physiological job as allele-specific regulation at imprinted control regions. If the non-AFF4 territory is enriched for imprinted loci in a plain HEK293T knockdown, the imprinting liability of a reader-directed drug stops being speculative.

This entry is filed status: run for the arithmetic and planned for all three of those; the data.status field records the arithmetic only.