ATF4's own transcript does NOT move under either menin inhibitor - and the whole ISR arm is reproduced by a DOT1L inhibitor, so it is not a menin-interface readout

analyses/does-atf4s-own-transcript-move-under-a-menin-inhibitor/panel-preregistration.json was written to disk first and contains the 72-symbol panel in nine named groups, the alias strings to be checked, the arms to be reported, and - importantly - the prespecified interpretation of each possible outcome:

if ATF4 mRNA falls by 1 log2 or more in both: transcriptional loss of ATF4, the simple story. if ATF4 mRNA is flat while targets collapse: loss of ATF4 ACTIVITY, which in the ISR is translational, so the graph's phrase describes an OUTPUT and not a cause. if the upstream kinases are also flat: narrows further, and a kinase's mRNA is a weak proxy for its activity, so this is a bound.

run.py imports that file and iterates its panel rather than defining its own list, so the panel in force is provably the panel declared. The second outcome is the one that occurred, and the third also occurred.

Four methodological choices, all made before the result and all recorded in the script's docstring.

1. Every probe, not the best one. The repo's earlier loader keeps one probe per symbol. That would have collapsed ATF5's four probes to one and hidden the four-probe agreement, and would have hidden that three of ATF3's and seven of CD84's probes are at background.
2. Alias resolution on the array annotation by ID SET, not count.
3. A detection floor printed beside every null, because "ATF4 does not move" is only a result if ATF4 is measured.
4. Both timepoints, both lines and both EPZ arms, which is what made Result 2 visible at all.

RS4;11, day 7, same experiment, same platform, same cells. Sorted by VTP-50469 effect 2:

Tally: 19 shared, 8 menin-selective, 7 differing by >=1 log2 in the same direction, 2 not moved, 1 DOT1L-selective.

The reading, stated at the strength the design supports. VTP-50469 and EPZ5676 hit different surfaces of the same complex - menin's pocket and DOT1L's SAM site - and they do discriminate: the HOX/MEIS/MEF2C arm separates them by 1.0 to 3.8 log2. So the assay has the resolution to detect a menin-specific effect, and on the ISR arm it detects none. Every ISR gene the graph used to infer the signature (CHAC1, TRIB3, ATF3, SESN2, STC2, plus ATF5 and DDIT3) is shared.

What this does NOT license. It does not show the ISR collapse is irrelevant, nor that it is a generic toxicity response - REH does not show it, and the specificity controls XBP1/ATF6/ERN1/HSPA5 do not move. The defensible conclusion is narrower and sharper: the ISR-output collapse marks disruption of the KMT2A-fusion transcriptional programme by EITHER route, and is a readout of fusion-programme loss rather than of the menin interface. That is a fusion-dependence claim, which is good for the lab's central hypothesis, and simultaneously a refutation of menin-interface specificity for this particular programme.

1. Four floor collisions. GDAP1L1, KCNK12, HOXA9 and PROM1 give a fold change exactly equal to minus their DMSO value, because the treated arm is zero counts. run.py flags these as floor_collision. GDAP1L1 and KCNK12 are two of the sixteen genes the prior entry lists as concordantly down, and for those two the RNA-seq half of the concordance is a floor artefact, not a measurement 3.

2. The ATF5 triple is SEM. The graph records "-3.02 / -2.89 / -2.77 on three MI-2-2 probes". Those are SEM; the same three probes in RS4;11 give -1.09 / -1.12 / -0.74 3. The prior entry says "in the earlier dataset" without naming a line and is not strictly false, but it sits in an RS4;11-scoped argument and the dispatch brief quoted it as RS4;11. What survives is better than the number: all four ATF5 probes move down in both t(4;11) lines and all four are flat-to-positive in REH, across a baseline range of 8.07 to 12.53 log2 RMA.

3. REH is not a universal negative control. The prior entry argues the 24 concordant genes are "expressed in the wild-type line, unresponsive in it", naming TSC22D3 (11.37), TRIB3 (10.31), SESN2 (9.99) and CHAC1 (8.95) - all four correct. But STC2 sits at 4.89 in REH, at background, so its flat response there is an absence of signal; and SOCS2 sits at 3.22 and 3.67 in SEM, i.e. one of the two t(4;11) lines does not express it at all. Three of the four array-measured cascade members also show a large RS4;11-versus-SEM discrepancy (CHAC1 -1.38 vs -0.38, ATF5 -1.09 vs -3.02, ATF3 -1.16 vs -0.36), which looks like a property of the GSE63664 experiment rather than of any one gene 3.

4. One irreconcilable platform disagreement. SGK1's array baseline is 8.94-10.89 log2 RMA and its RNA-seq baseline is 0.39-1.02 log2 CPM. Most likely a 3-prime-end window mismatch against a transcript variant. Both platforms agree on the direction and neither magnitude can be trusted 3.

Recorded rather than silently dropped, as the dispatch brief requires. All figures RS4;11 day 7 VTP-50469 unless stated 1, 2.

Declined because the measurement is a floor artefact: GDAP1L1 (-3.10 with DMSO exactly 3.10, treated arm zero counts) and KCNK12 (-1.08 with DMSO 1.08, same). Both are in the prior entry's concordant-down list and both have no usable RNA-seq measurement.

Declined because the gene is not detectably expressed where it matters: CD84 (7 of 8 array probes at background 2.93-5.39; the +4.18 rises from a 3.81 log2 CPM baseline), CASZ1 (4 of 5 array probes at background; RNA-seq baseline 1.26), RGS2 (RNA-seq baseline 1.08), SERPINE2 (2.76, and it moves +1.38 under EPZ5676 - opposite direction).

Declined because there is no ISR or fusion mechanism to attach: KIAA1598/SHTN1, KIAA0754, ZNF503, TTC28, EMID1, FAM69C/DIPK1C, FAM129A/NIBAN1, TSC22D3, RHEBL1, RGS16. These are real, reproducible movers with clean baselines and no published connection to the ISR, to ATF4 or to a KMT2A fusion; writing an entry per gene would have been padding. TSC22D3 deserves a note: it is the glucocorticoid-induced leucine zipper, dexamethasone is backbone therapy in this disease, and its -1.66 is worth an entry in a steroid beat rather than an ISR one.

Declined because the transcript is flat and the gene adds nothing beyond the two kinases written: EIF2AK1/HRI (max 0.26 anywhere), EIF2S1 (max 0.31), EIF2S2, EIF2S3, PPP1R15B, QRICH1 (max 0.23). EIF2AK2/PKR is declined for a different reason and it is the interesting decline: it is the ONE kinase that moves (-1.88 at RS4;11 day 7) and it sign-flips to +1.40 at day 2 in the same line and arm while sitting at +0.20 on a well-expressed array probe. A sign flip between adjacent timepoints on a low-baseline gene is noise, and an entry built on it would have been a false positive.

Declined as specificity controls that behaved: XBP1 (-1.31), ATF6 (+0.14), ERN1 (-0.38), HSPA5 (-0.23), EDEM1 (-0.72), DNAJB9 (-0.93). The IRE1 and ATF6 arms do not move while the ATF4 arm falls 1.4-5.1, which is the specificity Mungrue established for CHAC1 - their value is as controls inside gene-atf4, not as entries.

Declined but flagged as the strongest candidates for the NEXT beat: PHGDH (-3.24, and strongly menin-selective at -1.52 under EPZ5676) and PSAT1 (-2.40). These are the serine-biosynthesis genes the menin-ATF4 Ewing paper is built on; PHGDH is named 35 times in this corpus with no entry, PSAT1 zero times, and PHGDH is also one of the four genes in the AML prognostic signature whose top gene is SOCS2. A serine/one-carbon beat is the obvious successor to this one and would have been padding here. NUPR1 (-2.35) and CEBPB (-0.96, DOT1L-selective at -1.70) belong with them.