EIF2AK3 / PERK (Q9NZJ5) - the kinase whose transcript is flat under both inhibitors, whose arm IS functionally required in MLL-AF9 leukaemia stem cells, and whose common name collides with phospho-ERK

This is a prediction, not a measurement. No one has seen this protein in this shape. The colours say how much the software trusts itself, and that is worth reading before anything else — the orange stretches are places the model is effectively saying this part has no fixed shape.

• A PERK inhibitor RESCUES leukaemia stem cells in an MLL-AF9 model. A menin inhibitor lowers the output of the same PERK-eIF2alpha-ATF4 axis. If those two facts compose, menin inhibition is protecting the stem-cell compartment on this axis while killing the bulk population - which is exactly the shape of a relapse mechanism, and nothing in a bulk transcriptome or a MEN1 sequencing panel would reveal it.
• Any therapeutic attempt to ACTIVATE this arm (an ER-stress inducer, a chaperone inhibitor, a proteasome inhibitor) would be working against the menin inhibitor's measured effect on the same axis, and vice versa. The two strategies are candidates for mutual antagonism and neither has been tested against the other.
• In this project specifically, searching this gene by its common name imports papers about MEK inhibition in MLL-rearranged infant ALL, because pERK and PERK are one token. Any bounded absence, coverage count or citation audit in this graph that used PERK as a search string is contaminated by the most relevant-looking possible false positives.

1. Is phospho-eIF2alpha changed in RS4;11 under VTP-50469? Every mechanistic claim in this beat's ATF4 story turns on it, no menin-inhibitor study in any cell has measured it, and it is one lane on a blot.
2. Does menin inhibition protect the leukaemia-stem-cell compartment on this axis? A PERK inhibitor rescues LSC self-renewal in MLL-AF9 mice, a menin inhibitor lowers ATF4 output, and nobody has scored the stem-cell-enriched fraction separately from the bulk under a menin inhibitor.
3. Is DNAJC10 expressed and does it change in KMT2A::AFF1 cells? It was not in this beat's pre-registered panel, it is reportedly upregulated in AML and LSC-enriched cells, and it is the chaperone whose loss engages this arm.
4. Does GSK2606414 or a modern PERK inhibitor antagonise VTP-50469 in RS4;11? The prediction from Li 2024 is that it should, if the menin inhibitor is already acting on the same axis - and antagonism would be the clean demonstration that the ISR arm is doing something the drug does not want.

1. E1established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.dataset

   EIF2AK3's transcript does not move under either inhibitor in any line or at either timepoint: RS4;11 array MI-2-2 +0.32 (p 0.030, DMSO 8.98), SEM -0.13, REH +0.09, RNA-seq RS4;11 day 7 VTP-50469 -0.35 and EPZ5676 -0.09, day 2 +0.29, MOLM13 day 7 +0.08 - while in the same RS4;11 day-7 samples its downstream output falls 1.42 to 5.11 log2, and EIF2AK1, EIF2AK4 and EIF2S1 are likewise flat.

   Q106 pre-registered ISR panel run over the cached GSE63664 (MI-2-2 / EPZ004777, SEM / RS4;11 / REH, GPL570 RMA) and GSE127506+GSE127508 (VTP-50469 330 nM / EPZ5676 1 uM, RS4;11 and MOLM13, d2 and d7, 3-prime-end RNA-seq) matrices. Script analyses/does-atf4s-own-transcript-move-under-a-menin-inhibitor/run.py, results.json, panel frozen in panel-preregistration.json BEFORE any expression value was read. EIF2AK3 probes 218696_at and 241463_at; RNA-seq row EIF2AK3; the full eIF2alpha kinase module. · NCBI GEO, re-analysed here 2026
   ▸What this evidence cannot show

   THE BOUND IS TRANSCRIPTIONAL ONLY, and this is the entry's central caveat rather than a footnote: PERK is activated by oligomerisation and trans-autophosphorylation in response to ER client load, so its mRNA is not the regulated variable and a flat transcript is compatible with PERK ACTIVITY having risen, fallen or not moved. No claim about activity is made. The array probe is well expressed in all three lines (8.98-10.60 log2 RMA) so the null is a measurement and not a detection failure; the second probe (241463_at, 3.33) is at background and is reported as background. The +0.32 at p 0.030 is another nominally significant, biologically nil change. RS4;11 array DMSO arm n=2.

   ▸Evidence that argues against this (1)

   • Reading a flat kinase transcript as narrowing the mechanism assumes the ATF4-output collapse would have to be driven transcriptionally from the kinase layer if it were driven from there at all. It need not be: PERK could be inactivated by a change in ER client load, by a phosphatase, or by the fusion protein binding a feedback component - and this same beat found that KMT2A fusion proteins were reported in 1999 to bind GADD34, the eIF2alpha-phosphatase adaptor. The flat transcript is consistent with a fully reconfigured pathway. Adler HT et al. 1999
2. E2strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper

   DNAJC10, an HSP40-family member frequently upregulated in AML and in LSC-enriched cells, maintains leukaemia-stem-cell survival and self-renewal; its deficiency causes a dramatic increase in apoptosis in human leukaemia lines and LSC-enriched populations but is not required for normal haematopoiesis; Dnajc10 deficiency significantly abrogated AML development and suppressed LSC self-renewal in the MLL-AF9-induced murine leukaemia model; inhibition of DNAJC10 specifically induces ER stress and promotes activation of the PERK-eIF2alpha-ATF4 branch of the unfolded protein response; and blocking PERK with GSK2606414 or shRNA rescued the DNAJC10 loss-of-function phenotype both in vitro and in vivo.

   DNAJC10 maintains survival and self-renewal of leukemia stem cells through PERK branch of the unfolded protein response. · Li M, Wu X, Chen M, Hao S, et al. · Haematologica 2024
   PMID 37496439DOI
   ▸What this evidence cannot show

   MURINE MLL-AF9, myeloid, not the index B-lineage KMT2A::AFF1 (rule 73), plus human AML lines. ABSTRACT ONLY - the figures were not obtained, so the LSC assays, the PERK phosphorylation readouts and the GSK2606414 dosing are unexamined. The axis is engaged there by removing a chaperone, NOT by a menin inhibitor, and nothing in that paper or this one links the two perturbations. NOT PREVIOUSLY CITED IN THIS GRAPH (scripts/check-citation-reuse.py, 2026-09-22: 'NOT CITED HERE - genuinely new to this graph'). Reachable only under the PERK alias, not under EIF2AK3.

   ▸Evidence that argues against this (1)

   • This result points OPPOSITE to the menin-inhibitor observation: here ACTIVATION of PERK-eIF2alpha-ATF4 is what kills the leukaemia stem cell and blocking PERK rescues it, while a menin inhibitor LOWERS ATF4 output 1.4 to 5.1 log2. If both hold in the same cells, a menin inhibitor is doing on this axis what a PERK inhibitor does - which would protect the LSC compartment while killing the bulk population. That is an uncomfortable prediction and it is unfalsified because nobody has scored the stem-cell fraction separately under a menin inhibitor. NCBI GEO, re-analysed here 2026
3. E3established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.database

   PubMed returns 0 records for EIF2AK3 jointly with an eight-operand KMT2A-fusion list, with a non-empty querytranslation and 'No items found.', while PERK[tiab] against the same operands returns 5 including PMID 37496439; PERK[tiab] alone returns 12,317 records against EIF2AK3's 364, sharing 111 (Jaccard 0.0108); and one of the five fusion-scoped PERK hits is PMID 27588400, a MEK-inhibitor paper in MLL-rearranged infant ALL, which matches because 'pERK' (phospho-ERK1/2) and 'PERK' are the same string to a case-insensitive index.

   NCBI PubMed esearch, run 2026-09-22, with querytranslation and warninglist read back from the response for every query quoted · NCBI PubMed 2026
   ▸What this evidence cannot show

   All five PERK-scoped titles were fetched and read; none was assumed. Rule 3 was satisfied on the EIF2AK3 zero AND THE ZERO WAS STILL FALSE - the second instance of that in this beat after PPP1R15A/GADD34. Note the two halves of this trap pull opposite ways: you MUST widen to PERK to find the functional paper, and you MUST title-check every PERK hit because phospho-ERK contaminates exactly the papers most likely to look relevant in a leukaemia project. Rule 94 form; rule 95 observed; rule 80: titles and abstracts only. NOT in data/symbol-collisions.json.

   ▸Evidence that argues against this (1)

   • Calling PERK 'contaminated' is only half true and the distinction matters: unlike CHOP, where none of the twelve extra records was about the gene, MOST of PERK's 12,317 records ARE genuine PERK papers, because the field almost never writes EIF2AK3. So the canonical symbol is the deficient instrument here and the alias is the better one, subject to title-checking - the opposite prescription from the DDIT3/CHOP case in this same beat. NCBI PubMed 2026
4. E4strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.database

   UniProt Q9NZJ5 is the single reviewed human EIF2AK3 record, 1,116 aa with a protein-kinase domain at 593-1077, synonyms PEK and PERK, fifteen PDB entries, annotated as a metabolic-stress-sensing protein kinase that phosphorylates eIF2alpha in response to stresses including the unfolded protein response and as a key effector of the integrated stress response, localised to the endoplasmic reticulum membrane - a call carrying only ECO:0000250, by similarity to mouse Q9Z2B5, with a note that it is also present at mitochondria-ER contact sites; and DepMap 26.06 mean gene effect -0.0424 across 1,257 screens, 1.3% of lines below -0.5, leukaemia mean -0.0924.

   UniProtKB - Q9NZJ5, E2AK3_HUMAN, eukaryotic translation initiation factor 2-alpha kinase 3 / PERK · UniProt Knowledgebase 2026
   ▸What this evidence cannot show

   Resolved live 2026-09-22: exactly one reviewed record (rule 62 satisfied), re-fetched alone with name, length and primary symbol compared back: PASS. RULE 91 FLAG: the ER-membrane localisation for HUMAN PERK carries ONLY ECO:0000250 - by similarity to the mouse orthologue - so even the compartment of the textbook ER stress kinase is, in the human record, an inference rather than a human experimental observation. Recorded as uncoded/inferred rather than as a measurement. The fifteen PDB entries were NOT fetched. DepMap figures from the same session with the release read back (rule 83).

   ▸Evidence that argues against this (1)

   • A -0.04 DepMap gene effect with 1.3% of lines below -0.5 means PERK is dispensable in essentially every cancer cell line screened, which is difficult to reconcile with a paper reporting that PERK inhibition rescues a lethal phenotype - rescue requires the axis to be doing something. The resolution is that the screens never apply ER stress, but it does mean there is no independent functional-genomics support for PERK mattering in leukaemia. DepMap / Open Targets Platform 2026

• regulates →ATF4 / CREB2 (P18848) - the hub this graph inferred from its targets and never wrote, and its own transcript does not move under either menin inhibitor

  PERK phosphorylates eIF2alpha and is the key effector of the ISR to unfolded proteins, raising ATF4 translationally; its own transcript is flat under both inhibitors (max 0.35 log2) while ATF4 output falls 1.4-5.1

• related to →KMT2A::MLLT3 (MLL-AF9), t(9;11)(p21.3;q23.3)

  the PERK-eIF2alpha-ATF4 branch is functionally required in MLL-AF9 murine leukaemia stem cells: DNAJC10 loss activates it and blocking PERK with GSK2606414 RESCUES the phenotype (PMID 37496439)

• related to →The differentiation block — what it actually is, operationally, in KMT2A::AFF1 leukaemia

  the state that matters for this kinase is the leukaemia-stem-cell state, which bulk cell lines cannot address; a PERK inhibitor rescues LSC self-renewal in MLL-AF9 mice

• ← supportsATF4'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

  PERK's transcript is flat in all four lines on a well-expressed probe, so the null is a measurement and not a detection failure