P-TEFb (CDK9 + cyclin T1) — the deepest dependency and the clearest therapeutic-index wall in the interactome

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.

• Essential-function toxicity is the whole story and it is not fixable downstream. 99.3% of 1,208 cancer cell lines are depleted by CDK9 knockout and the KMT2A-r selectivity is -0.077, inside noise. Suppressing global Pol II pause release kills proliferating normal tissue — gut, marrow, skin — by exactly the mechanism that kills the tumour. Fifteen years and seven clinical compounds did not find a way around it: the ceiling is phase 3, with no phase 4.
• The druggable pocket is on the wrong subunit. Every clinical agent binds the CDK9 ATP site; every verified scaffold complex (4IMY, 4OR5, 4OGR, 5L1Z, 6CYT) places AFF4 on cyclin T1. So the only interface that is plausibly fusion-adjacent is a shallow, extended, largely disordered peptide track on a non-catalytic subunit — the hardest PPI class to drug — and the easy pocket is the one with no selectivity.
• The complex is explicitly shared with normal biology. PMID 20153263 states that the AF4-family/ENL-family/P-TEFb higher-order complex serves physiologic as well as oncogenic MLL-dependent transcription. There is no version of blocking P-TEFb recruitment that spares wild-type KMT2A function.
• Resistance and compensation have an easy route: partial CDK9 inhibition selects for cells with higher P-TEFb activity or altered pause-release setpoints, and the cyclin partners (T1, T2, K) offer combinatorial redundancy that a CCNT1-directed agent would run straight into — the near-benign CCNT1 knockout is already telling us that removing one cyclin is survivable.
• The CDK9 LOEUF of 1.162 is a trap for anyone building an automated safety score. Read naively it says CDK9 is a safe, LoF-tolerant gene. It is underpowered (39 expected LoF events in a 372-residue gene) and it measures heterozygote selection, not the homozygous-equivalent state a potent inhibitor creates. Any pipeline that ranks targets on LOEUF will rank the least tolerable target in this interactome as one of the most tolerable.

1. Why is cyclin T1 knockout nearly benign (-0.206, 8.4% of lines depleted) when its obligate kinase partner is required by 99.3% of lines? If paralogous cyclins substitute, then the cyclin T1 surface is the buffered half of P-TEFb and the AFF4-binding site on it becomes the only part of this node with a defensible therapeutic index. A double CCNT1/CCNT2 knockout would settle it, and we could verify no citation for the substitution hypothesis this cycle.
2. Does KMT2A::AFF1 recruit P-TEFb through its own AFF1 module, and if so does it hold P-TEFb with higher occupancy than wild-type complexes do? Every verified structure of scaffold-bound P-TEFb uses AFF4, not AFF1, and there is no KMT2A::AFF1-specific interactome paper. Without this, 'the fusion is a P-TEFb-recruiting factor' is an inference from paralogy.
3. Is there any dosing or scheduling regime in which the 0.077 selectivity gap becomes exploitable — for example if KMT2A-fusion target genes have shorter transcript half-lives or higher P-TEFb dependence than housekeeping genes, so that transient partial inhibition hits them first? That is a nascent-transcription time-course experiment and it is the only route by which CDK9 could work in this disease.
4. Would a CDK9 degrader behave differently from an ATP-site inhibitor here, or identically worse? Degradation removes the kinase from every cell, which is closer to the DepMap experiment that gives 99.3% depletion; the a priori expectation is that it widens the toxicity rather than the window.
5. Can the LOEUF-versus-DepMap discordance be generalised? CDK9 (tolerant LOEUF, essential) and AFF4 (extreme LOEUF constraint, dispensable) are mirror images within one complex. If gene length and expected-LoF count explain most of the discordance, the project should stop using LOEUF as a toxicity axis at all and say so explicitly.

1. E1strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.database

   CDK9 is DepMap common-essential with a mean Chronos of -1.536 across 1,208 lines and 99.3% of lines depleted; means are -1.610 across 64 leukemia lines, -1.687 across 10 KMT2A-rearranged lines (the deepest value computed for any gene in any subset in this interactome) and -1.552 across 4 KMT2A::AFF1 lines, giving selectivity_vs_leukemia -0.077 and selectivity_vs_all_lines -0.151. CDK9's gnomAD constraint is the outlier among deep dependencies here: pLI 0.0, LOEUF 1.162, 34 observed versus 39.0 expected loss-of-function variants, missense z 2.55. Cyclin T1 (CCNT1) is by contrast NOT common-essential: mean Chronos -0.206, 8.4% of lines depleted, leukemia -0.198, KMT2A-r -0.253, KMT2A::AFF1 -0.047, selectivity_vs_leukemia -0.055, with pLI 1.0, LOEUF 0.397, 18 observed versus 67.3 expected LoF and missense z 2.02. HPA classifies both subunits as low tissue specificity, detected in all tissues, low immune cell specificity.

   DepMap CRISPR (Chronos) gene effect via the Open Targets Platform GraphQL API; loss-of-function constraint via the gnomAD GraphQL API; normal-tissue expression via the Human Protein Atlas API · Broad Institute DepMap; gnomAD; Human Protein Atlas; Open Targets Platform · Databases 2026
   ▸What this evidence cannot show

   Means were computed by us from per-line API values rather than read from a DepMap release table, and the API does not state the release version. Only 10 KMT2A-rearranged lines (4 KMT2A::AFF1) contribute to the selective score, so -0.077 is inside our declared +/-0.15 noise band and must be read as no detectable selectivity. Cell lines are not normal tissue, and common-essentiality measured in immortalised proliferating lines overstates the requirement in quiescent tissue while saying nothing about it directly. Whole-gene knockout of CDK9 removes the kinase entirely and is not equivalent to partial pharmacological inhibition, which is what a drug does; conversely the near-benign CCNT1 knockout may reflect paralogous cyclin substitution rather than dispensability of the cyclin function. gnomAD LOEUF for CDK9 is computed on only 39.0 expected loss-of-function variants in a 372-residue gene and is therefore underpowered; its value of 1.162 should be read as uninformative rather than as evidence of tolerance.

   ▸Evidence that argues against this (1)

   • CDK9's pLI of 0.0 and LOEUF of 1.162 appear to contradict its DepMap essentiality. They do not: LOEUF is the upper bound of the 90% CI on observed/expected LoF and is inflated toward 1.0 for small genes with few expected events, and it measures selection against heterozygous loss over evolutionary time rather than cell-autonomous viability at zero copies. CDK9's missense z of 2.55 is the better-powered channel and does show constraint. The tolerant-looking LOEUF must not be taken as evidence that pharmacological CDK9 inhibition is safe. gnomAD Consortium 2026
2. E2established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.paper

   The AFF4 N-terminus (residues 2-73) binds human P-TEFb by meandering over the surface of cyclin T1 rather than contacting CDK9 stably, and HIV Tat enhances P-TEFb's affinity for AFF4; the complex was solved as PDB 4IMY at 2.94 A with CDK9 1-330, cyclin T1 1-264 and AFF4 2-73.

   The AFF4 scaffold binds human P-TEFb adjacent to HIV Tat · Schulze-Gahmen U, Lu H, Zhou Q, Alber T · eLife 2013
   PMID 23471103DOI
   ▸What this evidence cannot show

   A crystal structure of a truncated ternary assembly: CDK9 is modelled 1-330 of 372 and cyclin T1 only 1-264 of 726, so the full-length context is absent. The abstract reports no numeric dissociation constant for the AFF4/P-TEFb interaction, so no Kd may be attributed to it. The Tat findings are HIV biology and their relevance to a leukemia setting is by analogy to scaffold competition only.

   ▸Evidence that argues against this (1)

   • Meandering over cyclin T1 without stable CDK9 contact is a weak, extended interface — the hardest class to drug with a small molecule — and the structure was solved with a 72-residue AFF4 fragment rather than the protein. RCSB PDB 2026
3. E3established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.database

   P-TEFb has extensive verified structural coverage in two classes: ATP-site inhibitor co-crystals of CDK9/cyclin T1 (3BLQ 2.90 A with ATP; 3BLR 2.80 A with flavopiridol; 3MY1 with DRB; 3TN8 and 3TNH with CAN508; 3LQ5 with S-CR8; 4BCF, 4BCG, 4BCH, 4BCI, 4BCJ with 2-amino-4-heteroaryl-pyrimidines; 6GZH with compound A86 and full-length cyclin T1 residues 1-726; 6W9E with MC180295; 6Z45; 7NWK), and scaffold- or Tat-bound complexes in which AFF4 engages cyclin T1 (4IMY 2.94 A; 4OR5 2.90 A with HIV Tat and AFF4 32-69; 4OGR 3.00 A; 5L1Z 5.90 A with TAR RNA; 6CYT 3.50 A). Every clinical agent occupies the CDK9 ATP site and every scaffold complex places AFF4 on cyclin T1, i.e. the druggable pocket and the fusion-relevant interface are on different subunits.

   RCSB PDB entries 3BLQ, 3BLR, 3MY1, 3TN8, 3TNH, 3LQ5, 4BCF, 4BCG, 4BCH, 4BCI, 4BCJ, 6GZH, 6W9E, 6Z45, 7NWK, 4IMY, 4OR5, 4OGR, 5L1Z and 6CYT, with per-entity UniProt accessions, ligands, methods, resolutions and modelled residue ranges resolved against the RCSB APIs · RCSB Protein Data Bank · RCSB PDB 2026
   ▸What this evidence cannot show

   Residue ranges are RCSB's UniProt-mapped modelled segments, not independently curated domain boundaries. Most CDK9 entries model only residues ~1-330 of 372 and most truncate cyclin T1 at ~260-266 of 726, so the C-terminal halves of both subunits are structurally unaddressed in nearly every entry (6GZH is the exception for cyclin T1). The 5L1Z entry at 5.90 A resolves overall architecture only. Co-crystallisation with an inhibitor establishes binding mode, not cellular selectivity: several of these ligands are pan-CDK.

   ▸Evidence that argues against this (1)

   • Counting PDB entries measures how tractable a protein was to CRYSTALLISE, which correlates with being small, ordered and well-behaved rather than with being druggable. What decides design value is whether the deposited density covers the site that matters, at a resolution that supports modelling, in the relevant state — and this graph has repeatedly found structures described in the direction that overstates what is resolved. Extensive ATP-site coverage is coverage of the most conserved surface in the kinome, which is why CDK9 selectivity against other CDKs has been the limiting problem for two decades rather than potency. ChEMBL, EMBL-EBI 2026
4. E4strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.database

   CDK9 is the best-drugged node in this interactome: ChEMBL v34 target CHEMBL3116 has 1,948 activity rows at pChEMBL >= 6 with a best IC50 of 2.5 nM (CHEMBL424696) and seven curated mechanism-of-action INHIBITOR records — alvocidib/flavopiridol (CHEMBL428690, max phase 3), dinaciclib (CHEMBL2103840, phase 3), seliciclib (CHEMBL14762, phase 2), zotiraciclib (CHEMBL1944698, phase 2), riviciclib (CHEMBL3545283, phase 2), RGB-286638 (CHEMBL3545083, phase 1) and SNS-032 (CHEMBL296468, phase 1) — with none at phase 4. By contrast menin (CHEMBL1615381) carries a single curated inhibitor record at max phase 4, approved. Cyclin T1's own target CHEMBL2108 has 192 potent rows with a headline IC50 of 0.011 nM (CHEMBL3652915), but these are IMAP assays on the CDK9/cyclin T1 heterodimer attributed to the non-catalytic cyclin record and must be treated as CDK9 chemical space; CCNT1 has no curated mechanism drug.

   ChEMBL database version 34, targets CHEMBL3116 (cyclin-dependent kinase 9) and CHEMBL2108 (cyclin-T1): target_search, get_bioactivity and get_mechanism queries · EMBL-EBI ChEMBL · ChEMBL 2026
   ▸What this evidence cannot show

   Distinct-compound counts are page-local because of a 25-row API limit, so the 'best IC50 2.5 nM' is the best within one sampled page dominated by a single publication rather than a potency-ranked view of all 1,948 rows. Several of the seven curated inhibitors (alvocidib, dinaciclib, seliciclib) are multi-CDK agents whose clinical effects cannot be attributed to CDK9 alone, so the phase-3 ceiling is a ceiling for pan-CDK pharmacology as much as for CDK9 specifically. Max-phase fields record the highest phase reached for any indication, not success; absence of a phase-4 record is consistent with discontinuation, with ongoing development, or with curation lag. The cyclin T1 rows are not erroneous measurements, only mis-attributed ones.

   ▸Evidence that argues against this (1)

   • A ChEMBL activity-row count measures how much medicinal chemistry has been DONE, not how tractable the target is, and the best reported affinity belongs to the most optimised compound in whichever series someone chose to pursue. Zero rows means nobody tried; thousands of rows means a target had a decade of industrial attention. 1,948 rows and seven mechanism records make CDK9 the best-drugged node here and also the one with the longest record of clinical disappointment in leukaemia; flavopiridol and its successors reached patients repeatedly. Being drugged is not being useful. Blood 2018
5. E5strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper

   AFF4 is a component of the ELL/P-TEFb elongation complex and a shared subunit of MLL chimeras, linking transcription elongation to leukemia; AFF4 is required for the stability of that complex.

   AFF4, a component of the ELL/P-TEFb elongation complex and a shared subunit of MLL chimeras, can link transcription elongation to leukemia · Lin C, Smith ER, Takahashi H, et al. · Molecular Cell 2010
   PMID 20159561DOI
   ▸What this evidence cannot show

   Pan-MLL-chimera proteomics and RNAi, not KMT2A::AFF1-specific; we could verify no KMT2A::AFF1-specific published interactome paper. No affinity measurement is reported. Because KMT2A::AFF1 carries its own AFF-family module, the AFF4-mediated recruitment described here may not be the route P-TEFb takes in that particular fusion.

   ▸Evidence that argues against this (1)

   • Complex composition is not requirement. The AEP/SEC papers establish what co-purifies and what a reconstituted complex does; they do not establish that each subunit is rate-limiting in a patient's leukaemia, and this project's own dependency data separates the members by an order of magnitude. Nature 2012
6. E6strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper

   A higher-order complex containing AF4-family and ENL-family proteins together with P-TEFb facilitates both oncogenic and physiologic MLL-dependent transcription.

   A higher-order complex containing AF4 and ENL family proteins with P-TEFb facilitates oncogenic and physiologic MLL-dependent transcription · Yokoyama A, Lin M, Naresh A, Kitabayashi I, Cleary ML · Cancer Cell 2010
   PMID 20153263DOI
   ▸What this evidence cannot show

   The paper's own scope explicitly includes physiologic MLL-dependent transcription, so it establishes that the P-TEFb-containing assembly is shared between normal and oncogenic transcription rather than fusion-specific. It reports no affinity measurement and does not quantify P-TEFb occupancy differences between fusion and wild-type complexes.

   ▸Evidence that argues against this (1)

   • Complex composition is not requirement. The AEP/SEC papers establish what co-purifies and what a reconstituted complex does; they do not establish that each subunit is rate-limiting in a patient's leukaemia, and this project's own dependency data separates the members by an order of magnitude. Nature 2012

• ← related toAFF1 / AF4 — domain architecture and residue coordinates

  a surface loop of the AFF-family CHD is a CDK9 substrate

• ← part ofCCNT1 / cyclin T1 — the regulatory half of P-TEFb, and the one part of the elongation core with any dependency variance

  regulatory cyclin of P-TEFb

• ← bindsAFF2 (FMR2) — the SEC-L2 scaffold, and the AFF-family member with no dependency anywhere

  SEC-L2 contains P-TEFb

• ← related toSUPT5H (SPT5) — the pausing factor the KMT2A field has never studied, and the counted absence that undercuts 'the fusion hijacks pause release'

  SPT5 CTR1 (aa 754-817) and CTR2 (aa 844-950) are P-TEFb substrates; the graph's CDK9 axis had no substrate entry before this one

• ← bindsThe AFF-family N-terminus on cyclin T1: Kd 10 nM, a 2.94 A crystal, and every number in it measured on AFF4 rather than on AFF1

  cyclin T1 hot spots W207, W210, R165

• ← related toAFF1-AFF4 heterodimerisation through the C-terminal homology domain: the one SEC interface KMT2A::AFF1 certainly keeps, defined by a 2.2 A structure of the wrong subunit

  CDK9 phosphorylates a surface loop of the dimerisation domain it is recruited by

• part of →Mechanism class 1 — chromatin and elongation hijack (KMT2A-r, NUP98-r)

  the kinase that releases paused polymerase; deepest dependency in the complex

• ← inhibitsCDK9 inhibition in KMT2A-rearranged infant leukemia — the most direct hit on the hijacked machinery, and the narrowest window

  the deepest dependency, and the narrowest therapeutic window

• ← refutesCo-dependency scan: what else does a menin-dependent cell need? (run, not planned)

  pan-essential genes sit at r near zero; a flat profile cannot correlate, so this scan cannot find targets in the essential core

• ← refutesDiffing the core scaffold's curated interactome against the graph: 46 of 217 partners are things here, CDK9 is not one of them, and the filter I built to rank the rest was counting one platform twice

  a summary entry covering two proteins, where only one got an entry of its own

• ← recruitsWhat LEDGF actually does for the fusion: it brings AF4, CDK9, BRD4, P-TEFb and Ser2P Pol II to the target gene, and H4K16ac with them

  P-TEFb and CDK9 occupancy fall on LEDGF knockdown

• ← related toHypothesis: LEDGF licenses elongation at fusion target genes rather than anchoring the fusion, which predicts that a LEDGF agent is epistatic with BET and CDK9 inhibition and invisible in a fusion-ChIP assay

  predicts epistasis with CDK9 inhibition

• ← related toThe case that the KMT2A-fusion condensate is an artefact — what 1,6-hexanediol, overexpressed tags and in-vitro droplets actually prove

  1,6-hexanediol at droplet-dissolving concentrations inactivates the SEC kinase itself

• ← related toHypothesis: KMT2A::AFF1 does not require phase separation — recruitment biochemistry is sufficient, and the condensate model does not transfer from NUP98/ENL

  Methodological counter (Duster 2021, PMID 33814344): 1,6-hexanediol at droplet-dissolving concentrations inactivates CDK9/P-TEFb, so hexanediol-based condensate tests on SEC read out a poisoned kinase rather than a dissolved phase.

• ← recruitsKMT2A::MLLT1 engages P-TEFb through one 66-residue surface: the ENL AHD, residues 494-559, and L550E breaks it

  CDK9/cyclin T1 arrive in the AF4/AF5q31/ENL complex that the ENL AHD joins

• ← recruitsAFF1 AF4_int motif to AF9/ENL AHD: the index fusion's only documented elongation interface, and it competes with DOT1L

  route to P-TEFb via AEP/SEC

• ← depends onP-TEFb phosphorylates AFF4 and rearranges the complex the fusion hijacks

  CDK9 T186 autophosphorylation is induced by SEC or BRD4 binding

• ← related toThe super elongation complex, sampled from the RNA side: ELL, ENL, CDK9 and cyclin T1 are all in the proximity record - ELL2, EAF1/2, AF9 and AF10 are at zero

  CDK9 recovered by eleven baits and cyclin T1 by sixteen

• ← part ofCDK9 - the kinase the whole SEC story runs through, which this graph held only as half of a two-protein summary entry

  catalytic subunit of the P-TEFb module the graph previously held as a single entry

• ← regulatesThe 7SK snRNP - a stable RNA core with a reversibly docked kinase, which is why the reservoir is a switch and not a sink

  holds more than half of nuclear P-TEFb in an inactive state in HeLa

• ← part ofCCNT2 / cyclin T2 - expressed in every KMT2A-rearranged line and needed by none of them, so the cyclin T split is real rather than an expression artefact

  P-TEFb is CDK9 with cyclin T1 or cyclin T2

• ← related toThree hours after the fusion is degraded to 2.5-7.1% of its own sites, NOT ONE of the eight complex components profiled beside it falls below 68% retention genome-wide - menin 80.0%, DOT1L 68.7%, CDK9 83.6%, AFF4 92.7%, RNA Pol II 99.6%, pS2 98.3%, pS5 86.8% - so nothing in the complex is held on chromatin by the fusion at genome scale

  P-TEFb's catalytic subunit CDK9 is retained at 83.6% of its strongest sites when the fusion is removed, so P-TEFb's genome-wide chromatin residence is not fusion-dependent at 3 h.

• ← related toCDK9 peak widths more than double after fusion degradation, firing the pre-registered re-score trigger

  P-TEFb catalytic subunit measured at fusion target promoters after degradation

• part of →SEC-L2 — the AFF2-scaffolded super-elongation-like complex: P-TEFb plus ENL and AF9, and NO ELL

  P-TEFb (CDK9/cyclin T1-T2) is a SEC-L2 subunit

• part of →SEC-L3 — the AFF3-scaffolded complex, and the only SEC-family member with a known physiological job: genomic imprinting and X inactivation

  P-TEFb is a SEC-L3 subunit

• part of →SEC — the AFF1/AFF4-scaffolded super elongation complex, the particle KMT2A::AFF1 is a permanent subunit of

  P-TEFb is the catalytic module of SEC

• ← related toThe index fusion's 3' partner has ONE observed binary interaction in IntAct and forty-two residues of solved structure - every SEC contact attributed to AFF1 is an AFF4 result wearing AFF1's name

  AFF1-CDK9 is 16 of 16 spoke-expanded and AFF1-CCNT1 8 of 8, with no MI:0407 record and no shared structure

• ← related toCHUK (IKK-alpha) — the only kinase that sits in BOTH NF-kB arms, which makes it the one subunit whose inhibition has an ambiguous sign in this disease

  H3S10ph written by IKK-alpha recruits P-TEFb at induced promoters (PMID 12789342), which is an NF-kB-pathway route to the elongation machinery independent of the AFF1/AFF4 scaffold this graph is built on. Shown in fibroblasts and macrophages; never tested in a fusion cell (Q93 2026-09-21)

• ← recruitsWhat the reciprocal product assembles: P-TEFb, an H3K4 methyltransferase, an H3K79 methyltransferase, an arginine methyltransferase and an acetyltransferase

  P-TEFb kinase reported as a constituent; the elongation effect on ALOX5 is flavopiridol-sensitive.

• ← related toA two-compartment switch for a SEC cargo: nucleolar retention versus AFF4-induced nucleoplasmic droplets, set by one methylated arginine

  Tat's transactivation runs through P-TEFb within the SEC, the complex whose droplets are the active compartment