CBX8 — the one place in this interactome where two good papers flatly disagree, and DepMap takes a side

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.

• The central resistance risk is that the target may simply not be required. A mean Chronos of +0.101 means CBX8-null cells grow slightly better; selection in a patient would therefore favour CBX8 loss under a CBX8-directed drug, and it costs the leukemia nothing.
• Compensation is structurally pre-built: CBX2, CBX4, CBX6 and CBX7 occupy the same canonical PRC1 slot with related chromodomains. A chromodomain ligand potent enough to matter is likely to hit several of them, and one that is selective for CBX8 is likely to be buffered by the others.
• Essential-function toxicity from the wrong direction: CBX8 is part of Polycomb repression, and Polycomb loss is tumour-promoting in several lineages. Chronically blocking a Polycomb reader in a child who has been cured of leukemia is a second-malignancy risk that no cell-line assay will surface.
• The chemistry and the hypothesis are aimed at different surfaces. UNC3866-class ligands occupy the chromodomain (5EQ0); the contested interaction is at residues 327-349. A negative result with existing compounds would not refute the Tan hypothesis, and a positive result would not confirm it — a trap that will produce a misleading paper if nobody states it.
• Delivery: the only ligands with published potency here are peptidomimetics against a shallow aromatic cage, a chemotype with a poor record of oral exposure and nuclear penetration. The 1.18 A structure makes the pocket look more tractable than the pharmacokinetics have proved to be.

1. Are the AF9 AHD point mutants used by Tan 2011 and Schmidt 2020 actually CBX8-selective, or do they also perturb DOT1L and AFF1/AF4 binding? Because all four partners share one surface that folds on binding, this is the most plausible single explanation of the conflict, and it can be tested directly by measuring all four partner interactions for each published mutant.
2. Is CBX8 required for transformation but not for maintenance? DepMap can only ever see maintenance, so a requirement that operates when a progenitor becomes leukemic and then lapses would produce precisely the observed pattern. No reported experiment separates the two phases in a human KMT2A::AFF1 system.
3. Do CBX2, CBX4, CBX6 or CBX7 buffer CBX8 loss in these cell lines? A combinatorial CBX knockout would say whether +0.101 means 'not needed' or 'replaceable', and those have very different implications for a drug.
4. Does UNC3866-class chromodomain blockade phenocopy CBX8 knockout in KMT2A-rearranged lines? Nobody has run the pharmacology, and it is a two-week experiment with an existing probe.
5. Why is CBX8 constrained in humans (pLI 0.9999, LOEUF 0.363, 8 observed loss-of-function variants against 39.7 expected) if Cbx8-null mice are viable with no haematopoietic defect and knockout is mildly advantageous in 1,208 cancer lines? Something selects against human heterozygous loss that none of these assays detects.

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

   CBX8 is not DepMap common-essential; mean Chronos across 1,208 lines is +0.101 with only 0.1% of lines depleted, the leukemia mean is +0.055, the 10 KMT2A-rearranged lines average +0.096 and the 4 KMT2A::AFF1 lines +0.109, giving selectivity_vs_leukemia +0.041 and selectivity_vs_all_lines -0.005; gnomAD gives pLI 0.9999, LOEUF 0.363, 8 observed versus 39.7 expected loss-of-function variants and mis_z 0.46; HPA classifies CBX8 as low tissue specificity, detected in many tissues, low immune cell specificity, nucleoplasmic.

   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, so a selectivity of +0.041 is far inside the noise band (any absolute value under 0.15 at n=10) and must not be read as a protective effect. A positive mean Chronos across 1,208 lines is a robust statement that whole-gene CBX8 knockout is not anti-proliferative; it is not a statement about the CBX8-AF9 contact, because knockout removes the chromodomain and the 327-349 segment together. DepMap measures maintenance of established leukemia and can never see a requirement at transformation. Paralogue buffering by CBX2/4/6/7 is not controlled for. Cell lines are not normal tissue.

   ▸Evidence that argues against this (1)

   • Tan et al. 2011 report that CBX8 is essential for MLL-AF9-induced leukemogenesis and that MLL-AF9 point mutants losing CBX8 binding cannot transform. If correct, the DepMap result reflects the difference between transformation and maintenance, or paralogue buffering in human lines, rather than dispensability of the interaction. Tan J et al. 2011
2. E2moderate confidence 3 of 5: One strong study, or several smaller consistent ones.paper

   CBX8 is essential for MLL-AF9-induced leukemogenesis; MLL-AF9 point mutants that lose CBX8 binding cannot transform; and Cbx8-null mice are viable with no haematopoietic defect.

   CBX8, a polycomb group protein, is essential for MLL-AF9-induced leukemogenesis · Tan J, Jones M, Koseki H, et al. · Cancer Cell 2011
   PMID 22094252DOI
   ▸What this evidence cannot show

   Confidence is downgraded to moderate not because of the paper's quality but because a later study on the same interface reports the opposite result (see counter_evidence) and DepMap agrees with the later study. The system is mouse MLL-AF9 in primary haematopoietic progenitors, not human KMT2A::AFF1. The decisive experiment is an AHD point mutant, and because all four AHD partners share one surface that folds on binding, a mutant chosen to abolish CBX8 binding may also perturb DOT1L or AFF1/AF4 binding; the abstract does not establish that the mutation is clean. The viable, haematopoietically normal Cbx8-null mouse in the same paper sets a high bar for an obligate requirement in the transformation of those same cells.

   ▸Evidence that argues against this (1)

   • Schmidt et al. 2020 report that disrupting direct CBX8/MLL-AF9 binding does NOT affect proliferation, while disrupting BCOR binding to the same AF9 AHD does, acting through altered EYA1, SIX and MYC activity. Schmidt CR et al. 2020
3. E3strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper

   Disrupting the direct CBX8/MLL-AF9 interaction does not affect proliferation, whereas disrupting BCOR binding to the same AF9 ANC1 homology domain is essential for leukemia and acts through altered EYA1, SIX and MYC activity; the BCOR 1176-1207 peptide bound to the MLLT3 AHD 500-568 is solved as 6B7G.

   BCOR binding to MLL-AF9 is essential for leukemia via altered EYA1, SIX, and MYC activity · Schmidt CR, Achille NJ, Kuntimaddi A, et al. · Blood Cancer Discovery 2020
   PMID 32954361DOI
   ▸What this evidence cannot show

   Directly contradicts PMID 22094252 on the CBX8 arm; both are interface-mutant studies in MLL-AF9 systems and the discrepancy is unresolved in the literature. This paper's positive claim about BCOR is itself not supported by DepMap, where BCOR has a mean Chronos of +0.025 and selectivity_vs_leukemia of +0.063 — so the agreement between this paper and DepMap is one-sided. The AHD mutants share the caveat that all four AHD partners use one surface that folds on binding, so mutant specificity is hard to guarantee. No affinity measurement is reported in the abstract.

   ▸Evidence that argues against this (1)

   • DepMap does not support the positive half of this paper's claim: BCOR knockout is not anti-proliferative in any context measured, including the 10 KMT2A-rearranged lines (mean +0.003). Broad Institute DepMap; Open Targets Platform 2026
4. E4established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.database

   Verified structures cover two independent CBX8 surfaces: the chromodomain with the peptidomimetic antagonist UNC3866 (5EQ0, X-ray 1.18 A, CBX8 residues 7-61) and with an H3K9 methyl-lysine peptide (3I91, X-ray 1.55 A, residues 8-61), and separately the CBX8 327-349 segment bound to the MLLT3/AF9 ANC1 homology domain 500-568 (2N4Q, NMR).

   RCSB PDB entries 5EQ0, 3I91 and 2N4Q, with per-entity UniProt accessions and modelled residue ranges resolved against the RCSB search and data APIs · RCSB Protein Data Bank · RCSB PDB 2026
   ▸What this evidence cannot show

   Identifiers, resolutions and residue ranges were resolved against the RCSB API in this session; ranges are RCSB's UniProt-mapped modelled segments, not independently curated domain boundaries. No structure contains both the chromodomain and the 327-349 segment, so nothing here speaks to their relative orientation in the intact 389-residue protein. 2N4Q is a 23-residue peptide on a disordered partner domain and carries no published affinity. The 1.18 A resolution of 5EQ0 applies to an isolated 55-residue domain with a peptidomimetic, which is the easiest crystallographic case in this dossier and is not representative of the AHD problem.

   ▸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. Chromodomains are small, well-behaved and bind short methylated peptides, so they crystallise easily; the limiting problem for CBX8 is selectivity across paralogues that share the fold and the ligand, which no number of structures resolves. RCSB PDB 2026
5. E5strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.database

   ChEMBL v34 target CHEMBL3232684 (Chromobox protein homolog 8) has 15 activity rows at pChEMBL >= 6 with a best value of Kd 120 nM (CHEMBL4635096, binding assay), no curated mechanism of action, and all potent ligands being peptidomimetic chromodomain ligands.

   ChEMBL database version 34, target CHEMBL3232684 (Chromobox protein homolog 8): target_search, get_bioactivity and get_mechanism queries · EMBL-EBI ChEMBL · ChEMBL 2026
   ▸What this evidence cannot show

   All of this chemistry is directed at the chromodomain and therefore does not test the contested CBX8-AF9 AHD contact at all; nothing in ChEMBL addresses CBX8 residues 327-349. The CBX chromodomains are closely related, so within-family selectivity rather than potency is the historical limiting problem and ChEMBL rows do not report it. Counts are page-local under a 25-row API limit. Absence of a curated mechanism record partly reflects that mechanism annotations tend to live on family or protein-protein-interaction targets rather than single-protein 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. Fifteen rows against menin's 499 ranks attention, not druggability — and this graph's own DepMap analysis finds the CBX8 co-dependency in KMT2A-rearranged lines indistinguishable from zero, which is the fact that should govern. Oncofusion knowledge graph (internal analysis) 2026
6. E_q95_cbx8_loc1established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.database

   CBX8's entire UniProt subcellular annotation is one value, 'Nucleus', on one ECO:0000269 - PubMed 21282530 - and that paper is a tandem-affinity-purification and tandem-mass-spectrometry interaction-proteomics study of the CBX family whose abstract describes no imaging experiment at all; its only localisation statement is a closing inference, that differences in CBX protein interactions 'would account, at least in part, for distinct subnuclear localization of the CBX family members'. This is the SAME citation and the SAME defect that tranche 4 recorded on gene-ring1, where UniProt's 'Nucleus speckle' for RING1 rests on this one paper.

   UniProtKB Q9HC52 (CBX8_HUMAN), SUBCELLULAR LOCATION comment with evidence code and source, and gene synonym list, fetched live 2026-09-22 · UniProt Consortium · Databases 2026
   ▸What this evidence cannot show

   A statement about the database record. CBX8 is a chromodomain H3K9/K27me reader and a PRC1 subunit, so nobody disputes that it is nuclear; what is recorded is that the annotation's single experimental source measured interactions, not position, and that the same source has now produced a questionable subnuclear term for two different PRC1 subunits in this graph.

   ▸Evidence that argues against this (1)

   • THE PAPER IS BETTER THAN THIS AUDIT MAKES IT SOUND, AND THE DEFECT IS ASYMMETRIC BETWEEN THE TWO GENES IT AFFECTS - which is exactly brief rule 91's point that the audit is per GENE, not per paper. For CBX8, UniProt took only 'Nucleus' from it, and a mutually-exclusive-CBX-complex purification legitimately supports a nuclear call for a PRC1 subunit; the annotation is imprecise but not wrong. For RING1 the same paper was made to support 'Nucleus SPECKLE', a specific compartment the paper never images and which returns zero records in the independent literature. Same citation, one acceptable use and one unacceptable one, decided entirely by which term was drawn from it. This is the second such pair in this series - tranche 5 found PMID 25593309 to be a KMT2A miscitation and a correct KAT2A citation - and the generalisable rule is that a shared citation's quality has to be re-decided for every entry that carries it. Vandamme J et al. 2011
7. E_q95_cbx8_loc2strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper

   CBX8's subnuclear address is the POLYCOMB BODY and the measurement is CBX8-resolved with kinetics: fluorescently tagged CBX proteins with BMI1 and RING1 form foci of different sizes and intensities in human U2OS cells, fluorescence recovery inside the foci is substantially slower than the diffusion-plus-transient-binding behaviour of the surrounding nucleoplasm, CBX4-GFP foci are MORE STABLE than CBX8-GFP foci, and the presence of CBX4 or CBX8 in a focus has only a minor effect on BMI1 and RING1 recovery kinetics - so the paralogues differ from one another inside the same compartment.

   PRC1 components exhibit different binding kinetics in Polycomb bodies. · Vandenbunder B, Fourré N, Leray A, Mueller F, Völkel P, Angrand PO, Héliot L · Biol Cell 2014
   PMID 24460908DOI
   ▸What this evidence cannot show

   Human U2OS osteosarcoma cells with fluorescently tagged proteins - overexpression, no haematopoietic cell, and no endogenous CBX8 image anywhere. This is the ONLY paper in PubMed that puts CBX8 and a Polycomb body or focus in one title or abstract: CBX8[tiab] against a Polycomb-body and foci clause returns 1 record, and CBX8[tiab] against a broad body/foci/condensate/subnuclear/localisation clause returns 13, of which this and the interaction-proteomics paper are the only two that measure position.

   ▸Evidence that argues against this (1)

   • THE AUTHORS' OWN CONCLUSION ARGUES AGAINST THE CONDENSATE READING OF THEIR DATA, AND THIS GRAPH SHOULD FOLLOW THEM RATHER THAN THE DROPLET LITERATURE. They interpret the slow intra-focus recovery as 'PcG spreading onto chromatin inside foci' - that is, binding to a locally dense array of target sites - not as a liquid phase. Brief-recorded methodology says the two commonest assays cannot tell the models apart: classical FRAP recovery and 1,6-hexanediol sensitivity both fail to distinguish liquid-liquid phase separation from binding to spatially clustered sites, and only calibrated half-bleaching that measures internal mixing discriminates them. This paper used classical FRAP. So the defensible claim is a kinetically distinct, chromatin-anchored compartment; a CBX8 condensate claim is not supported by this measurement, and the clustered-binding model is the more interesting one here anyway because it predicts sequence-specific targeting where phase separation predicts bulk enrichment. Muzzopappa F et al. 2022
8. E_q95_cbx8_loc3moderate confidence 3 of 5: One strong study, or several smaller consistent ones.paper

   CBX8's cell-type and cell-state evidence is entirely functional, entirely myeloid-mouse, and internally contradictory, while its only positional evidence is in a human osteosarcoma line: CBX8 has been reported essential for MLL-AF9-induced leukaemogenesis with MLL-AF9 point mutants that lose CBX8 binding failing to transform, and separately reported dispensable, with disruption of the direct CBX8/MLL-AF9 interaction leaving proliferation unaffected while disruption of BCOR binding to the same AF9 AHD surface does not. Cbx8-null mice are viable with no haematopoietic defect, and DepMap gives CBX8 a mean Chronos of +0.101 across 1,208 lines with 0.1% of lines depleted.

   CBX8, a polycomb group protein, is essential for MLL-AF9-induced leukemogenesis. · Tan J, Jones M, Koseki H, Nakayama M, Muntean AG, Maillard I, Hess JL · Cancer Cell 2011
   PMID 22094252DOI
   ▸What this evidence cannot show

   This entry already carries the conflict as unresolved and this localisation pass does not resolve it - it adds the observation that the two positions are about a REQUIREMENT in mouse MLL-AF9 myeloid transformation while the only compartment data are in human U2OS, so no experiment in this literature asks whether CBX8's Polycomb-body residence changes in a leukaemic cell at all. Brief rule 73 applies to both arms: neither is infant B-lineage KMT2A::AFF1.

   ▸Evidence that argues against this (1)

   • THE STATE CLAIM AND THE COMPARTMENT CLAIM MAY BE THE SAME MEASUREMENT AND THIS IS TESTABLE. The Schmidt result is that the AF9 AHD surface matters through BCOR rather than through CBX8, and BCOR is a non-canonical PRC1.1 subunit - the same complex family as this tranche's TRIM27 entry. If the AHD-bound species is PRC1.1 rather than a CBX8-containing canonical PRC1, then the relevant compartment question is not whether CBX8 is in a Polycomb body but whether PRC1.1 is, and the FRAP paper measured canonical components (CBX2/4/6/7/8, BMI1, RING1) and no PRC1.1-specific subunit. That would make the graph's Polycomb-body address correct for canonical PRC1 and unmeasured for the complex the fusion actually engages - which is a sharper statement of the gap than either side of the requirement conflict. Schmidt CR et al. 2020
9. E_q95_cbx8_loc4strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.database

   Bounded 2026-09-22, and this gene's alias is a 117-fold retrieval trap that has to be reported with the count: CBX8[tiab] alone returns 146 records, while UniProt's curated gene synonym 'PC3' returns 17,070 - overwhelmingly the PC-3 prostate carcinoma cell line - so the naive alias-expanded gene clause returns 17,233 and is 99.2% noise. Inside the clean symbol: CBX8 with a subnuclear clause returns 42 records (contaminated by PC3), CBX8 against a Polycomb-body and foci clause returns 1, CBX8 against a nuclear-speckle clause returns 0, and CBX8 with a localisation clause AND the KMT2A/MLL fusion clause returns 1 with a subnuclear-plus-fusion count of 0.

   NCBI E-utilities esearch db=pubmed; every count stored with the querytranslation read back, in data.localization_bounded_absence · NCBI E-utilities · Databases 2026
   ▸What this evidence cannot show

   Titles and abstracts only; supporting figures not searched (rule 80). The PC3 measurement is a brief-rule-69-part-1 case: alias expansion here does not recover a hidden literature, it destroys the query. The safe alias is 'hPC3', which returns 18 records; 'RC1', CBX8's other UniProt synonym, was not searched because it is a two-letter-plus-digit string of the same class.

   ▸Evidence that argues against this (1)

   • THE ALIAS TRAP IS WORSE THAN A CELL-LINE COLLISION AND THE REGISTRY SHOULD SAY SO, BECAUSE THREE DIFFERENT OBJECTS ANSWER TO PC3. It names the PC-3 prostate carcinoma line, which is what most of the 17,070 records are; it is a documented alias of NR4A1/Nur77 in the immediate-early-gene literature; and PC3 also abbreviates proprotein convertase 3. That is the same four-way structure the registry already records for CD numbers in rule 69 part 4 - one string, several classes of referent - except that here the string is a CURATED UNIPROT GENE SYNONYM, so a pipeline that expands aliases from the canonical database is the thing that walks into it. The counter-consideration worth keeping: a bare CBX8 search is CLEAN, so unlike SEPT9 or SKIC8 this gene loses nothing by refusing alias expansion, and the correct prescription is simply to never expand it. Hernández-Muñoz I et al. 2005

• binds →KMT2A::AFF1 (MLL-AF4), t(4;11)(q21;q23)

  binds the partner side; two good papers disagree about what it does there

• ← depends onPolycomb bodies — indirect evidence via CBX8, undercut by the finding that CBX8-PRC1 condensates are dissolved rather than built by the rest of PRC1

  the only route by which a KMT2A fusion touches Polycomb machinery

• ← recruitsWhere KMT2A::MLLT1 actually sits: chromatin occupancy is mapped, subnuclear localisation of the endogenous fusion has never been imaged

  hPc3/CBX8 recruited by the ENL interaction domain retained in MLL-ENL

• ← bindsBCOR, not CBX8, is the AHD contact MLL-AF9 cannot lose - and the same two partners bind ENL far more tightly than AF9

  CBX8 binds the AF9 AHD weakly; the direct contact is not a requirement

• ← refutesPhase-2 dossier: CBX8 - the candidate this review would kill, and the reasons are all numeric

  the executed DepMap data sides against a CBX8 requirement; no-go this phase

• related to →Polycomb bodies — indirect evidence via CBX8, undercut by the finding that CBX8-PRC1 condensates are dissolved rather than built by the rest of PRC1

  CBX8-RESOLVED and kinetically characterised: tagged CBX8 forms nuclear foci in human U2OS whose internal FRAP recovery is substantially slower than the nucleoplasm, and CBX4-GFP foci are MORE STABLE than CBX8-GFP foci (PMID 24460908). Consistent with tranche 4's RING1 treatment. Not a condensate claim - the authors read it as PcG spreading onto clustered chromatin, and classical FRAP cannot separate the models (Q95 2026-09-22)

• related to →RING1 (RING1A) - the paralogue that is not RING1B, and the reason a substring search for 'RING1' in this graph returns the wrong gene

  SAME UNIPROT CITATION, DIFFERENT VERDICT - the per-gene nature of a rule-91 audit. PMID 21282530, a CBX-family TAP-MS study with no imaging, is UniProt's sole source for CBX8's bare 'Nucleus' (acceptable if imprecise) and for RING1's specific 'Nucleus speckle' (not adopted by tranche 4, and RING1 against nuclear speckles returns 0). One paper, one good use and one bad one (Q95 2026-09-22)

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

  The CBX8 focus is a worked case for this node's standard: the only CBX8 positional measurement is classical FRAP, which cannot distinguish liquid-liquid phase separation from binding to spatially clustered sites, and the authors themselves conclude chromatin spreading rather than a phase. A CBX8 condensate claim is not supported by the data that exist (Q95 2026-09-22)