Oncofusion

start / graph / ana-the-kmt2a-aff1-surfaceome-screen-and-its-zero

Analysismoderate confidence 3 of 5: One strong study, or several smaller consistent ones.Q138-kmt2a-aff1-surfaceome

A surfaceome-first screen for KMT2A::AFF1: 2,404 surface proteins, 169 expressed in the fusion's own cell lines, 15 fusion-enriched, 3 clean in normal tissue - and the two sets do not intersect

also called KMT2A::AFF1 surface target screen · Q138 surfaceome screen

We asked whether KMT2A::AFF1 leukaemia carries any protein on its outer surface that a CAR-T cell could attack without also attacking healthy organs, checking every surface protein in the human proteome rather than the handful already in use. The proteins that are specific to this leukaemia turned out to be present in normal blood vessels or brain, and the proteins absent from normal organs turned out not to be specific to this leukaemia.

The same thing, for a scientist

Starting from the surfaceome rather than from a curated target list, 2,404 reviewed human proteins carrying a UniProt Cell-membrane location **and** an annotated extracellular topological domain were scored against HPA cell-line nTPM in the three Cellosaurus-confirmed KMT2A::AFF1 lines (SEM, RS4;11, MV4-11), against a KMT2A-rearranged-other-partner arm, against a B-cell-precursor leukaemia comparator, and against 51 normal tissues. **169** genes cleared the expression filter in the KMT2A::AFF1 B-lineage arm; **15** cleared both fusion-enrichment thresholds; **3** (CD19, SIT1, S1PR4) were clean in normal parenchyma; and **the intersection is empty**. The pre-registered rule-of-three bound on that zero is **under 1.78%** of the 169 tested. The branch reached is the pre-registered **C - NEITHER**: surface, expressed and parenchymally safe genes exist, fusion-enriched genes exist, and they are disjoint. The binding constraint on this disease's antigen problem is therefore neither surface localisation nor normal-tissue safety taken alone - it is that the two cannot be had together. **And the zero has a boring explanation, tested the same day.** Only **3 of 169** index-expressed surface proteins are parenchymally clean at all - a base rate of **1.8%** - which by itself predicts 0.27 clean genes among any 15. An expression-matched permutation (20,000 draws) finds the 15 fusion-enriched genes **no more normal-tissue-exposed** than matched controls: 97.4 against 106.9 nTPM, ratio 0.911, p = 0.789. **The empty intersection is the base rate, not a property of fusion-driven antigens.** What this screen therefore measures is that the normal-tissue axis rejects 98.2% of everything a leukaemia cell puts on its surface - which is a statement about the filter and the surfaceome, and applies equally to lineage antigens.

Go deeper
The funnel, with what each filter removed
stepcriterionn survivingremoved
F1UniProt reviewed human, Cell membrane (SL-0039), and an extracellular TOPO_DOM or GPI anchor2,4041,630 membrane proteins with no extracellular domain - identities persisted in results.json.F1_dropped_symbols
F2bmean nTPM >= 20 across SEM + RS4;11, each line >= 101692,235
F4bparenchymal max <= 5 nTPM across normal tissues3166 (18 amber)
F3b>= 4x a B-cell-precursor leukaemia comparator and >= 2x the KMT2A-rearranged-other-partner arm03

Run the other way round, F3b alone leaves 15 genes - BST2, CD320, CLEC2B, CSPG4, ERBB2, ICAM2, IFITM2, IGFLR1, IL2RG, ITGB5, KLRK1, MCOLN2, MMP14, P2RY11, STX1A - and every one of the fifteen is killed or amber on normal tissue 1. That ordering is the result: fusion-enrichment and normal-tissue safety are each individually achievable and jointly are not.

The controls, and the first run that failed them

The pre-registration named CD19 as the positive control - it must be surface, expressed and normal-tissue clean, because blinatumomab is tolerated in infants - and FOLR1 as the filter control, because it must be killed by normal tissue 26.

The first run failed both, and the failure is recorded rather than smoothed away. CD19 was killed by the normal-tissue filter at 31.2 nTPM in small intestine, and it failed the expression filter because MV4-11, a monocytic AML line carrying t(4;11), has CD19 at 0.2 nTPM against SEM 229.6 and RS4;11 400.9 1. Both failures were diagnosed from control measurements and are written up as separate entries. After the two control-forced fixes, CD19 is clean (parenchymal max 3.7 nTPM, testis) and correctly not fusion-enriched at 2.45x a B-precursor comparator, below the 4x threshold - which is exactly right, because CD19 is a lineage antigen.

FOLR1 is killed at 2,860.9 nTPM in choroid plexus, with salivary gland 206.7, lung 182.4 and kidney 142.6 2. The filter does what it was built to do.

What a zero here does and does not license

With 0 of 169 surface, index-expressed genes clearing fusion-enrichment and the tissue filter, the true rate among that pool is below 1.78% at 95% confidence by the rule of three. That is a bound on this pool under these thresholds - not a proof that no such antigen exists.

It does not cover: proteins whose surface presentation is post-transcriptional; glycan and glycolipid antigens with no gene of their own, which this graph already records as a structural blind spot in its target-cd15-lewis-x and target-cd65s-sialylated-vim2-ganglioside entries; fusion-junction neoepitopes presented on MHC; and proteins UniProt does not yet annotate with an extracellular topological domain. 1,630 membrane proteins were excluded at F1 for exactly that last reason, and their identities are persisted so the exclusion can be re-derived rather than taken on trust 14.

The base rate, and why it deflates the headline

This beat wrote a hypothesis proposing that the disjointness was mechanistic - that the fusion's developmental output programme necessarily borrows surface genes the body uses elsewhere. The pre-written matched falsifier refuted it within the hour 8. Fusion-enriched antigens sit at a median parenchymal maximum of 97.4 nTPM against 106.9 for expression-matched controls; the ratio is 0.911 and the permutation p is 0.789.

So the honest headline is not 'KMT2A::AFF1 has no fusion-specific surface antigen'. It is:

  1. NG2/CSPG4 and LAMP5 are genuine fusion-driven surface antigens and both were already published - the corpus's claim that this disease has only borrowed and drug-induced antigens is wrong.
  2. Both fail a normal-tissue filter their discovery papers did not apply, at 124.5 nTPM in blood vessel and 297.8 in basal ganglia, and that is this beat's contribution.
  3. So does almost everything else - 98.2% of the pool - so the failure is not about them. A group choosing between a fusion-driven and a lineage antigen gains no normal-tissue advantage either way and should choose on escape resistance, where fusion-driven antigens do have a real argument.

An empty intersection of two rare properties is not evidence that they are anticorrelated, and this beat nearly wrote that it was.

What this cannot do

Expression is not surface presentation. Every expression number in this entry is mRNA. It says nothing about how much protein reaches the membrane, whether the epitope is accessible to an antibody, or whether the protein is shed. A cell line is not a patient, and the Human Protein Atlas consensus tissue panel is adult tissue while the population of concern is infants.

The index arm is three cell lines, none infant-derived, and this is overwhelmingly an infant disease. The B-lineage sub-arm that the controls forced is two lines. The patient arm reached (TARGET ALL Phase II) is keyed on MLL_STATUS, which does not name the fusion partner, so it tests KMT2A-rearranged versus not, not KMT2A::AFF1 versus not 7.

Facts and figures
scriptanalyses/kmt2a-aff1-surface-target-screen/run.py
statusrun
preregistrationanalyses/kmt2a-aff1-surface-target-screen/PREREGISTRATION.md
branch reachedC - NEITHER: surface+expressed+clean genes exist (CD19, SIT1, S1PR4) and fusion-enriched genes exist (15), and the sets are disjoint
rule of three upper bound PCT1.775
result summary0 of 169 surface, index-expressed genes are both fusion-enriched and normal-tissue clean; rule-of-three upper bound 1.78%.
base rate parenchymally clean3/169 = 1.8%
funnel
final_hits0.000final_hits: 0F4b_parenchymal_clean3.000F4b_parenchymal_clean: 3F3b_fusion_enriched15.0F3b_fusion_enriched: 15F4b_amber18.0F4b_amber: 18F2b_expressed_in_B_lineag…169.0F2b_expressed_in_B_lineage_index_arm: 169F1_surface_with_extracell…2,404F1_surface_with_extracellular_domain: 2404
inputs
[
  "UniProt REST SL-0039 (2026_03)",
  "HPA rna_celline.tsv v25.1",
  "HPA rna_tissue_consensus.tsv v25.1",
  "Cellosaurus API",
  "cBioPortal all_phase2_target_2018_pub",
  "GTEx v8 API v2"
]
funnel — the numbers behind the chart
{
  "F1_surface_with_extracellular_domain": 2404,
  "F2b_expressed_in_B_lineage_index_arm": 169,
  "F4b_parenchymal_clean": 3,
  "F4b_amber": 18,
  "F3b_fusion_enriched": 15,
  "final_hits": 0
}
fusion enriched but tissue failed
[
  "BST2",
  "CD320",
  "CLEC2B",
  "CSPG4",
  "ERBB2",
  "ICAM2",
  "IFITM2",
  "IGFLR1",
  "IL2RG",
  "ITGB5",
  "KLRK1",
  "MCOLN2",
  "MMP14",
  "P2RY11",
  "STX1A"
]
parenchymally clean but not fusion enriched
[
  "CD19",
  "SIT1",
  "S1PR4"
]
index arm
[
  "SEM CVCL_0095",
  "RS4;11 CVCL_0093",
  "MV4-11 CVCL_0064"
]
matched permutation
{
  "fusion_enriched_median_parenchymal_max": 97.4,
  "matched_control_median": 106.9,
  "ratio": 0.911,
  "permutation_p": 0.78851,
  "verdict": "the empty intersection is the base rate"
}
How this could bite us back
  • The headline reads as a discovery about fusion-driven antigens and it is not one. The matched permutation shows they behave like any other surface protein at the same expression level; what the screen really measures is that a 5 nTPM normal-tissue threshold rejects 98.2% of the leukaemia surfaceome.
  • The zero is a bound on an mRNA pool, not a proof of absence, and a clinician reading it as 'no target exists' would be over-reading it. The rule-of-three bound is stated so that the over-reading is checkable.
  • Three of the fifteen fusion-enriched genes - IL2RG, KLRK1 and ICAM2 - are immune-effector genes. Targeting them would damage the very T and NK cells a cell therapy depends on, so their fusion-enrichment is a trap rather than an opportunity, and a naive ranking would put them near the top.
  • If a future beat moves a threshold to make this screen produce a hit, the pre-registration makes that visible. The thresholds were fixed before any number was seen and the branch reached was the one the pre-registration named most likely.
  • A screen that starts from an annotation database inherits that database's blind spots. Both this screen and the published MLL-rearranged surfaceome screen rest on computational surface prediction, so a shared false negative would look like independent confirmation of an absence.
Still open
  1. Does cell-surface biotinylation plus mass spectrometry on an isogenic human KMT2A::AFF1 model return the same fifteen, or antigens invisible to mRNA? This screen cannot answer it and that is the experiment that would.
  2. Of the 1,630 membrane proteins excluded at F1 for having no annotated extracellular topological domain, how many are in fact accessible? The list is persisted; a topology-prediction pass over it is a day of work and could reopen the screen.
  3. Would the same funnel on KMT2A::MLLT3 or KMT2A::MLLT1 give a different zero, or is the disjointness a property of KMT2A-rearranged leukaemia generally rather than of AFF1?
Every source behind this entry (8)
  1. E1moderate confidence 3 of 5: One strong study, or several smaller consistent ones.dataset

    Of 2,404 reviewed human cell-membrane proteins with an extracellular topological domain, 169 are expressed at >= 20 nTPM mean and >= 10 nTPM per line in SEM and RS4;11; 15 clear both fusion-enrichment ratios; 3 have a normal parenchymal maximum <= 5 nTPM; the intersection of the last two sets is empty.

    What this evidence cannot show

    An mRNA screen over three cell lines with pre-registered but necessarily arbitrary thresholds. The thresholds were fixed before any value was seen and were not moved; a different threshold set would give a different funnel. The comparator arm was built from Cellosaurus disease labels because HPA ships no cell-line metadata file - a deviation from the pre-registration that is recorded in results.json rather than quietly substituted.

    Evidence that argues against this (1)
    • A screen bounded by mRNA cannot see an antigen created post-transcriptionally, by glycosylation, or by drug exposure. This graph already records menin-inhibitor-induced CLEC12A as exactly such an antigen, and it would be invisible to this design. Oncofusion Knowledge Graph 2026
  2. E2strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.database

    FOLR1, the CBFA2T3::GLIS2 CAR-T antigen, reaches 2,860.9 nTPM in choroid plexus, 206.7 in salivary gland, 182.4 in lung and 142.6 in kidney in the HPA consensus normal-tissue panel.

    What this evidence cannot show

    Bulk adult tissue mRNA. HPA consensus values combine HPA with external transcriptomic inputs per tissue, so a single tissue value is an aggregate rather than one measurement.

    Evidence that argues against this (1)
    • FOLR1-directed CAR-T against CBFA2T3::GLIS2 was nonetheless advanced preclinically, so high normal-tissue mRNA did not stop a credible programme. Expression is not accessible target density in vivo, and this is the clearest case in the beat where the filter may be too strict. J Clin Invest 2022
  3. E3established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.database

    SEM, RS4;11 and MV4-11 each carry a KMT2A-AFF1 gene fusion; MOLM-13, THP-1 and NOMO-1 carry KMT2A-MLLT3; KOPN-8 carries KMT2A-MLLT1; ML-2 carries KMT2A-AFDN.

    What this evidence cannot show

    Cellosaurus fusion comments are curated from primary reports and were not re-sequenced here. The run aborts on any mismatch between the fetched partner and the arm table, so no line was assigned from recall - which matters because the dispatch brief for this beat listed MOLM-13 among the lines to check for KMT2A::AFF1 and it is KMT2A-MLLT3.

    Evidence that argues against this (1)
    • Cell lines drift and are sometimes misidentified; Cellosaurus itself flags problematic and contaminated lines. None of the eight used here is so flagged, but an API record alone does not exclude the possibility. Cellosaurus (SIB) 2026
  4. E4established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.database

    The surface axis was defined as UniProt reviewed human proteins with subcellular location Cell membrane (SL-0039) plus an annotated extracellular topological domain or GPI anchor; 4,098 symbols carry the location and 2,468 also carry an extracellular domain.

    What this evidence cannot show

    UniProt release 2026_03. A TOPO_DOM annotation is frequently ECO:0000255 - sequence-analysis inference - rather than experimental, so this is a PREDICTED accessibility filter, not a measured one.

    Evidence that argues against this (1)
    • The published in silico human surfaceome uses a machine-learning classifier over a broader feature set and would not produce an identical list. It is also the instrument the already-published MLL-rearranged surfaceome screen used, so the two screens are not independent in their surface axis. Proc Natl Acad Sci U S A 2018
  5. E5established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.database

    Per-cell-line RNA nTPM for 1,206 human cell lines, including SEM, RS4;11, MV4-11, MOLM-13, THP-1, NOMO-1, KOPN-8 and ML-2, is available from the Human Protein Atlas.

    What this evidence cannot show

    HPA release v25.1. One nTPM value per line, no replicates, so no within-line variance exists and no statistical test on a single line is possible. All arm comparisons are therefore between-line means without error bars.

    Evidence that argues against this (1)
    • DepMap/CCLE expression for the same lines is an independent measurement that this beat did not cross-check; a disagreement between atlases would change the funnel and nothing here excludes one. Human Protein Atlas 2026
  6. E6strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper

    A fusion-specific surface antigen can be found for a paediatric fusion leukaemia: CBFA2T3::GLIS2 expressed in human cord-blood HSPCs yielded a fusion-specific gene library from which FOLR1 was developed as a CAR-T target.

    PMID 36136600
    What this evidence cannot show

    One fusion, one model. The model required endothelial coculture to recapitulate immunophenotype, so the antigen list is partly a property of the culture system rather than of the fusion.

    Evidence that argues against this (1)
    • FOLR1 is present in renal, pulmonary and choroid-plexus epithelium, so the same paper's target would fail the normal-tissue filter applied here - which says either that the filter is too strict or that FOLR1 carries an unretired on-target risk. Human Protein Atlas 2026
  7. E7established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.database

    The TARGET Pediatric ALL Phase II cohort labels samples with an MLL_STATUS attribute (Positive 48, Negative 1,464, Unknown 431 across all samples) and a MOLECULAR_SUBTYPE attribute with an 'MLL-Rearranged' value, but neither names the fusion partner.

    What this evidence cannot show

    Attribute counts are across all samples in the study; only a subset carry expression data (4 MLL-positive in the RNA-seq profile, 13 in the microarray profile). High-risk childhood B-ALL, not infant ALL.

    Evidence that argues against this (1)
    • Because KMT2A::AFF1 is the commonest KMT2A partner in B-ALL, MLL_STATUS-positive is enriched for it, but a partner-agnostic label cannot support any AFF1-specific claim and none is made from it here. cBioPortal for Cancer Genomics 2026
  8. E8moderate confidence 3 of 5: One strong study, or several smaller consistent ones.dataset

    Only 3 of 169 index-expressed surface genes are parenchymally clean (1.8%); an expression-matched permutation test over 20,000 draws finds the 15 fusion-enriched genes at a median parenchymal maximum of 97.4 nTPM against 106.9 for matched controls, ratio 0.911, p = 0.789.

    What this evidence cannot show

    n = 14 of 15 matched (IFITM2 dropped, reported). Matching is on expression level only. A null at this sample size excludes a 2-fold difference, not a small one.

    Evidence that argues against this (1)
    • The 1.8% base rate is itself a property of the pre-registered 5 nTPM cleanliness threshold. At 25 nTPM the clean fraction would be much larger and the base-rate argument correspondingly weaker, so both the zero and its explanation may be threshold artefacts. Oncofusion Knowledge Graph 2026
Connected to