The MHC was never invisible - the resolver was wrong - and once the classical chains are joined the class I machinery is bivalent, not unmarked
also called HGNC primary id · MHC rescue · class I bivalency · multi-mapping exposure · rule 78 remedy
I reported that a family of immune genes could not be examined because their names match several database entries, and drew a rule from it - the rule was half wrong, because a standard registry names one official entry per gene and brings them all back. Redoing the test shows the fragment-display machinery carries the activating and the silencing mark at once, which is a weak lead rather than a result.
The same thing, for a scientist
This morning's entry concluded that the GO antigen-presentation sets show **no skew** across Zhou's menin/H2AK119ub groups, and that the MHC is structurally unjoinable. **The second claim was wrong and it caused the first.** **MyGene returns every Ensembl gene id for a symbol, including alternative MHC haplotype scaffolds**, so an exact-match resolver correctly refuses. **HGNC assigns exactly one primary `ensembl_gene_id` per symbol**, on the primary assembly - which is what Zhou's table uses. Resolving through HGNC rescues **8 of the class I set**: B2M, HLA-A, HLA-B, HLA-C, HLA-E, TAP1, TAP2, TAPBP. **Re-run with them in, class I is no longer flat**: the bivalent (Menin + H2AK119ub) group holds **9 observed against 4.5 expected, nominal two-sided permutation p = 0.047**, and the members are **B2M, HLA-A, HLA-B, HLA-C, TAP1, TAPBP** plus IDE, LNPEP and PSMB10. Class II remains flat (10 genes, all p >= 0.93). **It does not survive correction and the genes are not independent** - see below. This is a lead, graded moderate.
Why it is a lead and not a result
Two problems, and both are large:
1. Six tests were run (two gene sets x three groups). Bonferroni takes p = 0.047 to p = 0.284. On a corrected reading there is nothing here.
2. The genes are not independent observations. Of the nine bivalent members, five sit in or beside the MHC:
| gene | cytoband |
|---|---|
| HLA-A | 6p22.1 |
| HLA-B | 6p21.33 |
| HLA-C | 6p21.33 |
| TAP1 | 6p21.32 |
| TAPBP | 6p21.32 |
A permutation that shuffles gene labels assumes genes are exchangeable. Genes inside one chromatin domain are not. The honest restatement is not 'nine genes are bivalent' but 'the MHC region is bivalent, and so are B2M, IDE, LNPEP and PSMB10' - which is roughly five independent observations, not nine, and the test was never designed for that.
So the defensible claim is a domain-level one: in this dataset the class I locus carries menin and H2AK119ub together. That is one fact about one region, stated at moderate confidence.
What the domain-level claim predicts, and it matches the published functional data
Bivalency is the configuration that predicts drug-releasable expression. A promoter carrying both an activating and a repressive mark is poised, and removing the repressive input raises it.
So if the class I locus is menin-bound and H2AK119ub-marked, menin inhibition should raise class I - which is exactly what the functional literature reports: VTP50469 raises cell-surface MHC class I dose-dependently, a chemically distinct menin inhibitor reproduces it, and bivalency at MHC-I promoters was confirmed by ChIP-reChIP in K-562.
That is an independent line of evidence arriving at the same place, and it is worth more than the p-value. The occupancy table and the pharmacology were produced by different groups, in different cell lines, for different purposes.
But note what it does NOT rescue. This morning's separate finding stands unchanged: NLRC5 is menin-bound and CIITA is not, and the class II machinery shows no skew even with B2M restored. The two arms still look different; what has changed is that the class I arm now has a chromatin signature consistent with its pharmacology, and the class II arm still does not.
How exposed is a symbol-to-Ensembl join, really
Measured rather than assumed, and it is reassuring:
| set | n | multi-mapping |
|---|---|---|
| this graph's gene entries | 241 | 22 (9.1%) |
| random protein-coding background | 400 | 38 (9.5%) |
This graph is not unusually exposed - about one protein-coding gene in eleven returns more than one Ensembl id, here and everywhere. The graph's 22 concentrate at 6p21 (7), with 22q11 and 6p22 next, which is the alt-haplotype geography rather than anything about this project.
The remedy is cheap and should be standard: resolve through the HGNC primary ensembl_gene_id when joining against a primary-assembly table, and reserve refusal for when no primary id exists. Refusing wholesale discards ~9% of the genome silently.
class I
{
"n_in_zhou_before": 16,
"n_in_zhou_after": 24,
"rescued": [
"B2M",
"HLA-A",
"HLA-B",
"HLA-C",
"HLA-E",
"TAP1",
"TAP2",
"TAPBP"
],
"observed": {
"H2AK119ub solo": 9,
"Menin/H2AK119ub shared": 9,
"Menin solo": 6
},
"p_nominal": {
"Menin solo": 0.3186,
"Menin/H2AK119ub shared": 0.0473,
"H2AK119ub solo": 0.6376
},
"bonferroni_6_tests": 0.284
}class II
{
"n_in_zhou_after": 10,
"observed": {
"Menin solo": 3,
"Menin/H2AK119ub shared": 2,
"H2AK119ub solo": 5
}
}multi mapping — the numbers behind the chart
{
"graph_n": 241,
"graph_multi": 22,
"background_n": 400,
"background_multi": 38
}bivalent members in mhc
[ "HLA-A", "HLA-B", "HLA-C", "TAP1", "TAPBP" ]
- This corrects an entry I wrote the same day, and the correction runs in the direction that makes the earlier caveat look prescient rather than wrong: the caveat said the null was about what remained after the interesting genes were dropped, and it was. **What was wrong was the conclusion that they HAD to be dropped.**
- Do not quote 'the class I machinery is bivalent, p=0.047' anywhere. Six tests were run, correction kills it, and five of the nine members share one chromatin domain. The claim that survives is that the MHC region carries both marks in one adult myeloid cell line, and that this is consistent with - not evidence for - the published pharmacology.
- Is the class I locus bivalent in a B-lineage KMT2A::AFF1 system, where the therapeutic question actually lives? Everything here is adult myeloid KMT2A::AFDN.
- Does an allele-aware or coordinate-interval join over chr6:29-33 Mb give a different answer from the HGNC primary id, which collapses the locus to one representative per gene?
- How many of this project's 88 analyses joined on a symbol and silently dropped ~9% of their input? None of them reported a dropped count.
- E1established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.analysis
Resolving gene symbols through the HGNC primary ensembl_gene_id rather than through a resolver that returns all Ensembl ids recovers 8 of the 44 GO MHC class I genes that were otherwise dropped for multi-mapping - B2M, HLA-A, HLA-B, HLA-C, HLA-E, TAP1, TAP2 and TAPBP - raising the number joinable against Zhou's called set from 16 to 24.
mhc-rescued-by-hgnc-primary-id: HGNC primary Ensembl ids used to join GO antigen-presentation sets against Zhou supplemental Table 4 · Oncofusion knowledge graph (internal analysis) 2026What this evidence cannot show
Run 2026-09-20 against the cached HGNC complete set and the live QuickGO and MyGene APIs. **HGNC's primary id is an editorial choice, not a measurement** - it is the correct id for joining to a primary-assembly table and it deliberately discards alternative-haplotype sequence, which for HLA is the biologically interesting part. So this buys comparability at the cost of the locus's actual diversity, and any allele-level question still cannot be asked this way.
Evidence that argues against this (1)
- Refusing a multi-mapping symbol is not a bug - rule 62 requires it, and this project has been burned by resolvers that guessed. The correct reading is that refusal and primary-id resolution answer different questions, and the earlier entry was wrong to treat refusal as the only safe option rather than to look for an authoritative single id. Zhou X et al. 2024
- E2moderate confidence 3 of 5: One strong study, or several smaller consistent ones.paper
In OCI-AML2, an adult KMT2A::AFDN AML line, the GO MHC class I gene set joined through HGNC primary ids shows 9 of 24 members in the bivalent menin/H2AK119ub group against 4.5 expected (nominal two-sided permutation p = 0.047 over 20,000 shuffles), with the bivalent members including B2M, HLA-A, HLA-B, HLA-C, TAP1 and TAPBP, while the class II set shows no skew in any group.
Epigenetic regulation of noncanonical menin targets modulates menin inhibitor response in acute myeloid leukemia · Zhou X, Zhang L, Aryal S, et al. · Blood 2024What this evidence cannot show
**Does not survive multiple-testing correction**: six tests were run and Bonferroni takes p = 0.047 to 0.284. **And the members are not independent** - HLA-A, HLA-B, HLA-C, TAP1 and TAPBP all lie within 6p21-6p22, one chromatin neighbourhood, so a label-shuffling permutation that assumes exchangeable genes overstates the evidence. The defensible statement is domain-level: the class I locus carries both marks. **Moderate for that; weak for any per-gene claim.** Occupancy is not regulation, the table has no condition column, and the system is adult myeloid KMT2A::AFDN rather than infant B-lineage KMT2A::AFF1 (rule 73).
Evidence that argues against this (1)
- The same paper that demonstrates menin-inhibitor-induced MHC class I also reports that reduced menin occupancy genome-wide produced comparatively few expression changes - so a bivalent occupancy signature at the class I locus does not establish that menin is the regulator there. It is equally consistent with menin being present and inert, which is the null this analysis cannot exclude. Sparbier CE et al. 2023
- E3strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.analysis
Multi-mapping of gene symbols to several Ensembl gene ids affects 22 of 241 (9.1%) of this knowledge graph's gene entries and 38 of 400 (9.5%) of a random protein-coding background sample, concentrated at 6p21, so the phenomenon is a general property of the human gene set rather than a peculiarity of this corpus.
Multi-mapping exposure of gene symbols to Ensembl gene ids, this graph versus a random protein-coding background · Oncofusion knowledge graph (internal analysis) 2026What this evidence cannot show
Background is a single random sample of 400 HGNC protein-coding symbols with a fixed seed; no confidence interval was computed and the two proportions were not formally compared, so 'not unusually exposed' means the point estimates are close rather than that equality was tested. Counts depend on MyGene's current content.
Evidence that argues against this (1)
- A matched overall rate can still hide a matched-where-it-matters problem. This graph's 22 multi-mappers are concentrated at 6p21, and this graph asks immunological questions - so an equal GLOBAL rate is compatible with much worse exposure on the specific questions being asked, which is exactly what happened this morning. Zhou X et al. 2024
CORRECTS ITS METHOD AND ITS RESULT: the MHC is joinable through HGNC primary ids, and the class I null does not survive the rescue.
BIVALENT IN ZHOU'S SET once joined properly - menin and H2AK119ub together at the class I locus, which is the configuration that predicts drug-releasable expression.
BIVALENT, and rescued from the drop list - B2M sits at 15q21, outside the MHC, so it is one of the few independent members of the signal.
AN ORTHOGONAL LINE TO THE SAME PLACE: a published occupancy table, read independently, puts the class I locus in the bivalent group.
UNCHANGED BY THE RESCUE: NLRC5 always resolved 1:1 and remains menin-bound.
ALSO UNCHANGED: CIITA remains H2AK119ub-only, and the class II set stays flat even with B2M restored.
THE JOIN THIS REPAIRS: that entry's 22 unresolved symbols were mostly this same alt-haplotype problem.
GIVES THAT RESULT ITS BACKGROUND: the bivalent group carries no immune enrichment at all, so the class I locus is not one instance of a general pattern.