The RNA layer scored for KMT2A-selectivity across 79 genes, lineage-split — PTBP1 is the only menin-shaped hit, and the IGF2BP family shows no co-dependency

PTBP1 is the only RNA-layer gene to reach the 0.0th percentile of the null - the percentile MEN1 and KMT2A occupy. It is -0.247 in the four B-lymphoid index lines and -0.382 in the seven myeloid ones, concordant in sign, with 50.7% of all lines depleted so it is well clear of floor compression 1.

It was not in the first version of this panel. The panel was assembled from m6A, splicing-factor, helicase, decay and 3'-end vocabulary; PTBP1 entered only after the literature route found a domain-focused CRISPR screen of 341 RNA-associated proteins that named PTBP1 and RBM15 as AML dependencies with a bias toward the KMT2A-rearranged subtype 5. On being added it outranked every gene that had been there from the start.

Two things follow, and the second is the more useful.

The ranking is not blind and must not be quoted as though a scan discovered PTBP1. What this analysis independently contributes is that the published bias reproduces in a different dataset with a lineage split the screen did not perform.

And the miss itself is the finding. A gene-list scan finds members of the categories the curator thought of; a domain-tiled screen defines membership by protein structure. This beat was dispatched to correct enumeration-from-complex-membership and reproduced the same error one level up, as enumeration-from-functional-category.

Below PTBP1 the table is modest. HNRNPA2B1 (-0.122 / -0.275, 1.7th percentile), SYNCRIP (-0.116 / -0.242, 2.4th) and ZC3H13 (-0.182 / -0.157, 2.8th) are the only other genes that are simultaneously lineage-concordant, not floor-compressed, and inside the 3rd percentile. The controls remain a different order of magnitude: MEN1 -0.871 / -0.863 and KMT2A -0.436 / -0.601, both 0.0th 1.

So the honest headline outside PTBP1 is a negative. The beat's premise - that a post-transcriptional layer is under-enumerated in this graph - is right about the literature and is not generally rescued by the dependency data. This graph should not acquire an RNA-layer target on the strength of a Chronos difference alone.

That negative has a hard limit the results file states and this entry repeats: Chronos measures two-to-three-week proliferation dropout in lines already selected for growth in culture. It cannot see a dependency acting on leukaemia initiation or self-renewal, which is exactly the phenotype the published IGF2BP3, MSI2, YTHDF2 and RBM15B work reports 4.

KHDRBS1/2/3 entered the panel only after the literature route found that they translationally repress the MLL-AF4 transcript itself 3. All three come back near zero and positive in the index arm: KHDRBS1 +0.087 / +0.059 (89.3rd null percentile), KHDRBS2 +0.068 / +0.041, KHDRBS3 +0.044 / +0.043, with 4.2%, 0.1% and 0.2% of lines depleted 1.

That is not a failure to replicate; it is what the published mechanism predicts. A dropout screen scores loss of fitness on knockout, so a gene whose product restrains the oncogene should score at or above zero. The KHDRBS rows demonstrate that a dependency screen cannot enumerate negative regulators in either direction, and this beat found them only because the literature route ran independently.

Every conclusion in this graph of the form 'flat DepMap, therefore dispensable' inherits that blind spot.

Thirty of the 79 panel genes have lymphoid and myeloid index-arm differences of opposite sign. For every one of them the pooled eleven-line number is an average of two contradictory results and must not be quoted 1.

Two matter most.

METTL3 is +0.097 lymphoid and -0.229 myeloid. Pooled, it would have read as a modest fusion-selective requirement and been quoted in a B-lymphoid entry. The published METTL3 leukaemia work is entirely myeloid 6, so the split reproduces the literature's own boundary rather than crossing it.

MSI2 runs the other way: -0.284 lymphoid, +0.154 myeloid, against a literature in which MSI2 is a myeloid MLL leukaemia-stem-cell dependency 4. The myeloid half disagrees in sign with the mouse genetics that motivated it.

A 30-of-79 discordance rate is itself a statement about the instrument. With four lymphoid and seven myeloid lines against a null SD of 0.075, sign agreement between halves is not the default, so concordance is the filter that should be applied before any single-gene claim - and PTBP1, HNRNPA2B1, SYNCRIP and ZC3H13 are the genes that pass it.

Pairwise Chronos correlation is 0.066 (1|2), 0.105 (1|3), 0.078 (2|3) across all 1,208 lines, turning negative inside the 54 leukaemia lines: -0.275, -0.063, -0.123. Ninety-eight lines are depleted for at least one member and four for two or more, against 2.52 expected under independence 2.

There is no positive co-dependency signal. The design cannot answer the question that was asked. Under genuine buffering every single knockout is near-neutral by construction and the profiles carry no signal at all - indistinguishable from the family not mattering. A positive correlation would have indicated shared dependency context, not redundancy; its absence is not evidence against redundancy. The decisive experiment is a combinatorial double and triple knockout and DepMap contains none 2.

What the analysis can add is a direction. In B-ALL, IGF2BP2 and IGF2BP3 are the pair over-expressed in MLL/AF4-positive disease while IGF2BP1 marks ETV6-RUNX1 7 - and here IGF2BP2 (-0.123) and IGF2BP3 (-0.135) are the two members with negative B-lymphoid differences while IGF2BP1 is +0.095. Two datasets with no shared machinery pick out the same two of three.

analyses/rna-layer-kmt2a-selectivity/run.py, seed 20260920. Gene effects from DepMap CRISPR (Chronos) through the Open Targets GraphQL API; that API does not state the DepMap release version, so the exact dataset cannot be pinned. Symbol-to-Ensembl mapping comes from the HGNC complete set rather than any hard-coded identifier, because this project has previously resolved an Ensembl id to the wrong gene. The index arm comes from kmt2a_rearranged_lines() in analyses/depmap-line-normalised-specificity/run.py, imported rather than duplicated.

The eleven index lines are MV4-11, OCI-AML-2, THP-1, NOMO-1, MOLM-13, MOLM-14 and SHI-1 (myeloid) and SEM, RS4;11, SEMK2 and HB11;19 (lymphoid). Three of the four lymphoid lines - SEM, RS4;11 and SEMK2 - are KMT2A::AFF1, so the lymphoid column is close to a t(4;11)-specific readout on four points. Thirty-one genes are flagged floor-compressed at 90-100% of lines depleted; YTHDF3 resolved in HGNC but returned too few comparator lines and was dropped.

The panel was extended twice during the beat, each time after the literature route found something the vocabulary had missed: KHDRBS1/2/3 and PTBP1/2/3. Both extensions are recorded as comments in the panel definition so the provenance of the ranking stays visible.