Oncofusion

start / graph / ana-fa-aldehyde-vs-replication

Analysisunrated confidence 0 of 5: unrated

Neither rival mechanism leaves its fingerprint, and the best correlate of Fanconi dependency is menin

also called aldehyde versus replication · FA co-dependency · ALDH2 test

Two explanations were on the table for why Fanconi-repair genes matter in this leukaemia: a build-up of toxic aldehydes, or stalled DNA copying. Across 54 leukaemia cell lines neither leaves the pattern it should, and the gene whose requirement tracks Fanconi most closely is menin.

The same thing, for a scientist

The test the king log specified — regress FA Chronos on ALDH2 **expression** and KMT2A status — **cannot be run from this environment**, and that is the first result. Per-line expression is not reachable: Open Targets serves gene effect only, Cell Model Passports returned nothing, the DepMap portal is gated and the bulk matrix is a ~500 MB download. The nearest runnable test is co-dependency. **Aldehyde axis: null** (ADH5 r = 0.106, ALDH2 r = 0.051, both intervals crossing zero). **Replication axis: weakly NEGATIVE** (ATR −0.289, CHEK1 −0.317), the opposite sign to the transcription-replication-conflict prediction. **MEN1 is the strongest correlate at r = 0.670**, and it survives deleting the KMT2A-rearranged lines entirely: **0.486 [0.218, 0.686] across the 43 non-KMT2A leukaemia lines**, against an ACTB control of 0.096.

Go deeper
The specified test could not be run, and substituting quietly would have been worse

Last cycle's king log said this goes first: regress FA Chronos on ALDH2 expression and KMT2A status, because UBE2T and FANCL are already published as selective AML dependencies through synthetic lethality with methylation-silenced ALDH2.

Per-cell-line expression is not obtainable here. Open Targets exposes DepMap depMapEssentiality but no expression field; the Cell Model Passports API returned an empty body; the DepMap portal sits behind a verification page; and the bulk expression matrix is a ~500 MB download this project has no reason to hold. The honest move is to say so and run the nearest test that the available data supports, marked as what it is.

What each mechanism predicts, and what was found

Both rivals make a co-dependency prediction, which needs only gene effect.

geneaxismean effectr with FA module95% CIr with KMT2A-r lines REMOVED
ADH5aldehyde−0.0750.106[−0.166, 0.364]0.160
ALDH2aldehyde+0.0620.051[−0.220, 0.315]0.083
ATRreplication−0.965−0.289[−0.517, −0.024]−0.227
CHEK1replication−1.683−0.317[−0.539, −0.054]−0.355
RAD51replication−1.1350.035[−0.235, 0.300]−0.000
CLSPNreplication−0.5620.210[−0.062, 0.452]0.313
TIMELESSreplication−1.491−0.107[−0.364, 0.166]−0.134
FANCD2substrate−0.2920.714[0.552, 0.824]0.738
MEN1control−0.4250.670[0.490, 0.795]0.486
ACTBcontrol−0.6550.067[−0.205, 0.329]0.096

The aldehyde signature is absent. If the module were required because endogenous aldehydes are accumulating, the other arm of aldehyde clearance should matter in the same lines. ADH5 does not.

The replication signature is absent and points the wrong way. ATR and CHEK1 correlate NEGATIVELY: the lines that need the FA module most need the checkpoint least. Both intervals exclude zero, barely. That is a trade-off pattern, not the reinforcement a shared fork-stress mechanism predicts.

FANCD2 co-varies strongly and is much less required — r = 0.714 at a mean of −0.292 against module members two to three times deeper. The Fanconi beat concluded the module must act on something other than FANCD2; this refines rather than confirms that. They move together across lines and differ in amplitude, which is what a shared pathway with a non-rate-limiting step looks like.

The result nobody was looking for

MEN1 is the strongest correlate of FA-module dependency in the panel, and it is not the KMT2A block. r = 0.670 across all 54 leukaemia lines; 0.486 [0.218, 0.686] across the 43 lines with the KMT2A-rearranged ones deleted. The ACTB control sits at 0.067 and 0.096, so this is not generic coupling between any two dependencies.

The FA module is also genuinely KMT2A-enriched in the same data — mean −0.555 in the 11 rearranged lines against −0.202 in the other 43, and 7 of the 10 most FA-dependent lines are KMT2A-rearranged against a base rate of 11 in 54. Both things are true at once: the module is enriched in KMT2A-rearranged lines, and its correlation with menin does not depend on them.

The obvious sceptical reading is that menin is required across many leukaemias and something third — proliferative rate, screen quality, a shared lineage composition — drives both. ACTB argues against the crudest version of that and does not exclude it. This is filed as a correlation worth explaining, not a mechanism.

The subset column is noise and is printed anyway

The results also report r within the 11 KMT2A-rearranged lines alone, and it should not be read. ACTB gives −0.624 there against 0.096 outside; RAD51 gives 0.685 against −0.000; ALDH2 gives −0.347 against 0.083. At n = 11 a Fisher interval is roughly ±0.55 wide, so those numbers are sampling and nothing else. It is printed because this project has twice quoted a subset statistic that looked like a finding, and an obviously unstable column sitting next to a stable one is a better guard than a footnote.

Still open
  1. Locate the DepMap Omics figshare record and run the test as originally specified. The 'DepMap <release> Public' articles hold only Achilles/CRISPR files; expression is elsewhere, and the figshare search API did not find it. The portal download page would give the direct URL if someone can get past the verification page.
  2. Test the ATR/CHEK1 anti-correlation directly: is it a trade-off between two routes to tolerating fork stress, or a normalisation artefact of two broadly required gene sets?
  3. The MEN1-FA relationship is carried by myeloid lines and is not significant in lymphoid lines once the index arm is removed (n=17). Since the index disease is B-lymphoid, that stratum needs more lines before the correlation is carried into any claim about infant leukaemia.
Every source behind this entry (5)
  1. E1established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.analysis

    The discriminating test recorded in the cycle-47 king log — regress FA Chronos on ALDH2 expression and KMT2A status — cannot be executed from this environment. Open Targets exposes DepMap gene effect without expression, the Cell Model Passports API returned an empty body, the DepMap portal is behind a verification page, and the bulk expression matrix is a ~500 MB download.

    What this evidence cannot show

    Run by the coordinator on 2026-09-22. FA-module score per line = mean line-normalised Chronos of UBE2T, FANCL, FAAP100, FANCA, FANCF, FANCG, FANCI and BRIP1, with offsets from an independent 1,200-gene random panel so the compared genes never set their own baseline. Index arm resolved from Cellosaurus by DepMap accession. 54 leukaemia lines, 11 of them KMT2A-rearranged. Pearson r with Fisher 95% intervals.

    Evidence that argues against this (2)
    • Unreachable from here is not unreachable, and the download was attempted rather than assumed. The figshare records titled 'DepMap <release> Public' (24Q2, 24Q4, 23Q4) each hold ten files and all of them are Achilles/CRISPR — AvanaLogfoldChange, AvanaRawReadcounts, CRISPRGeneEffect, guide maps and QC. The Omics expression matrix sits in a separate figshare record that the search API did not surface for any query tried, and the one candidate mirror that did surface turned out to be a hepatobiliary subset — its columns are CCLP1, HEPG2, HUH7, SNU-series, with no leukaemia lines at all. So the correct statement is that the matrix is public and was not located in a bounded attempt, not that it is out of reach and not that nobody looked. figshare / Broad Institute DepMap 2026
    • A co-dependency test is a weaker instrument than the regression it replaces, and not only because of the data. Vakoc's mechanism is ALDH2 SILENCING, so ALDH2's own gene effect is expected to be flat whichever model is right — there is nothing left to cut in a silenced gene. ADH5 dependency proxies the pathway being loaded, not ALDH2 being off. Cancer Discovery 2021
  2. E2moderate confidence 3 of 5: One strong study, or several smaller consistent ones.analysis

    Across 54 leukaemia lines, FA-module dependency shows no co-dependency with the aldehyde-clearance axis (ADH5 r = 0.106 [−0.166, 0.364]; ALDH2 r = 0.051 [−0.220, 0.315]) and a weakly NEGATIVE relationship with the replication checkpoint (ATR r = −0.289 [−0.517, −0.024]; CHEK1 r = −0.317 [−0.539, −0.054]), the opposite sign to a shared fork-stress mechanism.

    What this evidence cannot show

    Run by the coordinator on 2026-09-22. FA-module score per line = mean line-normalised Chronos of UBE2T, FANCL, FAAP100, FANCA, FANCF, FANCG, FANCI and BRIP1, with offsets from an independent 1,200-gene random panel so the compared genes never set their own baseline. Index arm resolved from Cellosaurus by DepMap accession. 54 leukaemia lines, 11 of them KMT2A-rearranged. Pearson r with Fisher 95% intervals.

    Evidence that argues against this (2)
    • A null across 54 lines is a weak null. The Fisher interval at n = 54 is about ±0.27 wide, so a true correlation of 0.25 would frequently be reported as non-significant here. Absence of a signature at this power removes a prediction; it does not refute the model that made it. Oncofusion knowledge graph (internal analysis) 2026
    • ATR, CHEK1, RAD51 and TIMELESS are all broadly required — mean effects of −0.96 to −1.68 — which compresses their dynamic range and makes the replication axis systematically harder to detect than the aldehyde axis, whose genes sit near zero. The two axes were not tested on equal footing. Oncofusion knowledge graph (internal analysis) 2026
  3. E3moderate confidence 3 of 5: One strong study, or several smaller consistent ones.analysis

    MEN1 is the strongest correlate of FA-module dependency in the panel at r = 0.670 [0.490, 0.795] across 54 leukaemia lines, and the relationship survives deletion of the KMT2A-rearranged lines: r = 0.486 [0.218, 0.686] across the remaining 43, against an ACTB control of 0.096. In the same data the FA module is KMT2A-enriched, at −0.555 in 11 rearranged lines against −0.202 in 43 others, with 7 of the 10 most FA-dependent lines KMT2A-rearranged.

    What this evidence cannot show

    Run by the coordinator on 2026-09-22. FA-module score per line = mean line-normalised Chronos of UBE2T, FANCL, FAAP100, FANCA, FANCF, FANCG, FANCI and BRIP1, with offsets from an independent 1,200-gene random panel so the compared genes never set their own baseline. Index arm resolved from Cellosaurus by DepMap accession. 54 leukaemia lines, 11 of them KMT2A-rearranged. Pearson r with Fisher 95% intervals.

    Evidence that argues against this (2)
    • Menin is required across many leukaemias, so a third variable shared by both — proliferative rate, screen quality, lineage composition of the panel — remains a live explanation. The ACTB control excludes only the crudest form of it, because ACTB is pan-essential and near its floor, which is exactly where a generic coupling would be hardest to see. Oncofusion knowledge graph (internal analysis) 2026
    • A correlation this size invites a mechanistic story and there is none. If the FA requirement were downstream of the fusion's elongation programme, menin inhibition would RELIEVE it and the two would be antagonistic, which predicts a NEGATIVE correlation, not a positive one. The result does not fit the hypothesis the Fanconi beat proposed for it. Oncofusion knowledge graph (internal analysis) 2026
  4. E4moderate confidence 3 of 5: One strong study, or several smaller consistent ones.analysis

    FANCD2 correlates with the FA-module score at r = 0.714 [0.552, 0.824] while being substantially less required than the module members (mean −0.292 against module means two to three times deeper).

    What this evidence cannot show

    Run by the coordinator on 2026-09-22. FA-module score per line = mean line-normalised Chronos of UBE2T, FANCL, FAAP100, FANCA, FANCF, FANCG, FANCI and BRIP1, with offsets from an independent 1,200-gene random panel so the compared genes never set their own baseline. Index arm resolved from Cellosaurus by DepMap accession. 54 leukaemia lines, 11 of them KMT2A-rearranged. Pearson r with Fisher 95% intervals.

    Evidence that argues against this (2)
    • This refines rather than confirms the Fanconi beat's conclusion that the module must act on something other than FANCD2. Strong co-variation with lower amplitude is what a shared pathway with a non-rate-limiting step looks like, so 'acts on something else' and 'acts on FANCD2, which is not limiting' both fit. Oncofusion knowledge graph (internal analysis) 2026
    • FANCD2 and FANCI are an obligate heterodimer and FANCI scored strongly in the same beat. A correlation between a module and one half of its own substrate dimer is close to a positive control for the score being coherent, which makes it weak evidence about mechanism. Oncofusion knowledge graph (internal analysis) 2026
  5. E5moderate confidence 3 of 5: One strong study, or several smaller consistent ones.analysis

    The MEN1 correlation survives both obvious categorical confounds. Residualising FA-module score and MEN1 effect within each (lineage x KMT2A status) cell and correlating the residuals gives partial r = 0.506 [0.274, 0.683] over 53 lines, against an ACTB control of −0.034 [−0.302, 0.238] under the identical treatment. By stratum: myeloid r = 0.737 [0.523, 0.864] (n=32), lymphoid r = 0.556 [0.164, 0.797] (n=21); with KMT2A-rearranged lines removed, myeloid r = 0.577 [0.236, 0.792] (n=25) and lymphoid r = 0.286 [−0.225, 0.674] (n=17).

    What this evidence cannot show

    Computed on the same 54-line panel. The lymphoid stratum with KMT2A-rearranged lines removed has 17 lines and its interval crosses zero, so the relationship is demonstrated in myeloid lines and only suggested in lymphoid ones.

    Evidence that argues against this (2)
    • Lineage and KMT2A status are the confounds that could be removed because they are recorded. The ones that matter most are not: proliferation rate, screen quality, passage history and culture medium are absent from this dataset, and any of them could couple two dependencies without a shared mechanism. A flat ACTB control makes the crudest version unlikely and leaves the subtle version untouched. Oncofusion knowledge graph (internal analysis) 2026
    • The effect is carried by the myeloid lines and the lymphoid arm does not reach significance once the index lines are removed. Infant KMT2A::AFF1 disease is B-lymphoid, so the stratum most relevant to this project is the one where the evidence is weakest — the opposite of the shape a finding should have before it is carried into the disease. Oncofusion knowledge graph (internal analysis) 2026
Connected to