Oncofusion

start / graph / ana-q193-planned-re-search-of-pxd009338-for-the-nurd-deacetylase-module

Analysismoderate confidence 3 of 5: One strong study, or several smaller consistent ones.Q193-the-nurd-subunits

PLANNED: re-search the public PXD009338 raw files for the twelve corepressor subunits the published table truncated away - one compute job that closes twelve measured-not-reported absences and adjudicates the module-split hypothesis

also called the PXD009338 re-search · Q193 planned analysis

The published MLL-fusion pulldown only listed its top three hundred hits per experiment, so a dozen proteins this beat cares about were measured and never written down. The raw machine files are public, so the missing numbers can be recovered with computing time and no laboratory work at all.

The same thing, for a scientist

The Skucha 2018 seven-fusion AP-MS reports the ~300 most significant preys per bait by combined p-value rather than applying a threshold [E1]; 291 rows are retained for the MLL-AF4 (= KMT2A::AFF1) column. Twelve NuRD and Sin3 subunits this beat measured as absent from the 960-protein network are therefore `measured-not-reported`, not zero: MTA1, MTA2, MTA3, MBD3, CHD5, HDAC1, HDAC2, RBBP4, RBBP7, SIN3B, SAP30, SAP18 [E2]. The raw files are deposited at ProteomeXchange **PXD009338** [E3]. This analysis is a re-search of those files for the corresponding peptides across all seven baits and the negative control, reporting per-protein spectral counts whether positive or negative. It is the decisive test between the two explanations of the module split - a real property of the association versus a bait-specificity filter removing abundant promiscuous proteins - and it needs compute, not wet work. **Status: planned. Not run by this beat, which did not have the compute budget for a raw MS re-search within its cycle.** The entry exists so that the absence is closable by whoever picks it up rather than being rediscovered.

Go deeper
The design, and the controls that make its output readable

Input: PXD009338 raw files (ProteomeXchange), all seven Strep-HA MLL-fusion baits plus the negative control, both replicates each.

Search: the same human database and, as far as the deposit's metadata allows, the same search parameters the authors used - this must be a REPRODUCTION before it can be a difference. This project has been burned once by rebuilding a tool to audit it, dropping one of its filters, and publishing a defect rate nearly double the truth.

First output, and it is a gate, not a result: the spectral counts for CHD4, CHD3, GATAD2A, GATAD2B and MBD2 in the columns where the published table reports them. If those do not reproduce the published counts within a stated tolerance, the re-search is wrong and nothing else it says may be used. CHD4 in MLL-AF4 is 7|9 with a 2|2 negative control; that is the calibration point.

Second output: per-protein spectral counts for the twelve truncated subunits in every bait column and the negative control.

Third output: for each of the twelve, whether the count is zero, at negative-control level, or substantial.

What each answer means. Counts near CHD4's in MLL-AF4 would mean the deacetylase module WAS there and the top-300 filter removed it - the module split becomes an artefact and hyp-mll-fusions-engage-the-nurd-remodelling-arm-without-its-deacetylase-arm is refuted. Counts at zero or at control level would mean the absence is real in this dataset and the hypothesis survives its first test. An intermediate result is the likeliest and must be reported as intermediate, not rounded toward whichever conclusion is tidier.

Why this analysis and not a wet experiment first

Three reasons, and the third is the one that generalises.

  1. It is the cheapest thing that can change a conclusion. No cells, no antibodies, no mouse.
  2. It closes twelve absences at once, each of which currently reads as a gap and is in fact an unopened file. This corpus has a standing observation that an absence closable by compute and left open is worse than one that needs a mouse, because it looks like a finding.
  3. It is the only test that separates the hypothesis from its artefact explanation using the SAME data. A co-IP in a different cell line answers a different question and would leave the interpretation of the published table exactly where it is - which matters because that table is the source of a large fraction of this graph's interactome claims.
Facts and figures
scriptanalyses/q193-nurd-module-split/ (the scoring half is run; the re-search half is NOT written)
statusplanned
result summaryNOT RUN. Planned design, calibration gate and decision rule are specified so that whoever runs it cannot report a result without first reproducing CHD4's published counts.
calibration gateReproduce CHD4 at 7|9 in MLL-AF4 with a 2|2 negative control, and CHD3, GATAD2A, GATAD2B and MBD2 in the bait columns where the published table reports them, within a stated tolerance. IF THE GATE FAILS, REPORT THE FAILURE AND STOP - do not report the target counts. A reimplementation must reproduce before it can differ.
absence classmeasured-not-analysed
absence class noteThis entry is the reason twelve absences in this beat are classed measured-not-analysed rather than never-measured: the data exist, are public, and the analysis is specified. If the PRIDE deposit turns out to hold only processed search results rather than raw files, the class reverts to measured-not-reported and this entry must be amended to say so.
what it would costCompute and a proteomics search pipeline. No cells, no antibodies, no animals. Days, not years.
relationship to existing entriesThe planned counterpart to ana-the-mll-fusion-interactome-recovers-corepressor-scaffolds-without-their-hdac1-2-catalytic-core, which is run and which this analysis would adjudicate. It is the named closing step for absence-no-experiment-has-ever-asked-whether-a-nurd-scaffold-subunit-binds-a-kmt2a-fusion and the first falsifier of hyp-mll-fusions-engage-the-nurd-remodelling-arm-without-its-deacetylase-arm. It does not duplicate any existing PXD009338 work in this corpus: an earlier beat found the deposit while bounding a fusion-junction-peptide absence, and no beat has searched it for these proteins.
inputs
[
  "ProteomeXchange PXD009338 raw files - NOT downloaded by this beat",
  "analyses/q193-nurd-module-split/results.json - the absent-member list and the calibration counts"
]
targets
[
  "MTA1",
  "MTA2",
  "MTA3",
  "MBD3",
  "CHD5",
  "HDAC1",
  "HDAC2",
  "RBBP4",
  "RBBP7",
  "SIN3B",
  "SAP30",
  "SAP18"
]
decision rule
{
  "hypothesis_refuted": "MTA2 or HDAC1 in the MLL-AF4 pulldown at a spectral count within 2-fold of CHD4's, with negative-control counts at or below CHD4's 2|2 - i.e. the module was present and rank-truncated.",
  "hypothesis_survives": "The twelve targets at zero or at negative-control level in the MLL-AF4 pulldown while the calibration set reproduces.",
  "intermediate": "Anything between. REPORT AS INTERMEDIATE. The likeliest outcome and the one most at risk of being rounded toward whichever conclusion is tidier."
}
measurements
[
  {
    "quantity": "absences_closable",
    "subject": "NuRD and Sin3 subunits absent from the Skucha table whose counts this analysis would recover",
    "value": 12,
    "units": "proteins",
    "n": 12,
    "universe": "the 14 canonical NuRD subunits plus the 17 canonical Sin3/HDAC subunits scored against the 960-protein network, deduplicated",
    "null": null,
    "note": "MTA1, MTA2, MTA3, MBD3, CHD5, HDAC1, HDAC2, RBBP4, RBBP7, SIN3B, SAP30, SAP18 - identities persisted, not only counted"
  }
]
How this could bite us back
  • The likeliest outcome is intermediate counts, and an intermediate result reported as a clean one would propagate straight into the module-split hypothesis and out into any therapeutic reasoning built on it. The decision rule is written down here, before the data, for that reason.
  • If the calibration gate is skipped because the target counts look interesting, this project will have repeated an error it has already made once: rebuilding a tool to audit it, dropping a filter, and publishing a defect rate nearly double the truth.
Still open
  1. Does PXD009338 hold raw files or only processed search results? That single check decides whether this analysis is possible at all, and it was not run this cycle.
  2. Can the authors' search parameters be recovered from the deposit's metadata? Without them the calibration gate is harder to pass and a tail-regime difference between searches becomes indistinguishable from biology.
  3. Should the same re-search also recover the BAF/SWI-SNF subunits the earlier corpus analysis scored at 6 of 17? If several multi-module complexes show the same truncation pattern, the filter explanation strengthens for all of them at once.
Every source behind this entry (3)
  1. E1strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper

    The Skucha 2018 seven-fusion AP-MS retained the 300 most significant interactions per bait by combined p-value rather than applying a p threshold, so 26.6% of the tabulated per-bait p-values exceed 0.05, to a maximum of 0.175, and absence from the table is a rank truncation rather than a measured zero.

    MLL-fusion-driven leukemia requires SETD2 to safeguard genomic integrity. · Skucha A, Ebner J, Schmoellerl J, Roth M, Eder T, Cesar-Razquin A, Stukalov A, Vittori S, Muhar M, Lu B, Aichinger M, Jude J, Mueller AC, Gyorffy B, Vakoc CR, Valent P, Bennett KL, Zuber J, Superti-Furga G, Grebien F · Nature Communications 2018
    PMID 29777171DOI
    What this evidence cannot show

    The verbatim Methods sentence and the 26.6% figure were measured by an earlier beat of this project from the cached supplementary xlsx and are recorded in gene-chd4; I confirmed independently only the per-bait retained counts (291-293) from the same cached file on 2026-09-24, and did not re-read the Methods text myself. The inference that a truncated protein has a non-zero count is not stated by the paper - it is the reason to run this analysis rather than a conclusion of it.

    Evidence that argues against this (1)
    • A rank truncation does not guarantee a non-zero count below the cut - the tail of an AP-MS prey list contains genuine zeros as well as small counts, and for a protein that was never detected in any replicate the re-search will return nothing. So this analysis could equally confirm that the deacetylase module is genuinely undetected, and its value is that it distinguishes those cases rather than that it will find the missing counts. This repository 2026
  2. E2moderate confidence 3 of 5: One strong study, or several smaller consistent ones.database

    Twelve NuRD and Sin3 subunits are absent from all seven bait columns of the published 960-protein network - MTA1, MTA2, MTA3, MBD3, CHD5, HDAC1, HDAC2, RBBP4, RBBP7, SIN3B, SAP30 and SAP18 - and the five NuRD subunits that are present provide a calibration set with published counts (CHD4 7|9 in MLL-AF4 against a 2|2 control) against which a re-search must first reproduce.

    What this evidence cannot show

    The subunit lists are my curation, written out in the analysis script and auditable in the diff; a reader who defines NuRD or Sin3 differently gets a different list of twelve. The absent-member identities are persisted rather than only counted, so this plan does not depend on re-deriving them.

    Evidence that argues against this (1)
    • Reproducing the published counts is a necessary gate and not a sufficient one: a re-search that matches CHD4 could still differ systematically in the low-count regime, which is exactly the regime this analysis is about. Any reported count near the detection floor must carry a false-discovery statement, and a difference between my search and the authors' in the tail would be indistinguishable from the biology it is meant to measure. ProteomeXchange/PRIDE 2018
  3. E3strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.dataset

    The raw mass-spectrometry data behind the seven-fusion AP-MS are publicly deposited at ProteomeXchange as PXD009338, so the twelve truncated subunits' spectral counts are recoverable by re-search without generating new data.

    What this evidence cannot show

    THE ACCESSION IS CARRIED FROM THIS GRAPH'S OWN gene-chd4 ENTRY, WHICH CITES IT AS THE DEPOSIT FOR THIS PAPER; I DID NOT FETCH THE PRIDE PAGE IN THIS CYCLE. Whether the deposit contains raw files or only processed search results, which instrument and acquisition method were used, and whether the negative-control runs are included are all UNVERIFIED here and are the first things whoever runs this must check - if the deposit holds only processed results, this analysis is not possible as designed and the absence class stays measured-not-reported rather than measured-not-analysed.

    Evidence that argues against this (1)
    • An earlier beat of this project recorded that IntAct has never curated this dataset - a query on pubid:29777171 returns 0 rows against a working positive control - so no independent curator has re-examined these calls, and there is no external check on what the deposit contains. A deposit existing is not a deposit that answers your question, which is the same class of caution this graph applies to a PDB entry existing. Skucha A et al. 2018
Connected to