Oncofusion

start / graph / risk-q141-a-genome-wide-peak-call-census-is-the-wrong-pharmacodynamic-assay-for-a-menin-inhibitor

Riskstrong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.Q141-menin-at-fusion-targets

A genome-wide peak-call census is the wrong pharmacodynamic assay for a menin inhibitor, in two independent ways

If you measure whether a menin drug is working by counting how many menin spots are still detectable across the whole genome, you will conclude it is not working even when it is - twice over: most of the genome is irrelevant, and the counting method cannot see the size of change the drug produces.

The same thing, for a scientist

Two independent defects compound. **(1) Denominator.** Fusion-occupied genes are ~1% of the measured gene universe (the authors' 202-gene list sits at the 0.7th percentile of HA occupancy over 17,343 genes), so a genome-wide summary is ~99% places where nothing should happen. Measured: MENIN dVHL/DMSO 0.503 at target genes, 0.974 genome-wide-ish. **(2) Dynamic range.** A binary peak call saturates at strong sites: MEN1 summit retention rises from 7% in the weakest control-peak decile to 83% in the strongest, so at exactly the sites the drug acts on, the assay cannot resolve a 2-fold change. Measured: menin loses ~50% of signal at target genes while retaining 83% of its strongest peak calls. **Either defect alone produces a false negative; together they make the assay uninformative.** The correct endpoint is quantitative occupancy on a defined target-gene set, spike-in normalised, with the peak-width ratio reported if any peak-derived figure is quoted.

Go deeper
How to tell whether an assay in a protocol has this problem

Ask two questions. What fraction of the measured units are places the drug should act? If it is under a few percent, the summary statistic is a dilution. What is the smallest change the readout could have detected at the signal level where the drug acts? If that number is not stated, a null result bounds only near-total loss.

Distinct from the risk already recorded in this corpus

This corpus already records that menin displacement is a poor pharmacodynamic proxy for disabling the fusion - a claim about what is being measured. This entry is about how it is measured: even if menin displacement were the right thing to measure, a genome-wide peak census would fail to measure it. The two risks are independent and both apply.

Facts and figures
kinddelivery
severityhigh for endpoint design; no direct patient risk
mitigationUse quantitative occupancy on a pre-specified fusion-target gene set with a spike-in; report the assay's demonstrated dynamic range at that signal level; if any peak-derived figure is quoted, report the treated/control peak-width ratio beside it and flag anything above 1.3.
affects
[
  "drug-revumenib",
  "drug-ziftomenib",
  "drug-bleximenib",
  "drug-pinometostat"
]
measured defects
{
  "denominator": "target genes are the 0.7th percentile of 17,343",
  "dynamic_range": "summit retention 7% to 83% across strength deciles; 50% signal loss leaves 83% of calls"
}
How this could bite us back
  • An insensitive endpoint does not only miss activity - it also makes an inactive analogue look equivalent to an active one, which is the failure mode that matters when ranking compounds in a class with four clinical agents.
Still open
  1. Has any menin-inhibitor trial used a chromatin-occupancy pharmacodynamic endpoint, and if so which form?
  2. Can the dynamic range be calibrated by read subsampling, converting every retention figure in this corpus into a quantified bound?
Every source behind this entry (3)
  1. E1strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.dataset

    The authors' 202-gene MLL::AF9 target list sits at the 0.7th percentile of HA occupancy across 17,343 genes, so a genome-wide summary is dominated by non-target genes.

    What this evidence cannot show

    Computed on one cell line; the fraction of the genome a fusion occupies varies between fusions and cell states.

    Evidence that argues against this (1)
    • A pharmacodynamic assay is not required to be sensitive at every site - a genome-wide readout deliberately trades sensitivity for robustness, and a target-restricted readout carries its own risk of circularity if the gene set is chosen from the same experiments. Olsen SN et al. 2022
  2. E2strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.dataset

    MEN1 summit-containment retention rises from 7% to 83% across control-peak strength deciles, so the binary measure saturates exactly where the drug acts.

    What this evidence cannot show

    MACS2 q<0.01 with --slocal 1500 in paired-end mode; the saturation is a property of this analysis pipeline, which is the one deposited.

    Evidence that argues against this (1)
    • A peak caller run with a fold-enrichment cutoff rather than a q-value cutoff would have a different dynamic range, so the defect is fixable within the peak framework and is not an argument against peak calling in general. Olsen SN et al. 2022
  3. E3strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper

    MEN1 interface mutations confer clinical resistance to revumenib by preventing inhibitor-induced eviction of menin and MLL1 from chromatin, which makes chromatin eviction the mechanistically correct endpoint - and therefore makes measuring it correctly consequential.

    MEN1 mutations mediate clinical resistance to menin inhibition · Perner F, Stein EM, Wenge DV, Singh S, Kim J, Apazidis A, et al., Levine RL, Armstrong SA, Cai SF · Nature 615(7954):913-919 2023
    PMID 36922589DOI
    What this evidence cannot show

    Clinical trial samples plus models; the eviction claim is at target loci, not genome-wide.

    Evidence that argues against this (1)
    • Eviction from chromatin is necessary for response but may not be sufficient, and clinical responses to menin inhibitors are frequently partial and non-durable even where eviction occurs, so a perfect chromatin endpoint would still not predict benefit. Perner F et al. 2023
Connected to