Oncofusion

start / graph / ana-the-weak-inhibitor-was-doing-the-same-thing-weakly-and-the-selectivity-claim-strengthens-under-filtering

Analysismoderate confidence 3 of 5: One strong study, or several smaller consistent ones.

MI-2-2 was doing a weak version of the same thing - 5.8x enriched overlap with the potent inhibitor, 75% concordant - and the selectivity claim it left behind gets STRONGER when low-expressed probes are filtered out, 9.6x to 22.1x

also called MI-2-2 versus VTP-50469 overlap · selectivity floor sweep · GSE63664 GSE127506 intersection

A weak and a strong version of the same kind of drug were tested on the same leukaemia cells in two separate public studies, and the genes the weak one moved are mostly a subset of the genes the strong one moved, in the same direction. A separate worry about the earlier analysis - that its headline number came from barely-detectable genes - turned out to be wrong in the reassuring direction.

The same thing, for a scientist

Two questions left open by the cycle-56 correction, both answered here against the same cell line, **RS4;11**, the only one both datasets contain. **1. Was MI-2-2 doing a weak version of the same thing, or something else?** A weak version of the same thing. Within a universe of **9,595 genes expressed in BOTH DMSO arms** (of 18,111 shared symbols), MI-2-2 moves **61** genes by >=1 log2 and VTP-50469 moves **870**. The overlap is **32 against 5.5 expected under independence - a 5.79x enrichment - and 24 of the 32 agree in direction (75%)**. The subset test is the informative part: of the **top 50** genes by VTP-50469 effect size, MI-2-2 also moves **22%**, against **1.0%** of ranks 51-150 and 1.9% of ranks 401-870. So the weak compound preferentially reached the genes the potent compound hits hardest, which is what partial target engagement looks like. It is not a clean compression, though: across the 899 genes moved by either, **Pearson r is only 0.184** with a slope of 0.67, so most of each response is invisible to the other. **2. Was cycle 55's selectivity ratio an artefact of unfiltered low-expressed probes?** No - and this is the opposite of what the first result made me expect. That analysis applied **no expression floor at all**, and this script's cross-platform filter cut RS4;11's MI-2-2 movers from 692 probes to 61 genes, so the worry was real. Re-tested by sweeping an array expression floor: | ratio | no floor | >=6 | >=7 | >=8 | |---|---:|---:|---:|---:| | SEM MI-2-2 / REH | 9.61 | 13.12 | **22.07** | 22.50 | | SEM EPZ004777 / REH | 3.71 | 2.09 | **2.23** | 2.40 | | RS4;11 MI-2-2 / REH | 8.65 | 9.47 | **12.36** | 12.88 | | RS4;11 EPZ004777 / REH | 1.81 | 1.68 | **1.57** | 1.80 | **The menin-inhibitor selectivity more than doubles under the floor while the DOT1L-inhibitor control falls.** If the ratio had been carried by background probes both would have collapsed together. The selectivity is concentrated in well-expressed genes, and the one cycle-55 claim the correction left standing now stands on a test designed to break it.

Facts and figures
scriptanalyses/was-the-weak-inhibitor-doing-the-same-thing-weakly/run.py
resultsanalyses/was-the-weak-inhibitor-doing-the-same-thing-weakly/results.json
restricted toRS4;11, the one cell line GSE63664 and GSE127506 share
universe18,111 symbols measured by both platforms; 9,595 of them expressed above floor in BOTH DMSO arms (array log2 RMA >= 6.0, RNA-seq log2CPM >= 1.0)
moversMI-2-2 61 (21 up / 40 down); VTP-50469 870 (489 up / 381 down)
overlap32 observed, 5.5 expected under independence, 5.79x enrichment, 24 concordant / 8 discordant (75%)
effect size relationPearson r = 0.184 over the 899 genes moved by either; slope 0.67 VTP per MI-2-2 unit
what mi22 moved that vtp did notTop MI-2-2-only movers by effect size: SOCS2-AS1 -2.17, MPEG1 +1.95, MYO6 -1.81, PARVB -1.69, DSE -1.65, DYNC1I1 -1.63, RNF175 +1.60, IRX3 +1.58, TMEM45A -1.54, CTSG -1.53, TXNIP -1.40, ZMYND11 -1.17. Several sit at or near the RNA-seq detection floor (SOCS2-AS1 1.06, MYO6 1.32, DSE 1.26, SMAGP 1.06, DYNC1I1 1.06), so 'VTP did not move it' and 'the RNA-seq barely sees it' are not separated for those.
shown for fusiontrue
absence classmeasured-not-analysed
subset test by vtp effect rank
{
  "ranks 1-50": "11 of 50 also moved by MI-2-2 (22.0%)",
  "ranks 51-150": "1 of 100 (1.0%)",
  "ranks 151-400": "11 of 250 (4.4%)",
  "ranks 401-870": "9 of 470 (1.9%)"
}
selectivity floor sweep
{
  "SEM MI-2-2 / REH": "9.61 (no floor) -> 13.12 (>=6) -> 22.07 (>=7) -> 22.50 (>=8)",
  "SEM EPZ004777 / REH": "3.71 -> 2.09 -> 2.23 -> 2.40",
  "RS4;11 MI-2-2 / REH": "8.65 -> 9.47 -> 12.36 -> 12.88",
  "RS4;11 EPZ004777 / REH": "1.81 -> 1.68 -> 1.57 -> 1.80",
  "raw_mover_counts_no_floor_to_floor_8": "SEM MI-2-2 769/446/309/180; RS4;11 MI-2-2 692/322/173/103; REH MI-2-2 80/34/14/8; REH EPZ 248/94/47/20",
  "reading": "The drug-of-interest ratio rises and the control ratio falls. A low-expression artefact would have moved both the same way."
}
How this could bite us back
  • Nothing here concerns a patient or a treatment decision. It is a methodological comparison of two public datasets, and the only clinical-adjacent statement it supports is that a first-generation tool compound partially engaged the same axis as a potent one.
  • The 5.79x enrichment should not be quoted without the 0.184 correlation beside it. The two numbers say different things - shared axis, small shared variance - and the first alone overstates the agreement.
  • The strengthened selectivity ratio is a statement about MI-2-2 in GSE63664 and says nothing about whether menin inhibitors are selective in patients. It also rests on probe counts with a denominator as small as 8 at the highest floor.
Still open
  1. What are the 32 concordant genes? They are the only genes in this graph known to respond to two chemically unrelated menin inhibitors in the same line, which makes them the closest thing available to a compound-independent menin-dependence signature. The script reports the counts and the MI-2-2-only list but does not name the intersection, which is a gap in the instrument rather than in the data.
  2. Does the enrichment survive a same-platform comparison? The cleanest version of this test is VTP-50469 against MI-2-2 on ONE platform, and GSE127508 has further subseries including a MEN1 knockout RNA-seq (GSE137813) that would give a genetic rather than pharmacological comparator on the same assay.
  3. Is the 0.184 correlation limited by the array's n=2 DMSO arm in RS4;11? SEM has the same problem and REH does not. Re-running the intersection with REH as a negative control would say how much of the weak correlation is noise rather than biology.
  4. Should the cycle-55 analysis be re-run with an expression floor throughout? Its probe counts are quoted in the ledger and the king log, and 692 versus 322 at a floor of 6 is a large enough difference that the published figures should either be re-stated with a floor or explicitly labelled as unfiltered.
Every source behind this entry (2)
  1. E1moderate confidence 3 of 5: One strong study, or several smaller consistent ones.database

    Within 9,595 genes expressed above floor in both DMSO arms of RS4;11, MI-2-2 (GSE63664, 10 uM, 6 days, array) moves 61 genes by at least 1 log2 and VTP-50469 (GSE127506, 330 nM, day 7, RNA-seq) moves 870. Their overlap is 32 genes against 5.5 expected under independence from the marginal counts - a 5.79-fold enrichment - with 24 of the 32 concordant in direction and 8 discordant. Of the 50 genes with the largest VTP-50469 effect, MI-2-2 also moves 11 (22.0%), against 1 of 100 (1.0%) at VTP ranks 51-150 and 9 of 470 (1.9%) at ranks 401-870. Across the 899 genes moved by either compound the Pearson correlation of effect sizes is 0.184 with a slope of 0.67.

    What this evidence cannot show

    A cross-platform gene-level overlap is a LOWER BOUND on agreement: probe-to-symbol mismatch, annotation drift between a 2014 array annotation and the RNA-seq submitter's, and different detection floors all destroy real signal, so a 5.79x enrichment is consistent with substantially more shared biology than it shows. The MI-2-2 DMSO arm in RS4;11 is n=2 because the submitter excluded RS411-DMSO-1 for QC, which inflates MI-2-2 fold changes and therefore its mover count - biasing the enrichment DOWNWARD. The expression floors are conventions chosen before the overlap was computed but informed by the cycle-55 observation that housekeeping probes sit near log2 14 and absent ones near 3-5. Day 6 versus day 7 is close but not identical. No p-value is attached to the enrichment; the expected count under independence is the null and 32-against-5.5 is reported as a ratio rather than a test.

    Evidence that argues against this (2)
    • The correlation of 0.184 argues against the tidy reading. If MI-2-2 were simply a compressed version of VTP-50469 the effect sizes should track, and they barely do - most of each compound's response is invisible to the other. The enrichment and the subset test say the responses share an axis; the correlation says that axis explains a small part of either. Both are true and an entry quoting only the 5.79x would be overstating it.
    • Several of the largest MI-2-2-only movers sit at the RNA-seq detection floor (SOCS2-AS1 log2CPM 1.06, DYNC1I1 1.06, SMAGP 1.06, DSE 1.26, MYO6 1.32). For those genes 'VTP-50469 did not move it' is indistinguishable from 'the 3'-end assay cannot see it', so the MI-2-2-only list is not a clean off-target set and must not be read as one.
  2. E2moderate confidence 3 of 5: One strong study, or several smaller consistent ones.database

    Cycle 55's fusion-versus-wild-type selectivity ratio is not an artefact of unfiltered low-expressed probes, and strengthens under an expression floor. Probes moving at least 1 log2 with no floor, then requiring a DMSO log2 RMA of at least 6, 7 and 8: SEM MI-2-2 769/446/309/180, RS4;11 MI-2-2 692/322/173/103, REH MI-2-2 80/34/14/8, REH EPZ004777 248/94/47/20. The resulting ratios are SEM MI-2-2 over REH 9.61 -> 13.12 -> 22.07 -> 22.50 and RS4;11 MI-2-2 over REH 8.65 -> 9.47 -> 12.36 -> 12.88, while the DOT1L-inhibitor control ratios FALL or stay flat (SEM EPZ 3.71 -> 2.09 -> 2.23 -> 2.40; RS4;11 EPZ 1.81 -> 1.68 -> 1.57 -> 1.80).

    What this evidence cannot show

    The floors are arbitrary cut-points on a continuous quantity and the ratios are ratios of counts, so they inherit the instability of small denominators - REH's MI-2-2 count falls to 8 probes at the highest floor, and a ratio computed on 8 is not a stable number. That is exactly why the DOT1L control is reported at every floor: it shares the denominator problem and does not show the same trend. The comparison remains within one dataset, one platform and one normalisation, so it tests whether the cycle-55 ratio was a filtering artefact and nothing else - it does not address compound potency, which is settled elsewhere.

    Evidence that argues against this (1)
    • This test was run because the intersection analysis made the opposite outcome look likely: the same filter that produced the 5.79x enrichment cut RS4;11's MI-2-2 movers from 692 to 61, which is an 89% loss and looks exactly like a result carried by background. The floor sweep shows the loss is spread across both fusion and wild-type lines rather than concentrated in the fusion lines, so the RATIO survives what the COUNT does not. Anyone citing the 692 figure as a count of responding genes should stop; anyone citing the ~9x ratio may continue.
Connected to