Oncofusion

start / graph / ana-atlas-confidence-tracks-stoichiometry-not-abundance

Analysisstrong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.

Atlas confidence class tracks the prey:bait ratio, not abundance - so the proposed copy-number triage of the atlas-only partners does not exist

also called atlas confidence tracks stoichiometry not abundance · the copy-number triage does not exist

When a large survey reports that two proteins stick together, one worry is that it only found the protein because there is a lot of it lying around. We tested that on 31,944 measurements and it is not what is happening: what marks a confident result is the ratio of the two proteins pulled out together, not how abundant the partner is.

The same thing, for a scientist

Runs `exp-abundance-versus-confidence-regression-over-the-hein-atlas` and tests `hyp-atlas-only-spokes-are-predicted-by-prey-abundance-not-by-the-bait` over all 31,944 interaction rows of Hein 2015 joined to its HeLa proteome (97.3% join). Abundance barely moves the lowest confidence class (standardised beta +0.041, odds 1.04 per SD) and NEGATIVELY predicts A+ (-0.288). The dominant predictor of A+ is log10 prey:bait ratio (+0.884, odds 2.42 per SD). Unadjusted, B- share FALLS across abundance deciles, 28.6% to 24.0%. The DOT1L spike settles it: MLLT6 is A+ at 169 copies while NIN is B- at 1,070 and EBNA1BP2 is B- at 1.7 million. Outcome by prediction: 1 fails in its absolute form and holds comparatively only because the A+ coefficient is negative; 2 is contradicted (promiscuity beta -0.134 on B-); 3 is UNTESTED. The operative result is that the 15 atlas-only partners cannot be triaged on abundance and must be read.

Go deeper
A sign test with no effect-size floor gave the wrong answer first
What this costs the project
Prediction by prediction, because 'refuted' was too flat
Facts and figures
analysis diranalyses/hein-abundance-versus-confidence/
runs experimentexp-abundance-versus-confidence-regression-over-the-hein-atlas
tests hypothesishyp-atlas-only-spokes-are-predicted-by-prey-abundance-not-by-the-bait
outcomeNOT SUPPORTED
n interaction rows31944
n joined to copy number31085
join rate PCT97.3
unadjusted b minus share by abundance28.6% in the lowest decile falling to 24.0% in the highest
shown for fusionNOT APPLICABLE - this is a property of a proteomics dataset's confidence scoring, not a claim about any fusion protein. It bears on the fusion only through the 15 atlas-only partners of the core complex.
dot1l spike
MLLT6169.0MLLT6: 169MLLT10394.0MLLT10: 394NIN1,070NIN: 1070MAPK73,260MAPK7: 3260MLLT33,700MLLT3: 3700DOT1L3,780DOT1L: 3780PARP1612,000PARP16: 12000IBTK38,300IBTK: 38300EBNA1BP21,730,000EBNA1BP2: 1730000
standardised log odds
{
  "P(class = B-)": {
    "log10_copy_number": 0.0413,
    "prey_promiscuity": -0.1343,
    "log10_prey_bait_ratio": -0.382
  },
  "P(class = A+)": {
    "log10_copy_number": -0.2877,
    "prey_promiscuity": 0.0072,
    "log10_prey_bait_ratio": 0.8836
  }
}
dot1l spike — the numbers behind the chart
{
  "MLLT3": {
    "class": "A+",
    "copy_number": 3700
  },
  "MLLT6": {
    "class": "A+",
    "copy_number": 169
  },
  "MLLT10": {
    "class": "A+",
    "copy_number": 394
  },
  "DOT1L": {
    "class": "A+",
    "copy_number": 3780
  },
  "PARP16": {
    "class": "A-",
    "copy_number": 12000
  },
  "IBTK": {
    "class": "B",
    "copy_number": 38300
  },
  "EBNA1BP2": {
    "class": "B-",
    "copy_number": 1730000
  },
  "MAPK7": {
    "class": "B-",
    "copy_number": 3260
  },
  "NIN": {
    "class": "B-",
    "copy_number": 1070
  }
}
outcome by prediction
{
  "1_abundance_predicts_low_class": "absolute claim FAILS (odds 1.04 per SD); comparative ordering holds only via the negative A+ coefficient",
  "2_promiscuous_preys_fill_low_classes": "CONTRADICTED (promiscuity beta -0.134 on B-)",
  "3_atlas_only_partners_replicate_less": "UNTESTED - needs BioPlex or a second atlas"
}
How this could bite us back
  • This is a methods result about a proteomics dataset and implies nothing about any treatment. It should not be cited to support or oppose targeting any protein named in it.
  • The result makes the graph's atlas-derived interaction claims HARDER to dismiss, not easier to believe. It removes one specific reason for doubt and supplies no positive evidence that any of those interactions is direct - and spoke-expanded rows in the same dataset are curator-inferred pairs that were never observed.
Still open
  1. If abundance does not explain the low-class spokes, what does? The remaining candidates are genuine weak or transient binding, indirect membership of the same complex, and bait-specific artefacts such as tag proximity. This analysis rules out one and does not discriminate the others.
  2. The 15 atlas-only bait-supported partners of the core complex cannot be discounted on abundance grounds, which means they need reading rather than filtering. That was the practical purpose of the experiment and the answer is the more expensive one.
  3. Does the same hold in BioPlex, whose confidence model is built differently? If confidence tracks stoichiometry there too it is a property of AP-MS; if not, it is a property of Hein's scoring.
Every source behind this entry (3)
  1. E1strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.analysis

    Across 31,085 interaction rows joined to a copy number, prey abundance does not predict membership of the lowest confidence class. The standardised log-odds is +0.041, an odds ratio of 1.04 per standard deviation of log10 copy number, and unadjusted the B- share FALLS across abundance deciles from 28.6% to 24.0%. The hypothesis predicted abundant preys accumulating in weak classes; they do not.

    hein-abundance-versus-confidence: does prey abundance predict low confidence class specifically, or every class equally? · This repository (reproducible from the Hein 2015 supplementary workbooks) 2026
    What this evidence cannot show

    ONE ATLAS, ONE CELL BACKGROUND (HeLa), ONE PIPELINE. This describes how Hein's confidence classes were assigned, not a general property of affinity purification. Confidence class is Hein's own construct, so the analysis inherits whatever went into building it. The full class-by-decile table is printed rather than summarised for that reason. No confidence interval is placed on the coefficients. What carries the argument is the sign and the relative magnitude of the abundance and ratio terms, not their exact values (rule 87). 634 of 31,085 rows were dropped for non-finite predictors, where a bait or prey intensity was zero. The count is reported rather than the rows silently coerced.

    Evidence that argues against this (1)
    • The ADJUSTED coefficient is weakly positive (+0.041) while the UNADJUSTED trend is negative, so the two disagree in sign. A reader could take the adjusted number as weak support. The reason to prefer the marginal here is that the adjustment conditions on prey promiscuity and prey:bait ratio, both of which are themselves abundance-related, so the model is partialling out the very pathway the hypothesis proposes. Either way the effect size is ~1.04 per SD and cannot carry an argument. Hein MY et al. 2015
  2. E2strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.analysis

    What confidence class actually tracks is the prey:bait ratio. For A+ membership the standardised log-odds on log10 prey:bait ratio is +0.884, an odds ratio of 2.42 per standard deviation, against -0.288 for abundance and +0.007 for promiscuity. Confidence is a statement about STOICHIOMETRY - how much prey came out per unit of bait - not about how much of the prey the cell contains.

    What this evidence cannot show

    The prey:bait ratio is one of the quantities Hein's own confidence class was built from, so this is partly circular: the model recovers a construction rule rather than discovering a fact about biology. That is still the useful answer to the question asked, which was whether ABUNDANCE drives the class. Standardised coefficients are comparable to one another and are not odds ratios per copy.

    Evidence that argues against this (1)
    • Because class was constructed partly from stoichiometry, finding that stoichiometry predicts class is close to tautological, and the honest conclusion is narrower than 'stoichiometry determines whether an interaction is real'. What the result does establish is the negative: abundance is not the hidden driver, which is what the hypothesis asserted. Hein MY et al. 2015
  3. E3strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.analysis

    The DOT1L bait settles it without a model. Within one pulldown, A+ spans 169 to 3,780 copies and B- spans 1,070 to 1,730,000. MLLT6 is A+ at 169 copies - one of the least abundant preys in the experiment - while NIN is B- at 1,070 and EBNA1BP2 is B- at 1,730,000. Abundance cannot be what separates the DotCom core from the low-class spokes, because the core contains the rarest protein in the set.

    What this evidence cannot show

    Nine preys of one bait is a small spike and is offered as an illustration that survives inspection, not as the statistical result. The 31,085-row model is the result. DBC1 appears in the DOT1L rows with no copy number because the prey has no entry in the study's own 8,805-protein HeLa proteome; a separate beat declined to write it an entry for that reason.

    Evidence that argues against this (1)
    • A single bait can be unrepresentative, and DOT1L was chosen precisely because this project already knows the answer for it - which makes it a positive control, not independent evidence. Had the spike agreed with the hypothesis it would equally have been one bait. Hein MY et al. 2015
Connected to