Regress prey abundance against interactor confidence class across all 1,330 baits of the Hein atlas - the data are already downloaded and the run is an afternoon
also called abundance-versus-class regression · the 31,944-row test
A computer-only experiment: take the whole published screen, and ask whether the proteins it calls weak hits are simply the most common proteins in the cell. Nothing has to be done at a bench.
The same thing, for a scientist
A dry experiment over two spreadsheets that are already in hand. Join the atlas's 31,944 interaction rows (bait cell line, prey, confidence class, both stoichiometries, enrichment, p value, prey-bait correlation) to its 8,804-protein HeLa proteome (copy number per prey), then model confidence class against log copy number with the number of baits a prey appears under as the second term. The question is whether **prey abundance predicts low-class membership specifically**, or predicts membership in every class equally, which would mean it is about detectability rather than contamination. This is the test named by `hyp-atlas-only-spokes-are-predicted-by-prey-abundance-not-by-the-bait`, and it is the only thing that converts this beat's two-pulldown observation into a measurement.
Design, and the controls that make it interpretable
Inputs. mmc3.xlsx sheet interactions (31,944 rows, 16 columns) and mmc4.xlsx sheet HeLa proteome (8,804 protein groups with proteomic-ruler copy numbers), both already fetched from the Elsevier static CDN and checked for magic bytes 1.
Model. Ordinal outcome = confidence class (A+ > A > A- > B+ > B > B- > C). Predictors = log10 prey copy number, promiscuity (number of distinct bait cell lines in which the prey appears), and log10 prey:bait abundance ratio. Report the class-specific marginal effect, not one pooled coefficient - brief rule 45's 'split a sign-dependent quantity before pooling it' applies directly, because abundance and enrichment move in opposite directions across the class ladder.
Controls that decide whether the result means anything.
- The detectability control. Fit the same model restricted to A+ rows. If abundance predicts A+ membership as strongly as B- membership, the effect is about what the instrument can see and the hypothesis is dead.
- The null for promiscuity. A prey appearing under many baits is the classic contaminant definition, and it is also what an abundant protein does by construction. Report the base rate of appearance per abundance decile before quoting any enrichment - brief rule 84.
- The spike. Take a set of preys that are unambiguously true complex members in independent structural literature (the DotCom subunits under the DOT1L bait, the RFC subunits under a replication bait, proteasome subunits under a proteasome bait) and require the model to NOT classify them as abundance-driven. If it does, the model is measuring the wrong thing and the analysis should refuse to report - brief rule 67's spike requirement.
- Report what the join dropped. Preys with no HeLa proteome row (BRINP1 is one) cannot enter the regression, and if the dropped set is concentrated in the low classes the result is about the remainder - brief rule 78.
Readout. Two numbers for the ledger: the marginal effect of abundance on low-class membership, and the fraction of this project's 15 atlas-only bait-supported partners that fall above the abundance threshold the model identifies.
Why this is worth doing rather than reading partner by partner
Fifteen atlas-only partners across the six core baits is roughly fifteen beats' worth of work if each is read individually, and this beat took a full cycle to do six. A model that tells the coordinator which of the fifteen are abundance-explained would let the remaining ones be read properly instead of all of them being read shallowly.
And it has a second use the graph needs more. analyses/interactome-bait-versus-prey and analyses/directed-interactome-coverage-recheck both grade partners by role. Neither grades by source quality, because no instrument existed. This would be that instrument, computed rather than asserted, which is exactly what rule 90 demands of any number that decides whether a beat runs.
| design | Ordinal regression of interactor confidence class on log10 prey copy number, prey promiscuity across bait cell lines, and log10 prey:bait abundance ratio, over all 31,944 interaction rows of PMID 26496610 joined to its 8,804-protein HeLa proteome. |
|---|---|
| readout | Class-specific marginal effect of abundance, plus the fraction of this project's 15 atlas-only bait-supported partners above the identified abundance threshold. |
| persist | Write the resolved prey identities per class, not only the coefficients - rule 90. A count that cannot be re-derived is a claim. |
| cost tier | low |
| timeline | one working day; both input files already downloaded and parsed in this beat |
| wet lab required | false |
controls
[ "Detectability control: refit restricted to A+ rows; if abundance predicts A+ as strongly as B-, the hypothesis is dead", "Base rate: appearance frequency per abundance decile reported before any enrichment claim (brief rule 84)", "Spike: DotCom subunits under the DOT1L bait plus two further bait/complex pairs must NOT be classified as abundance-driven, or the analysis refuses to report (brief rule 67)", "Dropped-row accounting: name and count preys with no HeLa proteome row and report the class distribution of the drop (brief rule 78)" ]
- A model that identifies abundance-driven spokes will immediately be used as a filter, and the AF9 row shows a true partner can look exactly like a weak one. The output must be a flag for reading, never an exclusion.
- Running this and reporting only a coefficient would repeat rule 90's defect inside the analysis written to address rule 90. The prey identities per class must be persisted.
- Is there a public contaminant repository (CRAPome or equivalent) that already answers most of this, and would checking it first make the regression unnecessary or merely confirmatory?
- Should the same design be run against BioPlex, whose baits are HEK293T and HCT116, to ask whether the abundance signature is cell-line-specific?
- E1established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.dataset
The two inputs are already in hand and parse cleanly: Hein et al.'s interactions table has 31,944 data rows across 16 columns including cell.line.ID, prey.IDs, interactor.confidence.class, log10.prey.bait.ratio, log10.prey.bait.expression.ratio and core.interactor, and the HeLa proteome sheet has 8,804 protein groups with per-protein copy numbers and concentrations.
Hein et al. 2015, supplementary workbooks mmc3.xlsx (interactions) and mmc4.xlsx (HeLa proteome) · Cell 163(3), supplementary material 2015What this evidence cannot show
Both downloaded 2026-09-21 from the Elsevier static CDN after the article HTML returned 403, with magic bytes checked (PK). The Elsevier CDN path is not a guaranteed-stable endpoint and the files should be archived locally before anyone depends on them. The join key is a semicolon-delimited protein group on both sides and is not one-to-one; resolving it is the main implementation cost and the dropped count must be reported.
Evidence that argues against this (1)
- A join between two semicolon-delimited protein-group columns can silently drop exactly the preys with the most ambiguous identifications - which are disproportionately the low-confidence ones this analysis is about. If the drop is concentrated in the B- class the regression answers a different question than the one asked, which is brief rule 78's failure shape. Cell 163(3), supplementary material 2015
- E2strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.dataset
The DOT1L pulldown already provides the internal spike this design requires: MLLT3, MLLT6 and MLLT10 are A+ in the same experiment and are independently established DOT1L complex members, so a model that classifies them as abundance-driven is demonstrably measuring the wrong quantity.
Hein et al. 2015, supplementary Table S2 (mmc3.xlsx) sheet 'interactions', cell line MCP_ky_0003877 · Cell 163(3), supplementary material 2015What this evidence cannot show
Three spike proteins from one bait is a thin positive control and the design calls for at least two more bait/complex pairs. MLLT3 is also the weakest possible spike on the stoichiometry axis, at 1:172, so a model tuned to pass it may be tuned too permissively.
Evidence that argues against this (1)
- MLLT3's own numbers - enrichment 3.74 SD, correlation 0.275, stoichiometry 0.0058 - sit close enough to the B- rows that a spike built on it may be unable to distinguish a well-calibrated model from a permissive one, which would leave the control uninformative rather than reassuring. Cell 163(3), supplementary material 2015
THE EXPERIMENT, RUN. It was specified last cycle as a one-day dry analysis over two workbooks already downloaded, and this is its execution: 31,944 interaction rows joined to the HeLa proteome at 97.3%, base rates printed per abundance decile before any model as the design required, the A+ refit run as the named control, and prey identities persisted per class rather than counts alone.
The regression over all 1,330 baits that would confirm or kill the hypothesis, with a DotCom spike required before it may report
Uses the DOT1L pulldown's A+ DotCom subunits as the internal spike control