PLANNED: refit prey recovery on accessible lysine count instead of protein length, which separates the two mechanisms that could produce the size effect
also called lysine-count refit · is the BioID size effect a lysine effect
Bigger proteins are picked up more often in proximity screens. There are two reasons that could be true: a bigger protein has more of the chemical handles the labelling enzyme attaches to, or a bigger protein simply lingers nearby longer. Counting the handles would tell us which.
The same thing, for a scientist
`ana-prey-recovery-tracks-size-in-proximity-labelling-and-not-in-ap-ms-and-the-tail-was-scored-on-the-wrong-side-of-the-pair` measured a standardised negative-binomial coefficient of **+0.289 (IM-30059, K562 RBP BioID2)** and **+0.205 (IM-28767, HEK293 TF BioID)** for log10 protein length on distinct-bait recovery, and **+0.030 (p = 0.27, not significant)** in the AP-MS arm of the same publication, same baits, same cells [E1]. Size predicts **proximity** recovery and not **affinity-purification** recovery. BioID chemistry labels **primary amines, overwhelmingly lysine epsilon-amines**, so a mechanical reading is available: a larger protein presents more lysines within the labelling radius per encounter. The competing reading is residence time - larger, more disordered, more multivalent proteins linger. **These make different predictions and are separable with one refit.** This analysis is PLANNED and not run.
The design
Refit the same negative-binomial model on the same two depositions, replacing log10(length_aa) with log10(lysine count) and then with both, and compare:
- If lysine count subsumes length - the lysine coefficient stays near +0.29 and length falls toward zero when both are in the model - the size effect is chemical, and the correct covariate for every future proximity analysis in this graph is lysine count, not length.
- If length survives lysine count, size is doing something beyond presenting handles, and residence time or multivalency is live.
- If neither survives, the two are collinear at this n and the test is uninformative; report that rather than picking the one with the smaller p.
A second, sharper version uses solvent-accessible lysines from AlphaFold models rather than total lysines, because a buried lysine cannot be biotinylated. That is a heavier lift and should only be attempted if the crude version is promising.
Controls. Lysine fraction (lysines per residue) as a third covariate, to separate 'more lysines because bigger' from 'lysine-rich'. And the AP-MS arm as the negative control: lysine count should NOT predict AP-MS recovery, because affinity purification does not label amines. If it does, the whole framing is wrong and the result is about something else - abundance, stickiness or a shared confound.
Why it is worth an afternoon
This graph's most-cited deflationary finding about its own interactome is that KMT2A's proximity promiscuity is explained by size - and KMT2A, at 3,969 residues, is the largest protein in the project's core. If the mechanism is lysine count, the deflation is a statement about a chemical bias with a known fix (normalise per accessible lysine), and every proximity row in the graph can be rescored with a better covariate. If the mechanism is residence time, the deflation is a statement about biology - big disordered chromatin proteins really do spend more time near everything - and no normalisation helps.
The two readings also differ for the fusion. KMT2A::AFF1 retains roughly the N-terminal third of KMT2A, so it is a much smaller protein than wild-type KMT2A. Under the lysine reading the fusion should be recovered proportionally less in any future proximity map; under the residence-time reading it depends on what the AFF1 portion contributes, which is a different prediction and a testable one the first time anyone uses the fusion as a bait.
| script | analyses/interactome-tail-size-and-abundance-adjudication/run.py (to be extended) |
|---|---|
| status | planned |
| cost tier | low |
| timeline | one afternoon for the crude version; two days for the solvent-accessible version |
| result summary | NOT RUN. Planned as the discriminating refit for the size effect measured by the parent analysis. |
inputs
[ "the two PSICQUIC depositions already cached by run.py", "UniProt reviewed human proteome sequences (for lysine counts)", "optionally AlphaFold models for solvent-accessible lysine counts" ]
- If lysine count explains the size effect, every proximity-based claim in this graph becomes rescoreable - including the ones this graph currently relies on to DECLINE partners. A better covariate would move promotions in both directions, not just downward.
- Are lysine count and protein length too collinear in the human proteome for this test to resolve them at n = 2,627? That is checkable before running anything, by computing their correlation, and if it is above about 0.95 the test should be declared uninformative in advance rather than after.
- Does the BioID2 used in IM-30059 and the BirA* used in IM-28767 have the same amine preference? If not, the two depositions should not be pooled and the coefficient difference between them may be a reagent difference.
- E1strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.dataset
Standardised negative-binomial coefficients for log10 protein length on distinct-bait prey recovery are +0.289 in IM-30059 (BioID2, K562, n = 2,627 preys) and +0.205 in IM-28767's BioID arm (HEK293, n = 933), while in IM-28767's AP-MS arm on the same baits and cells the Poisson coefficient for log10 length is +0.030 with p = 0.27 and abundance is +0.239.
What this evidence cannot show
Pearson dispersion is 4.95 and 13.24 in the two proximity arms, so standard errors from the Poisson fits are understated; the coefficients themselves are point estimates and are not affected. The AP-MS arm was not separately checked for overdispersion, so its p = 0.27 is if anything an understatement of the uncertainty - which strengthens rather than weakens the claim that the length effect is absent there.
Evidence that argues against this (1)
- The BioID and AP-MS arms of IM-28767 differ in more than the labelling chemistry: they have different numbers of recovered preys (946 vs 641), different filtering thresholds applied by the authors before deposition, and the paper states the BioID arm yielded more high-confidence interactions. A coefficient difference between two differently filtered lists is not automatically a chemistry difference. source 2022
if the size term is a lysine-count artefact the predicted values move and the ratio moves with them
the discriminating refit for the size coefficient measured by the parent analysis
- supports →A BioID hit is a neighbour, not a complex member - and the two methods are known to disagree
size predicts BioID recovery and not AP-MS recovery in the same publication; lysine count would say whether that is chemistry