Oncofusion

start / graph / concept-proximity-labelling-is-not-co-purification

Explainerestablished confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.Q20-proximity-interactome

A BioID hit is a neighbour, not a complex member - and the two methods are known to disagree

also called proximity labelling caveat · BioID interpretation · TurboID caveat · MI:1314

Attaching a tagging enzyme to a protein marks everything that comes close to it for a few hours. That is a list of who was nearby, not a list of who was holding on.

The same thing, for a scientist

Proximity-dependent biotinylation fuses a promiscuous biotin ligase to a bait; the ligase releases reactive biotin that labels accessible lysines on proteins **within roughly a 10 nm radius** over a labelling window of minutes (TurboID) to many hours (BirA*/BioID2) [E1][E2]. The output is an abundance-weighted, compartment-structured census of the bait's vicinity. A direct head-to-head against affinity purification on chromatin-associated complexes found the two recover **few shared partners** and enrich for **distinct biological processes**, with BioID preferentially recovering lower-abundance proteins [E3]. And in the largest human survey, a prey's bait profile is dominated by **subcellular compartment** rather than by binding [E4]. Four practical rules follow, and this graph's proximity entries all state them.

Go deeper
The four rules

1. A hit is a neighbour. 'X was recovered in Y's BioID' means X came within labelling radius of Y at some point. It does not mean X binds Y, that they are in one complex, or that they are simultaneously present.

2. Recovery is abundance-weighted and compartment-structured. Nuclear proteins label nuclear proteins. A splicing bait recovering elongation machinery is what co-transcriptional co-residence predicts 4.

3. Counts are detectability, not membership. In the one dataset this graph has, obligate complexes split wildly across the same fifty baits: WDR5 11, ASH2L 9, RBBP5 0, DPY30 0; CDC73 25, RTF1 15, PAF1 6, CTR9 3, LEO1 0. If proximity recovery tracked membership, those would cluster.

4. Spoke expansion is not a binary interaction. Curated proximity rows are usually bait-to-prey assignments expanded from a pulldown (PSI-MI MI:1060), so a deposited 'interaction' between two preys of the same bait does not exist as an observation at all.

What proximity labelling is uniquely good for

It is not a weaker interactome method; it answers different questions 3.

  • Transient and dynamic contacts. A transcription-factor BioID/AP-MS comparison found BioID gave more high-confidence interactions and attributed that to how transient TF interactions are 5.
  • Localisation. The strongest result in the published oncofusion TurboID work was that CBFB::MYH11's partners are cytoplasmic - a fact about where, which fractionation-plus-pulldown is poorly placed to find.
  • Insoluble and chromatin-bound compartments, where lysis conditions dictate the answer.

So the right use in this graph is to ask where is the fusion and what is around it, not what is the fusion made of.

How this concept should be applied to the entries in this folder

Every entry in data/graph/nodes/proximity-interactome/ names its bait, its prey, its cell line and its method code, and none of them uses the edge type binds. They use co_occurs_with and related_to. If a future cycle promotes one of these to binds, that is a claim requiring a new experiment, not a re-reading.

Facts and figures
labelling radius NM10
labelling window bioid h18
labelling window turboid min10
method codes
{
  "proximity-dependent biotin identification": "MI:1314",
  "spoke expansion": "MI:1060",
  "bait": "MI:0496",
  "prey": "MI:0498"
}
edge types permitted for proximity
[
  "co_occurs_with",
  "related_to"
]
edge types forbidden
[
  "binds",
  "part_of"
]
How this could bite us back
  • The rules in this entry are easy to state and are routinely violated by summarisation. Any downstream figure that draws a proximity hit as an edge between two proteins has already broken them.
  • Treating proximity data as unreliable is the opposite error and is equally wrong: the method found cytoplasmic RUNX1 sequestration by CBFB::MYH11, which nuclear-extract pulldowns had missed for years.
Still open
  1. Can compartment be subtracted from a single-study proximity dataset without a 192-bait reference map, and if not, should this graph treat every single-study proximity claim as compartment-confounded by default?
  2. Is there any published case of a proximity hit later shown to be a direct binder by structure, for a chromatin elongation complex specifically?
  3. What would a negative control bait look like for a KMT2A::AFF1 proximity experiment - nuclear-localised inert protein, or the reciprocal AFF1::KMT2A fusion?
Every source behind this entry (5)
  1. E1established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.paper

    BioID fuses a promiscuous mutant biotin ligase to a bait protein, which locally activates biotin and biotinylates proximal and interacting proteins in living mammalian cells; the method identifies proteins near the bait as well as those bound to it.

    PMID 22412018DOI
    What this evidence cannot show

    The founding methods paper, demonstrated on lamin A. Labelling radius is an inference from that system and is conventionally quoted as about 10 nm; the paper does not establish a universal radius, and the effective radius depends on the ligase, the biotin concentration and the labelling time.

    Evidence that argues against this (1)
    • A later engineered ligase (TurboID) labels in 10 minutes rather than 18-24 hours, which changes what 'at some point during the window' means; a conclusion drawn from a long-window BioID experiment may not hold for a short-window TurboID one, and vice versa. Branon TC et al. 2018
  2. E2established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.paper

    TurboID and miniTurbo are engineered biotin ligases that achieve efficient proximity labelling in living cells and organisms in 10 minutes rather than the 18 hours required by BirA*, giving much greater temporal resolution.

    PMID 30125270DOI
    What this evidence cannot show

    Abstract-level use. Faster labelling also means higher background biotinylation and potential toxicity from biotin depletion; the paper discusses trade-offs this entry does not enumerate.

    Evidence that argues against this (1)
    • Short labelling windows reduce integration over time but do not remove the compartment bias, which is spatial rather than temporal; a 10-minute TurboID neighbourhood is still a neighbourhood. Go CD et al. 2021
  3. E3established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.paper

    A direct comparison of BioID against an affinity-purification protocol optimised for chromatin-associated protein complexes found the two approaches identified few shared interaction partners and enriched for distinct biological processes, while both permitted recovery of biologically meaningful interactions; BioID allowed purification of proteins of lower cellular abundance.

    PMID 25281560DOI
    What this evidence cannot show

    HEK293, on core histones and Mediator subunits. The generality of the low-overlap finding across other complexes and cell types is asserted here rather than demonstrated.

    Evidence that argues against this (1)
    • The paper's own conclusion is complementarity, not incompatibility. Reading it as 'proximity data are unreliable' inverts its message; the correct reading is that neither method alone gives a complete roster. Lambert JP et al. 2015
  4. E4established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.paper

    A proximity-dependent biotinylation map of a human cell built from 192 baits demonstrates that proximity data resolve subcellular compartments, including compartments refractory to microscopy and fractionation, so a prey's profile across baits is organised by localisation.

    PMID 34079125DOI
    What this evidence cannot show

    HEK293 only. The study was designed to resolve compartments, so it is both the best evidence for the compartment effect and the study most invested in it.

    Evidence that argues against this (1)
    • The same map resolves complexes within compartments, so compartment structure does not exhaust the information in proximity data - it means compartment must be controlled for before an interaction claim, which most individual studies do not do. Go CD et al. 2021
  5. E5strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper

    In a study of 109 human transcription factors, proximity-dependent biotinylation identified 6,703 interactions against 1,536 by affinity purification mass spectrometry, and the authors attribute the greater number of high-confidence BioID interactions to the transient and dynamic nature of many transcription-factor interactions.

    Human transcription factor protein interaction networks. · Göös H, et al. · Nat Commun 2022
    PMID 35140242DOI
    What this evidence cannot show

    Abstract read; full text not read. 'More interactions' is a count, and the attribution to transience is the authors' interpretation rather than a demonstrated result.

    Evidence that argues against this (1)
    • A larger hit list can equally reflect a higher false-positive rate: the two methods have different background structures, and the chromatin-complex comparison found each enriched for processes the other missed in both directions. Lambert JP et al. 2015
Connected to