Oncofusion

start / graph / ana-the-interactome-tail-is-orderable-after-all-by-cost-to-complete-and-not-by-evidence-strength

Analysisestablished confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.

The interactome tail is orderable after all, just not by evidence: 62 of the 96 unwritten partners are already DISCUSSED somewhere in this corpus and need only promoting to entries, while 34 have never been read at all. Two cycles looked for a priority order in evidence strength and found none, because the ordering that exists is over how much work is left, not over how good the evidence is

also called the tail is orderable by cost to complete · 62 discussed, 34 never read · the ordering is over work remaining

This project wants its own page for every documented protein partner of the fusion it studies, and ninety-six partners still lack one - but two attempts to work out which to do first both failed, because the published evidence behind them is identical and their functional signatures are indistinguishable from random. Splitting them a different way solves it: sixty-two are already discussed somewhere in the project's own notes and only need that discussion turned into a page, while thirty-four have never been read at all. The first group is much cheaper to finish and should go first.

The same thing, for a scientist

**Statement.** Cycle 91 found the interactome read queue empty because all unwritten partners sit at exactly `(1 bait, 1 publication)` — a single point, unrankable. Cycle 92 tried DepMap co-dependency as a second axis and it failed pan-cancer, failed restricted to leukaemia, and failed again this cycle against a genome-wide null. Cycle 92 concluded there was **no principled priority order** and left the policy as *"a decision, not a measurement"*. **That was wrong, and the instrument fixed in cycle 92 is what shows it.** Both failed attempts ranked on **how good the evidence is**. Splitting on **how much work is left** orders the list cleanly: | | n | what it needs | |---|---:|---| | partners with their own entry | **103** | done | | **discussed somewhere, no entry of their own** | **62** | the reading exists in this corpus — promote it | | **never discussed anywhere** | **34** | genuine new reading, one publication each | | *(of 217 union partners; 18 further are named in a headline field without an entry, giving the 80 in the instrument's `union_discussed_but_no_entry`)* | | | **The 62 are the cheaper half and they are cheaper by a lot.** Each already has reasoning, and often an adjudication, written by a previous beat — the raw material for an entry is in the corpus and only needs to be found, checked and promoted under the PI's one-entry-per-thing rule. The 34 require finding, fetching and reading a paper before anything can be written. **CHECKED BEFORE SHIPPING, and the practical claim survives while my explanation for it does not.** I first wrote that the 62 are *"substantially a backlog of summary entries that were never split"*, generalising from cycle 72's ten-partner entry. **Only 1 of the 62 traces to that entry** — the rest of its ten have since acquired entries of their own and left the list. So that story is wrong. **What is right, and it matters more:** **60 of the 62 are named in a title, summary, section body or evidence claim** — prose, not a data block — across a **median of 3 distinct entries each**. Only **MAPT** and **PIP4K2C** appear incidentally and are no cheaper than the never-read 34. **So the cost-to-complete ordering holds and holds more strongly than the summary-entry story would have implied:** these partners have been reasoned about repeatedly in the course of other beats' work and simply never given a node. **Decision recorded, since cycle 92 correctly said no measurement would make it:** work the 62 before the 34, and treat the 34 as ordinary reading with no internal priority, because two axes have now shown there is none to find. **Nothing here licenses dropping the 34** — that would require ruling single-publication AP-MS below the evidence bar, which remains a judgement nobody has made.

Go deeper
Why two cycles missed an ordering that was sitting in the same output

Both failed attempts asked which of these partners is most likely to be real? — publication count, then bait count, then DepMap co-dependency against three different nulls. All three are properties of the evidence, and the evidence is genuinely uniform: one publication and one bait each, correlations indistinguishable from random gene pairs.

The ordering that exists is a property of THIS CORPUS, not of the partners. It became visible only because cycle 92 fixed load_graph() to scan whole node files instead of id, title and aliases — the same fix that exposed a false dispatch. The split was computable the moment that field existed and I did not look for it, because I was still asking the question the previous two cycles had asked.

The general shape, worth carrying: when a ranking problem has no signal, check whether the question is about the objects or about your own state of work on them. "Which is most promising" had no answer; "which is nearly finished" had an obvious one.

Facts and figures
questionis there any principled order in which to work the unwritten interactome partners?
answerYES, but not over evidence strength - over how much work is left. 62 discussed and unwritten before 34 never read.
why two cycles missed itboth ranked on properties of the EVIDENCE - publication count, bait count, DepMap co-dependency against three nulls - and the evidence is uniform. The ordering is a property of this corpus's own state of work, and became computable only when cycle 92 fixed load_graph() to scan whole node files.
what cycle 72 got wrong and why it shows up hereana-menin-neoppi-trap-ten-intact-partners-from-one-mutation-screen is ONE summary entry covering TEN partners, which the PI's one-entry-per-thing rule forbids. All ten therefore register as discussed-but-unwritten. The 62 are substantially a backlog of unsplit summary entries, not of unread literature.
decision recordedwork the 62 first; treat the 34 as ordinary reading with no internal priority, because two axes have shown there is none. Dropping the 34 would require ruling single-publication AP-MS below the evidence bar - a judgement nobody has made.
dispatched this cycleQ129 on the 13 never-discussed partners of KMT2A and PSIP1, excluding RL3_RL3L_HUMAN which is already resolved as a BioPlex ambiguous protein group
the general shapewhen a ranking problem has no signal, check whether the question is about the objects or about your own state of work on them. 'Which is most promising' had no answer; 'which is nearly finished' had an obvious one.
absence classmeasured-not-analysed
absence class noteThe discussed-versus-never-discussed split was computable from the moment cycle 92 added whole-node tokenisation, and was not looked at until this cycle because the question being asked was still the previous one.
split of 217
of_those_never_discussed34.0of_those_never_discussed: 34of_those_discussed62.0of_those_discussed: 62discussed_no_entry_total80.0discussed_no_entry_total: 80absent_from_headline_fiel…96.0absent_from_headline_fields: 96own_entry103.0own_entry: 103
split of 217 — the numbers behind the chart
{
  "own_entry": 103,
  "discussed_no_entry_total": 80,
  "absent_from_headline_fields": 96,
  "of_those_discussed": 62,
  "of_those_never_discussed": 34
}
never discussed 34
[
  "ABT1",
  "ANKRD37",
  "ANOS1",
  "BTG3",
  "CDK15",
  "CT45A6",
  "EDA",
  "FAM133A",
  "FGFBP1",
  "FMR1",
  "GDF5",
  "GRK6",
  "H1-5",
  "IGHMBP2",
  "IRF2",
  "KRR1",
  "LRRTM4",
  "MAGEA8",
  "MED10",
  "NFKBIL1",
  "OASL",
  "PLCD4",
  "PPP1R13B",
  "PRKD2",
  "PRKRA",
  "RASA2",
  "RL3_RL3L_HUMAN",
  "RNF151",
  "RXRB",
  "SMOC1",
  "SYT2",
  "TMCO1",
  "TMEM214",
  "UTP3"
]
discussed quality check
{
  "ran": "before this entry shipped, because its own counter-evidence called it the cheapest refutation of the main practical claim",
  "n_discussed": 62,
  "named_in_prose_title_summary_section_or_claim": 60,
  "incidental_only": 2,
  "incidental_only_list": [
    "MAPT",
    "PIP4K2C"
  ],
  "median_distinct_entries_mentioning_each": 3,
  "traceable_to_the_cycle72_summary_entry": 1,
  "what_it_refuted": "my claim that the 62 are substantially a backlog of unsplit summary entries. Only 1 of 62 traces to cycle 72's ten-partner entry; the other nine acquired their own entries in cycle 92 and left the list.",
  "what_it_strengthened": "the cost-to-complete ordering. 60 of 62 have prose about them in a median of 3 entries, so the reading really is largely done - for a different reason than I gave.",
  "output": "analyses/core-complex-interactome-gap/discussed_quality.json"
}
Still open
  1. ANSWERED IN THIS ENTRY: 1 of 62, not ten - the summary-entry story is refuted. 60 of 62 are named in prose across a median of 3 entries, which is why the ordering survives. The live question that replaces it: does a prose mention mean an entry could be WRITTEN from what is already there? Sampling ten and attempting a draft is the only way to calibrate the word 'cheaper'.
  2. Is 'discussed' a useful proxy for 'the reading is done'? It is a word-boundary token match, so a symbol in a table cell or an accession string counts. Sampling ten of the 62 and checking whether an entry could actually be written from what is already in the corpus would calibrate the whole ordering.
  3. Should the 34 go first after all? They are the only partners that can contain a surprise, because nobody here has read them. Ordering by cost optimises for finishing the list, and the PI's instruction to saturate the biology is a coverage goal only if the existing reading is sound.
Every source behind this entry (1)
  1. E1established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.analysis

    Of 217 human partners curated by IntAct across the six baits KMT2A, MEN1, PSIP1, AFF4, DOT1L and AFF1, 103 have their own entry. Of the 96 that appear in no entry id, title or alias, 62 are discussed somewhere in the text of at least one node file and 34 are not mentioned anywhere in the corpus. Counting all partners without their own entry, 80 are discussed somewhere. The never-discussed 34 are ABT1, ANKRD37, ANOS1, BTG3, CDK15, CT45A6, EDA, FAM133A, FGFBP1, FMR1, GDF5, GRK6, H1-5, IGHMBP2, IRF2, KRR1, LRRTM4, MAGEA8, MED10, NFKBIL1, OASL, PLCD4, PPP1R13B, PRKD2, PRKRA, RASA2, RL3_RL3L_HUMAN, RNF151, RXRB, SMOC1, SYT2, TMCO1, TMEM214 and UTP3.

    What this evidence cannot show

    'DISCUSSED' IS A WORD-BOUNDARY TOKEN MATCH over whole node files and is therefore an UPPER BOUND on how much reading has really been done. A symbol can appear in a table cell, an accession string or an unrelated sentence and count as discussed, so some of the 62 will turn out to need as much work as the 34. The split was computed with Q128's 14 files on disk, and those files discuss nine partners, so the 62 is flattered by one beat. Symbols that are not genes as written - RL3_RL3L_HUMAN, and TCTE1 which does not resolve in HGNC - sit inside the 217 denominator, so every percentage here is soft by at least two. The claim that the 62 are cheaper is an inference about effort, not a measurement of it: no one has timed promoting a discussed partner against reading a new paper.

    Evidence that argues against this (3)
    • Ordering by cost-to-complete optimises for finishing the list, not for learning anything, and those can diverge sharply. The 34 never-read partners are the only ones that can contain a surprise - by construction, nobody here has looked at them - while the 62 are largely re-packaging conclusions this project has already reached. A supervisor who cared about discovery rather than coverage would reverse this order, and the PI's instruction is to saturate the biology, which is a coverage goal only if you assume the reading behind it is already right.
    • I wrote that the 62 are 'a backlog of summary entries that were never split', generalising from one example. CHECKED BEFORE SHIPPING and it is FALSE: only 1 of 62 traces to cycle 72's ten-partner entry, because the other nine acquired their own entries in cycle 92 and left the list. The practical claim survived the same check and strengthened - 60 of 62 are named in prose across a median of 3 entries - so the ordering stands on a different and better-measured basis than the one I first gave it. What remains untested is whether prose mention means an entry could actually be WRITTEN from what is there; a token match in a section body is not a drafted entry.
    • This entry resolves an impasse by reframing the question, which is a move that can dress up a failure as a finding. The blunt reading is that two cycles of coordinator time went into ranking attempts that produced nothing, and the thing that unstuck it was a field added incidentally while fixing an unrelated defect. The reframing is genuinely useful, but it was not the product of the two cycles that preceded it.
Connected to