Oncofusion

start / graph / ana-q104-the-read-queue-instrument-sees-64-percent-of-intact-and-runs-on-six-baits-not-five

Analysisstrong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.Q104-interactome-read-queue

The interactome coverage figure is computed on 651 of the 1,015 IntAct rows available - 64% - and the gap is worst for menin and LEDGF, the two baits the menin-interface hypothesis rests on

also called Q104 instrument audit · the read queue is a floor · PSICQUIC versus IntAct REST, measured per bait

The number this project uses to say how much of the leukaemia machinery's interaction map it has written is calculated from a database connection that returns about two thirds of the available records. The missing third is worst for the two proteins the project's main hypothesis is about.

The same thing, for a scientist

`analyses/core-complex-interactome-gap/run.py` queries **PSICQUIC**. Measured live on 2026-09-22 against IntAct's own REST `findInteractions` endpoint for the same six accessions [E1]: | bait | PSICQUIC (the instrument) | IntAct REST | ratio | |---|---:|---:|---:| | KMT2A Q03164 | 218 | **376** | 58% | | AFF4 Q9UHB7 | 179 | 207 | 86% | | MEN1 O00255 | 62 | **137** | **45%** | | PSIP1 O75475 | 50 | **114** | **44%** | | AFF1 P51825 | 91 | 104 | 88% | | DOT1L Q8TEK3 | 51 | 77 | 66% | | **total** | **651** | **1,015** | **64%** | **So `union 217 / 95 with an entry / 43.8% / read queue = 15` are all computed on 64% of the record**, and the shortfall is not uniform: it is worst for **MEN1 (45%) and PSIP1 (44%)** - menin and LEDGF, the two proteins this project's central hypothesis and its `ledgf-axis` beat are about. **This graph already knows.** Rule 93 in the standing brief records the same defect, measured at the coordinator, with the menin-KMT2A case worked: PSICQUIC returns 3 menin-KMT2A records with **zero** direct-interaction (MI:0407) rows where REST returns 16 with **seven** direct ITC, x-ray and molecular-sieving records. **The rule was written and the instrument was never migrated.** Two corrections to premises follow. **First, the instrument is NOT making rule 93's exact mistake**: run.py queries the bare accession as a free-text PSICQUIC query, not `id:<accession>` - which is why KMT2A returns 218 rather than the 161 rule 93 records for the `id:` form. It is a better query on a worse endpoint, and the 64% figure is the honest measurement of what remains [E1]. **Second, this beat's dispatch described the instrument as having five baits - KMT2A, AFF1, AFF4, MEN1 and PSIP1. It has SIX. DOT1L is in `BAITS` and contributes 24 human partners** [E2]. And a third defect, structural: **the instrument computes all 217 partners and persists only the 122 missing ones**, discarding the identities of the 95 with entries one line before the write - which is precisely the failure rule 90 was written about, in the instrument that measures the PI's first-stated priority [E3].

Go deeper
The endpoint, per bait, and why the shape of the gap matters more than its size

A 64% overall figure would be tolerable as a stated floor. The problem is the distribution 1:

  • AFF1 (88%) and AFF4 (86%) are nearly complete. These are the elongation-side scaffolds.
  • MEN1 (45%) and PSIP1 (44%) are less than half. These are menin and LEDGF.
  • KMT2A (58%) sits between, and is the largest absolute gap: 158 missing rows.

So the instrument is systematically better at the partner-side scaffolds than at the menin-interface proteins, and the menin-interface hypothesis is what this whole project turns on. Any statement of the form "menin's interactome is N% covered" is currently drawn from a sample that omits 55% of menin's curated rows, and rule 93 already established that what the omitted rows contain is not random: the direct-interaction records - MI:0407, the ITC and crystallographic ones - are among them.

What this does NOT mean. It does not mean the 95 written entries are wrong, and it does not mean the read queue's fifteen members are spurious - every one of them was verified row by row in this beat and every row exists. It means the queue is a floor of unknown depth: the true number of unwritten, multiply-supported partners is at least 15 and the instrument cannot say how much more.

And it means the direction of a future correction is predictable. Migrating to REST will make the denominator grow faster than the numerator, so the headline coverage figure will FALL while the graph gets strictly better. Rule 93's fourth clause requires that such a change be announced as an instrument change carrying both numbers rather than as a corrected figure; this entry exists so that the before-number is on the record with a date.

Three premise corrections, with the measurements

1. The instrument is not making rule 93's exact mistake, and the rule's number does not transfer. Rule 93 measures id:Q03164 at 161 rows. run.py uses the URL form .../search/query/%s?format=tab25&maxResults=4000 with the bare accession, which is a free-text query across fields, and it returns 218 - more than the id: form, less than REST's 376 1. A beat that had quoted rule 93's 161 against this instrument would have overstated the defect by 57 rows. The defect is the endpoint, not the query form, and it should be reported with the number actually measured against this instrument.

2. Six baits, not five. This beat's dispatch stated the instrument covers "human proteins with an IntAct interaction against KMT2A, AFF1, AFF4, MEN1 or PSIP1". BAITS in run.py is 2:

("KMT2A", "Q03164"), ("MEN1", "O00255"), ("PSIP1", "O75475"),
("AFF4", "Q9UHB7"), ("DOT1L", "Q8TEK3"), ("AFF1", "P51825")

DOT1L is the sixth, with 51 PSICQUIC rows and 24 human partners, 17 of which have their own entry (70.8% - the best-covered bait of the six). It is also the bait that contributes H3C13 to the union, via a methyltransferase assay - the enzyme-on-its-own-substrate row that is the clearest single demonstration that this census counts substrates as partners. A five-bait description of a six-bait instrument understates its scope and omits the bait that produces the most instructive artefact in it.

3. The instrument discards the identities it computes. run.py builds rows_out for all 217 partners, then writes union_with_own_entry: sum(1 for r in rows_out if r["own_entry"]) and persists only missing = [r for r in rows_out if not r["own_entry"]] 3. The 95 identities exist in memory and are thrown away at the write.

This is rule 90 verbatim - "the identities were computed and thrown away... a count you cannot re-derive is a claim, not a measurement" - and rule 90's worked example is the previous version of this same measurement, whose stale figure suppressed interactome work for seven cycles. The fix is the same one line rule 90 asks for: persist rows_out, or at minimum the 95 symbols and their node ids. The practical cost right now is that the question "how many of the 95 written partners rest on prey-side rows only?" - the natural follow-up to this beat's direction audit - cannot be asked without re-running the instrument.

Credit where it is due: the same file already does the hard version of this right. It reports unresolved_participants_total: 116 with the distinct forms named, precisely because that skip had been silent and had hidden 47 fusion-construct rows until cycle 54, and it replaced a hardcoded generated date after a stale stamp nearly caused a misreading. The discipline is present; it was applied to the skipped rows and not to the covered ones.

What the read queue's numbers should be quoted as

Until the endpoint is migrated, any quotation of this instrument's output should carry the qualifier. Concretely:

Interactome coverage: 95 of 217 union partners of six core proteins have their own entry (43.8%), measured 2026-09-22 against IntAct via PSICQUIC, which returns 64% of the rows IntAct REST returns for the same six accessions (45% for MEN1, 44% for PSIP1). Read queue = 15, of which 2 were already adjudicated in cycle 54 and 1 is a substrate blocked in the registry since 2026-09-20. The figure is a floor.

That sentence is longer than the number it replaces and every clause in it is load-bearing. Three of them - the endpoint ratio, the per-bait shortfall and the adjudicated-but-requeued members - were not available before this beat, and the fourth, the substrate block, was in the registry and not in the figure.

The order of operations for a future beat or for the coordinator:

  1. Migrate run.py to IntAct REST findInteractions, and announce it with both numbers rather than shipping a corrected figure (rule 93 clause 4).
  2. Persist rows_out in full, not just missing (rule 90).
  3. Add an adjudicated set so documented rejections stop re-queueing, reading it from the ana-* rejection entries or from data/symbol-collisions.json's block_autogeneration flag, which already does the job for H3C1 and which scripts/check-named-but-unwritten.py already consults.
  4. Request MITAB2.7 rather than 2.5, so the experimental-role, host-organism and MI-score columns are available. The instrument currently cannot see that every row in its own queue is prey-side, because tab25 does not carry the column.
  5. Consider excluding substrate-class rows by detection method rather than by symbol - a methyltransferase assay against a histone is an enzyme-substrate measurement whatever the symbol.
Facts and figures
scriptanalyses/core-complex-interactome-gap/run.py (audited, not modified)
statusrun
headlineThe interactome coverage figure and the read queue are computed on 651 of 1,015 available IntAct rows (64%), and the shortfall is worst for MEN1 (45%) and PSIP1 (44%).
what the instrument already does rightIt reports unresolved_participants_total (116) with the distinct participant forms NAMED, because that skip was silent and had hidden 47 fusion-construct rows until cycle 54; and it replaced a hardcoded generated date after a stale stamp nearly caused a refreshed 41.0% to be read as an old 21.2%. The discipline is present in the file - it was applied to the skipped rows and not to the covered ones.
epigenetic readoutOne instrument consequence is epigenetic and concrete: because DOT1L is a bait and the census counts any curated interaction, the H3.2 cluster gene H3C13 enters the union as a DOT1L partner via a METHYLTRANSFERASE ASSAY - the H3K79 methyltransferase acting on its own substrate. An interactome instrument that cannot distinguish an enzyme-substrate assay from a partnership will always import the histone-mark layer into the partner census.
localizationNot applicable - this entry audits an instrument, not a protein. Recorded explicitly rather than left blank.
subnuclear localizationNOT APPLICABLE, and recorded as such rather than as a gap: the subject is a database query and a Python script. The one localisation-adjacent consequence of the endpoint defect is that the rows known to be missing from PSICQUIC include the taspase1 processed-chain identifiers (Q03164-PRO_0000390949 and -390950), whose UniProt location comments are the fragment-scoped ones this graph's rule 91 flags as the interesting ones - so the endpoint defect and the localisation-annotation defect touch the same two identifiers.
cell typeNot applicable to the instrument. Measured for the rows it returns, and reported in ana-q104-every-row-in-the-read-queue-is-prey-side-and-seven-members-are-rejected: HEK293T 16, HeLa S3 4, K562 3, HEK293 Flp-In 1, and zero B-lineage or fusion-expressing hosts.
cell stateNot applicable to the instrument. No row it returns for the read queue carries a cell-state condition - every one is an unperturbed proliferating culture.
shown for fusionNO, and structurally so: the instrument queries canonical accessions, and rule 93 established that PSICQUIC cannot match the processed-chain identifiers Q03164-PRO_0000390949 and -390950 - the taspase1 N320 and C180 fragments - which are the species this project argues about. Fusion-construct interactors were separately found to be invisible to accession queries altogether (47 rows, cycle 54).
inputs
[
  "https://www.ebi.ac.uk/intact/ws/interaction/findInteractions/{accession}?page=0&pageSize=1",
  "https://www.ebi.ac.uk/Tools/webservices/psicquic/intact/webservices/current/search/query/{accession}?format=tab25|tab27",
  "analyses/core-complex-interactome-gap/run.py",
  "analyses/core-complex-interactome-gap/results.json (generated 2026-09-22)"
]
endpoint comparison
{
  "measured": "2026-09-22",
  "per_bait": {
    "KMT2A (Q03164)": {
      "psicquic": 218,
      "rest": 376,
      "pct": 58
    },
    "AFF4 (Q9UHB7)": {
      "psicquic": 179,
      "rest": 207,
      "pct": 86
    },
    "MEN1 (O00255)": {
      "psicquic": 62,
      "rest": 137,
      "pct": 45
    },
    "PSIP1 (O75475)": {
      "psicquic": 50,
      "rest": 114,
      "pct": 44
    },
    "AFF1 (P51825)": {
      "psicquic": 91,
      "rest": 104,
      "pct": 88
    },
    "DOT1L (Q8TEK3)": {
      "psicquic": 51,
      "rest": 77,
      "pct": 66
    }
  },
  "total_psicquic": 651,
  "total_rest": 1015,
  "overall_pct": 64,
  "worst_baits": [
    "MEN1 45%",
    "PSIP1 44%"
  ],
  "why_the_shape_matters": "The instrument is systematically better at the elongation-side scaffolds (AFF1 88%, AFF4 86%) than at menin and LEDGF (45%, 44%) - the two proteins the menin-interface hypothesis and the ledgf-axis beat are about.",
  "not_strictly_commensurable": "PSICQUIC MITAB returns one line per binary-interaction evidence row; REST findInteractions reports totalElements over its own binary-interaction index. The ratio is a strong indication of substantial under-coverage, not a precise count of missing evidence rows.",
  "what_is_known_to_be_missing": "Rule 93, measured at the coordinator: the processed-chain identifiers Q03164-PRO_0000390949 and -390950 (the taspase1 N320 and C180 fragments this project argues about) and the MI:0407 DIRECT-INTERACTION records for the menin-KMT2A contact - PSICQUIC returns 3 menin-KMT2A records with zero direct rows, REST returns 16 with seven ITC, x-ray and molecular-sieving records."
}
premise corrections
[
  {
    "premise": "The instrument queries PSICQUIC as id:<accession>, per rule 93, and therefore sees 161 KMT2A rows.",
    "correction": "FALSE. run.py uses the bare accession in the query path, which is a free-text query, and returns 218. The instrument is a BETTER query on a WORSE endpoint. Quoting rule 93's 161 against this instrument would overstate the defect by 57 rows.",
    "measured": "PSICQUIC bare-accession query for Q03164 = 218 rows, 2026-09-22"
  },
  {
    "premise": "The read-queue instrument covers five baits: KMT2A, AFF1, AFF4, MEN1, PSIP1 (this beat's dispatch brief).",
    "correction": "FALSE. It has SIX. DOT1L/Q8TEK3 is in BAITS, contributes 51 rows and 24 human partners, and is the BEST-covered bait at 70.8% own-entry. It is also the bait that contributes H3C13 to the union by methyltransferase assay - the enzyme-on-its-own-substrate row that best demonstrates the census counts substrates as partners.",
    "measured": "run.py BAITS list read directly, 2026-09-22"
  },
  {
    "premise": "results.json records which partners the project has covered.",
    "correction": "FALSE. rows_out is built for all 217 and only the 122 missing are persisted; union_with_own_entry is written as a bare count. The 95 identities are discarded one line before the write. This is rule 90's defect in the instrument that measures the PI's first-stated priority, and rule 90's own worked example is the previous version of this same measurement.",
    "measured": "run.py rows_out/missing construction and the complete key list of results.json, read directly 2026-09-22"
  }
]
consequences for quotation
{
  "the_number_as_currently_quotable": "95 of 217 union partners of SIX core proteins have their own entry (43.8%), measured 2026-09-22 against IntAct via PSICQUIC, which returns 64% of the rows IntAct REST returns for the same six accessions (45% for MEN1, 44% for PSIP1). Read queue = 15, of which 2 were already adjudicated in cycle 54 and 1 is a substrate blocked in the registry since 2026-09-20. The figure is a FLOOR.",
  "direction_of_a_future_correction": "Migrating to REST will grow the denominator faster than the numerator, so the headline coverage figure will FALL while the graph gets strictly better. Rule 93 clause 4 requires announcing that as an instrument change carrying both numbers, not as a corrected figure. This entry puts the before-number on the record with a date so that is possible."
}
recommended fixes in order
[
  "1. Migrate run.py to IntAct REST findInteractions, and announce it with BOTH numbers (rule 93 clause 4).",
  "2. Persist rows_out in full, not just missing - the 95 covered symbols and their node ids (rule 90).",
  "3. Add an ADJUDICATED set so documented rejections stop re-queueing; read it from the ana-* rejection entries or from data/symbol-collisions.json's block_autogeneration flag, which already does this for H3C1 and which scripts/check-named-but-unwritten.py already consults. Currently 2 of 15 queue slots are already-answered questions.",
  "4. Request MITAB2.7 (or move to REST JSON) so experimental role, host organism and MI-score are available. The instrument currently cannot see that EVERY row in its own queue is prey-side.",
  "5. Consider excluding substrate-class rows by DETECTION METHOD rather than by symbol regex - a methyltransferase assay against a histone is an enzyme-substrate measurement whatever the symbol, and a method filter would also catch non-histone substrates a prefix match misses."
]
How this could bite us back
  • Publishing a coverage figure with a 64% qualifier attached invites the qualifier to be dropped in the next summary, which is how the previous version of this number went stale and suppressed interactome work for seven cycles. The safest form is to keep the endpoint ratio inside the sentence that carries the percentage, as this entry's suggested wording does, rather than in a caveat field.
  • Migrating the endpoint will make the headline number fall. A reader who sees coverage drop from 43.8% will read it as the project losing ground, and the honest explanation - the instrument got better - is exactly the explanation that sounds like an excuse. It has to be announced before the migration, not after.
  • This entry is itself a number-in-the-header risk: '64%' is a figure produced by one ad-hoc comparison in one session, with no script that regenerates it. Rule 90 says such a figure needs either a regenerating script or a visible as-of stamp. It has the stamp (2026-09-22) and not the script, and it should not be quoted without the date.
  • Fixing the queue filter to exclude adjudicated members will SHRINK the read queue from 15 to 12 without any new work being done, which is the mirror image of the coverage figure falling. Both changes make the numbers move for reasons unrelated to research effort, and both need announcing as such.
Still open
  1. What is the union, the covered count and the read-queue size against IntAct REST? That is the measurement this entry says is needed and does not provide - it requires re-running the instrument, which this beat did not do because changing a shared analysis mid-cycle while another agent is writing is the wrong move.
  2. What is in the 364 rows PSICQUIC does not return? Rule 93 names two classes - processed-chain identifiers and MI:0407 direct-interaction records - but nobody has enumerated them. If they contain multiply-supported partners with no entry, the read queue is materially larger than 15.
  3. How many of the 95 covered partners rest on prey-side rows only? Unanswerable today because the instrument discards their identities, and it is the natural follow-up to this beat's direction audit.
  4. Does the per-bait shortfall correlate with anything other than endpoint behaviour - curation age, source database, interaction type? MEN1 at 45% and AFF1 at 88% is a large spread for one endpoint against one database, and the cause has not been established.
  5. Are there other analyses in this repository still querying PSICQUIC? Rule 93 says both of this project's interactome analyses do. If the other one also feeds a quoted figure, the same qualifier applies to it and nobody has checked.
Every source behind this entry (4)
  1. E1strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.database

    Measured live on 2026-09-22, IntAct REST findInteractions returns totalElements of 376 for Q03164 (KMT2A), 207 for Q9UHB7 (AFF4), 137 for O00255 (MEN1), 114 for O75475 (PSIP1), 104 for P51825 (AFF1) and 77 for Q8TEK3 (DOT1L) - 1,015 in total - against the 218, 179, 62, 50, 91 and 51 rows (651 total, 64%) that analyses/core-complex-interactome-gap/run.py obtains from PSICQUIC for the same accessions; the shortfall is worst for MEN1 (45%) and PSIP1 (44%), and a direct PSICQUIC query for the bare accession Q03164 returns 218 rows rather than the 161 that the id:Q03164 form returns.

    What this evidence cannot show

    Both endpoints queried live 2026-09-22. THE TWO COUNTS ARE NOT STRICTLY THE SAME QUANTITY and this is the item's main limitation: PSICQUIC MITAB returns one line per binary-interaction EVIDENCE row, while REST findInteractions reports totalElements over binary interactions in its own index, and the two may differ in how they expand complexes, self-interactions and replicate evidences. The ratio is therefore a strong indication that the instrument sees substantially less than the database holds, and NOT a precise statement that exactly 364 evidence rows are missing. What is independently established, by rule 93 measured at the coordinator, is that the missing rows include the processed-chain identifiers (Q03164-PRO_0000390949 and -390950, the taspase1 N320 and C180 fragments) and the MI:0407 direct-interaction records for the menin-KMT2A contact - so the omission is not random with respect to the rows this project cares about. The PSICQUIC per-bait counts are run.py's own reported intact_rows from results.json generated the same day, re-derived for KMT2A by direct query.

    Evidence that argues against this (1)
    • Migrating the endpoint has a cost that cuts against doing it casually: it changes every n unevenly between baits, and because the denominator (union partners) will grow faster than the numerator (partners with entries), the headline coverage figure will FALL while the graph is strictly better informed. Rule 93's fourth clause exists for this case and requires the change to be announced as an instrument change carrying both numbers; two analyses in this repository already open with such a banner. this repository 2026
  2. E2established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.dataset

    analyses/core-complex-interactome-gap/run.py defines BAITS as six entries - KMT2A/Q03164, MEN1/O00255, PSIP1/O75475, AFF4/Q9UHB7, DOT1L/Q8TEK3 and AFF1/P51825 - and results.json records DOT1L with 51 IntAct rows, 24 human partners and 17 of them having their own entry (70.8%, the highest own-entry percentage of the six baits); DOT1L is also the bait contributing H3C13 to the union, by methyltransferase assay.

    What this evidence cannot show

    Read directly from the source file and from results.json by this beat on 2026-09-22. THE URL IS A PLACEHOLDER - both sources are files in this repository with no external URL; the paths are the identifiers and are given in the title. This is a correction to a premise in THIS BEAT'S OWN DISPATCH BRIEF, which described the instrument as covering five baits; it is not a defect in the instrument, which is doing what its source says. The 70.8% own-entry figure for DOT1L is results.json's own computed field and was not re-derived.

    Evidence that argues against this (1)
    • The omission has a direction worth noting: DOT1L is the BEST-covered bait of the six at 70.8% own-entry, so describing the instrument as five-bait understates its scope while making its coverage look slightly worse than it is. It is a small error in the conservative direction on the coverage figure and a larger one on what the instrument can be asked - a five-bait description cannot answer any DOT1L question. this repository 2026
  3. E3established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.dataset

    analyses/core-complex-interactome-gap/run.py builds rows_out containing all 217 union partners with their baits, row counts, publications, methods and own_entry node id, then persists only the missing subset - missing = [r for r in rows_out if not r['own_entry']] - while writing union_with_own_entry as a sum over rows_out, so the identities of the 95 partners that DO have entries are computed and discarded at the write, and cannot be re-derived from results.json.

    What this evidence cannot show

    Read directly from the source file by this beat on 2026-09-22 and confirmed against results.json, whose top-level keys are generated, baits, unresolved_participants_total, unresolved_participant_forms, unresolved_note, union_human_partners, union_with_own_entry, union_mentioned_only, union_absent_entirely, read_queue_size, read_queue, all_missing and note - there is no field containing the 95 covered partners. THE URL IS A PLACEHOLDER; the file path is the identifier. This is rule 90's defect, and rule 90's own worked example is the PREVIOUS version of this same measurement, so this is a recurrence in the same analysis family rather than a new class. The per-bait blocks DO record has_own_entry counts, so the aggregate is re-derivable per bait; it is the SYMBOLS that are lost.

    Evidence that argues against this (1)
    • The same file demonstrates the discipline it is missing here, which argues this is an oversight rather than a design choice: it reports unresolved_participants_total with the distinct participant forms named, precisely because that skip had been silent and had hidden 47 fusion-construct rows until cycle 54, and it replaced a hardcoded generated date after a stale stamp nearly caused a refreshed 41.0% to be read as an old 21.2%. The author of those fixes would have persisted rows_out if asked. this repository 2026
  4. E4strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.database

    PSICQUIC MITAB2.5 - the format analyses/core-complex-interactome-gap/run.py requests - does not carry the experimental-role, host-organism or MI-score columns, which MITAB2.7 does; consequently the instrument structurally cannot observe that in all 24 read-queue rows inspected the candidate is the experimental bait and this project's scaffold protein is the prey.

    What this evidence cannot show

    Established by this beat on 2026-09-22 by running the same query in both formats and reading the returned columns: the tab25 response stops at the confidence column while tab27 continues through biological role, experimental role, interactor type and host organism. The column INDICES used to read roles (19, 20), host (29), method (7) and MI-score (15) were determined by inspecting the returned values rather than from a specification document, and the readback control is that the role values are the expected controlled-vocabulary terms MI:0496(bait) and MI:0498(prey) and the hosts are recognisable cell-line names - so the indices are right, but they were verified empirically and not against the PSI-MITAB2.7 spec.

    Evidence that argues against this (1)
    • Requesting tab27 is not free: the responses are substantially larger, the instrument already fetches up to 4,000 rows per bait, and the extra columns are sparsely populated for older curation. A migration should therefore be to REST rather than to tab27 - REST carries the same fields in JSON and fixes the 64% shortfall at the same time, so the tab27 finding is best read as a reason to migrate the endpoint rather than the format. EMBL-EBI IntAct 2026
Connected to