Oncofusion

start / graph / ana-named-but-unwritten-the-queue-behind-four-beats

Analysisunrated confidence 0 of 5: unrated

Four beats in a row were picked by noticing the same shape by hand; this is the instrument that finds it - 171 gene symbols named in six or more entries with no entry of their own

also called named but unwritten · the mechanism is written and the proteins are not · standing gene-entry queue

We kept noticing that this project writes up a mechanism and never writes down the proteins inside it. That had been spotted four times by hand. This finds them all at once, and the top of the list is the cell-surface markers the whole disease is diagnosed and treated by.

The same thing, for a scientist

Four consecutive beats were chosen because a coordinator grepped on a hunch after reading a hand-back: **HDAC named in 81 entries with 0 written, EZH2/PRC2 in 63 with 0, TOP2A/TOP2B in 90 with 0, KMT2B in 57 with 0.** Each time the mechanism existed and the proteins named inside it did not. **That is an instrument, not an insight.** `scripts/check-named-but-unwritten.py` tokenises every node file, keeps only tokens that are **HGNC-approved protein-coding symbols**, and reports those with no `gene-`, `protein-` or `protein-*-domains` entry. Of **19,297** HGNC symbols, **242** have an entry here; **171 are named in six or more entries and have none**. **It independently rediscovered two of the four beats it was built from** - KMT2B (32 files), KMT2C (26), KMT2D (24), CTCF (33) and ATR (31) are all in the queue and all are being written by agents dispatched this cycle on a hunch. That is the validation that matters: the hunch was reproducible and now does not need to be had.

Go deeper
The top of the queue is not the chromatin layer, and that is the finding
symbolentries naming itHGNC name
CD34127CD34 molecule
CD19114CD19 molecule
MYC95MYC proto-oncogene
TP5384tumor protein p53
KRAS44KRas proto-oncogene
NRAS36NRAS proto-oncogene
PICALM34phosphatidylinositol binding clathrin assembly protein
CTCF33CCCTC-binding factor
KMT2B32lysine methyltransferase 2B
MCL132MCL1 apoptosis regulator
ATR31ATR checkpoint kinase
CD3329CD33 molecule
DNMT3A23DNA methyltransferase 3 alpha
SMARCA421BAF chromatin remodeling ATPase
IKZF118IKAROS family zinc finger 1
PAX518paired box 5

CD34 and CD19 lead the whole queue and neither has an entry. CD19 is the target of blinatumomab and of CAR-T, both of which this graph covers in detail, and the antigen whose loss defines the lineage-switch escape it keeps returning to. CD34 is how the cell of origin is defined in every experiment here. The graph has entries for the therapies and the escape and not for the molecule they act on.

The chromatin layer this project has spent four beats on is mid-table. That is worth sitting with: the hand-found gaps were all in the subject the coordinator was already thinking about, and the largest gaps are in the part nobody was looking at.

Two symbol collisions that would corrupt any symbol-keyed automation here

PRC1 — 78 entries. This graph means Polycomb Repressive Complex 1. HGNC's PRC1 is protein regulator of cytokinesis 1, an unrelated mitotic spindle protein. Every one of those 78 mentions is the complex; none is the gene. An automated 'write the missing gene' pass would have created gene-prc1 for the wrong protein, and it would have looked plausible - a cell-cycle protein in a leukaemia graph.

PZP — 21 entries. This graph means the PHD-Zinc finger-PHD cassette of KMT2A. HGNC's PZP is PZP alpha-2-macroglobulin like, a plasma protein.

Both are now the clearest examples in this project of a hazard rules 62, 63 and 65 only partly cover: those rules are about a query returning the wrong record. Here the graph's own vocabulary collides with the nomenclature, and no database lookup can detect it because both readings resolve.

The instrument reports them rather than filtering them, because a filter would hide the hazard from the next reader.

What the count cannot do, stated before anyone quotes it

Frequency is not importance. REST tops the raw list at 195 entries and is the English word; EFS at 123 is event-free survival; ACTB and GAPDH are loading controls. The HGNC filter removes the acronyms that are not symbols at all, and cannot remove the symbols that are also ordinary words.

And a high count can mean the opposite of a gap. A protein named in sixty entries because one roster was copied sixty times is less of a gap than one named in five because it is the subject of a single decisive paper. The count ranks the queue; it does not justify an entry.

Some absences are deliberate and the instrument must be told. This graph files histones as marks and interfaces, which is the right unit for a chromatin reader, so H3C1 is carried in an explicit exclusion list with its reason rather than being flagged every run. An instrument that keeps proposing a known false positive is wrong rather than persistent - a prior analysis in this project had H3C1 at rank 2 of its queue and had to re-adjudicate it by hand each time.

Still open
  1. Read the top twenty and decide which deserve entries. CD34 and CD19 lead the queue, are the molecules this disease is diagnosed and treated by, and have no entries while the therapies targeting them do.
  2. Add a second axis - how many DISTINCT beats name a symbol, rather than how many entries. A symbol named across many beats is load-bearing; one named sixty times inside one beat is a copied roster.
  3. Run the same instrument over non-gene things: complexes, cell lines, drugs and trials. The PRC1 collision suggests the complex layer has the same shape.
  4. Record the symbol collisions somewhere an automated pass would read before writing entries. A future agent told to 'write the missing genes' would create gene-prc1 for a mitotic spindle protein, and it would look plausible.
Every source behind this entry (3)
  1. E1established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.analysis

    Of 19,297 HGNC protein-coding symbols, 242 have an entry of their own in this graph, and 171 symbols are named in the text of six or more entries without one. The highest-ranked are CD34 (127 entries), CD19 (114), MYC (95), TP53 (84), KRAS (44) and NRAS (36).

    What this evidence cannot show

    Run 2026-09-20 against the graph on disk while two agents were writing, so it is a snapshot and files mid-write are skipped rather than counted. Established as a fact about the corpus. **The count is of entries mentioning a symbol, not of mentions**, so a symbol named ten times in one entry counts once - which is the right unit for 'how much of the graph talks about this' and the wrong one for 'how load-bearing is it'.

    Evidence that argues against this (1)
    • The queue contains symbols that are ordinary English words (REST at 195, EFS at 123) and laboratory boilerplate (ACTB, GAPDH), which the HGNC filter cannot remove because they are genuinely approved symbols. Any number quoted from this without reading the rows will be wrong in the same way the earlier interactome count was. HUGO Gene Nomenclature Committee 2026
  2. E2established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.database

    The graph's own vocabulary collides with HGNC nomenclature in at least two places: PRC1, named in 78 entries as Polycomb Repressive Complex 1, is the HGNC symbol for protein regulator of cytokinesis 1; and PZP, named in 21 entries as the KMT2A PHD-Zinc finger-PHD cassette, is the HGNC symbol for PZP alpha-2-macroglobulin like.

    What this evidence cannot show

    Both collisions were confirmed by reading the HGNC approved name against the graph's usage on 2026-09-20. **Neither is detectable by any lookup**, because both readings resolve to real records - which is what distinguishes this from the traps in rules 62, 63 and 65, where a query returns a wrong record that a careful reader can catch.

    Evidence that argues against this (1)
    • Two collisions found in a queue of 171 is a floor, not a count. The instrument only sees symbols the graph happens to use; a collision on a symbol this graph does not mention would not appear, and the rate among the 19,297 HGNC symbols is unmeasured. Oncofusion knowledge graph (internal) 2026
  3. E3strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.analysis

    The instrument independently returns four targets that were dispatched this cycle on a coordinator's hunch - CTCF, ATR, KMT2B, KMT2C and KMT2D are all in the queue - which is the reproducibility check on the hand-found method it replaces.

    What this evidence cannot show

    **Strong, not established, because it is a check on the method and not on the biology.** Agreeing with a hunch shows the hunch was reproducible; it does not show either was right about what deserves an entry. The four beats those hunches produced have to be judged on their entries, not on this agreement.

    Evidence that argues against this (1)
    • The agreement is partly circular: the instrument was written after those beats and its exclusion list and thresholds were chosen by the same coordinator who had the hunches. An instrument tuned until it reproduces the answer already reached has demonstrated tuning, not discovery. Oncofusion knowledge graph (internal analysis) 2026
Connected to