PrimeKG merged nodes: review table for #438 item 2

Refs #438 item 2. Richard decided on 2026-10-05 to “fix all the wrong merges”, starting from this table. Related: roadmap D13 (#611), the Jaccard safety report (#59). The fix lands before Run 3’s export.

Decided 2026-10-05: D-Q4-22 A (#819). Rules J0–J9 were approved as proposed, and the six unclear groups split along the CSV’s partition column, as the review leans. kg_split_primekg_merges.py carries it out (#832).

With that partition the split restores 109,412 rows, not the 109,440 quoted below. §4’s ”+ the 6 unclear” column splits D-glucose from dextrose, and J5 keeps those two together. The difference is 28 enzyme and transporter rows. Fed the approved partition, the replay in §8.5 gives the same 109,412.

Nothing here changes data. Whether two nodes are one entity is a mapping decision, so under the CLAUDE.md hard rule this table goes to Richard before any entity is split. Each verdict is a proposal. Where a verdict rests on a rule rather than on evidence, the rule is named (J0–J9, §3), and those rules are what he approves. Counts were read from prod, read-only and PrimeKG rows only, at 15:06 PDT on 2026-10-05, and replayed from PrimeKG’s own kg.csv.

Files. This page, and kg_primekg_merge_review.csv beside it: one row per merged node (401 rows), with every measured column and the proposal. §8 has the two scripts that produce it.

Summary

• 401 PrimeKG nodes sit on 183 canonical entities. The audit’s “218” is the surplus: 401 − 183 = 218 nodes that lost an entity of their own. By node type:
  • drug: 237 nodes in 101 groups, a surplus of 136 (the audit’s “136”);
  • gene: 122 nodes in 61 groups;
  • pathway: 36 nodes in 18 groups;
  • anatomy: 4 nodes in 2 groups;
  • disease: 2 nodes in 1 group.
• The audit’s edge counts reproduce exactly. Replaying kg.csv under these merges gives 7,988,012 relations. That is what prod stores, on every one of the 33 relation keys and on every one of the 183 entities. The merges collapsed 112,486 rows: drug_drug 108,174, contraindication 1,566, indication 428, and nine smaller relations (§1).
• Three mechanisms made the groups (§2).
  • 63 groups (61 gene, 2 anatomy): PrimeKG ships two nodes for one id. Either two symbols of one gene (C16orf58 and RUSF1, both NCBI Gene 64755) or two labels of one UBERON term. These were one entity from the first backfill. The merges are right and cost no rows.
  • 101 groups (all drugs): Node Normalizer gave their DrugBank ids one preferred CURIE. apply_canonical_entity_merges.py pass 1 then merged them in May 2026. Node Normalizer folds salts, esters, enantiomers and radiotracers into one chemical, and twice it folded together unrelated drugs.
  • 19 groups (18 pathway, 1 disease): the names matched, and the same script’s pass 2 merged them. Ten of them join a SARS-CoV-1 pathway to its SARS-CoV-2 namesake.
• Proposed verdicts: 110 distinct, 67 same, 6 unclear.
  • Same (67): the 63 one-id groups, two hydrate or “unspecified form” pairs (J5), and two pairs that are one product under two DrugBank entries (J8): CT-011 with pidilizumab, and PEG-uricase with pegloticase.
  • Distinct (110): 91 drug groups, all 18 pathway groups and the disease group. 29 are distinct on evidence alone: unrelated drugs, different molecules, different viruses. The other 81 rest on one of J1–J4, J6, J7 or J9.
  • Unclear (6): drug groups where the evidence does not settle it (§3).
  • Both known examples are distinct. D- and L-tartaric acid are enantiomers (J1). Choline C-11 and choline salicylate sit in a group of four different drugs.
• What a split restores. Splitting the 110 distinct groups restores 106,818 rows. That includes every lost contraindication (1,566), indication (428), drug_effect (888) and off-label use (46) row. Splitting the 6 unclear groups as well restores 109,440. The remaining 3,046 rows stay merged on purpose. They sit in the same groups and in the two kept pairs of the sodium phosphate group, where each is one assertion made twice (§4).
• What Richard approves: the ten rules in §3. Each comes with the number of groups it decides. A rejected rule flips its own groups and no others, and the CSV names the rule behind every node.

1. Reconciliation with the audit

Quantity	Audit (#438)	Measured	Why they differ
PrimeKG nodes on a shared entity	218	401	The audit counted the surplus: 401 − 183 = 218.
Entities they form	183	183	—
Drug nodes	136	237, in 101 groups	Also the surplus: 237 − 101 = 136.
drug_drug rows lost	108,174	108,174	—
contraindication rows lost	1,566	1,566	—
indication rows lost	428	428	—
drug–protein rows lost	1,272 (2026-10-03 comment)	1,250	1,272 counted (drug, protein) pairs, as stored before #696. 1,250 counts the role rows stored since: target 364, enzyme 438, transporter 346, carrier 102.
All rows lost	112,508 (same comment)	112,486	The 22 drug–protein rows above.

How “lost” is counted. A row is lost when two kg.csv rows map to one stored key, (subject, object, predicate, source, source_predicate), after their endpoints are merged. A row whose two endpoints land on one entity becomes a self-loop and is deleted too. The unit is the directed kg.csv row, which is the unit mods_kg_relation stores; PrimeKG writes every edge in both directions.

What prod holds. PrimeKG’s 129,375 nodes are all stored, as 129,375 prime_kg external ids on 129,157 entities. No external id sits on two entities. kg.csv has no duplicate rows and no native self-loops.

Relation	kg.csv rows	Stored on prod	Lost to merges	Of which self-loops
drug_drug	2,672,628	2,564,454	108,174	164
drug_effect	129,568	128,680	888	0
pathway_protein	85,292	85,184	108	0
disease_disease	64,388	64,386	2	2
contraindication	61,350	59,784	1,566	0
target	32,760	32,396	364	0
indication	18,776	18,348	428	0
enzyme	10,634	10,196	438	0
transporter	6,184	5,838	346	0
off-label use	5,136	5,090	46	0
pathway_pathway	5,070	5,046	24	0
carrier	1,728	1,626	102	0
all 33 relation keys	8,100,498	7,988,012	112,486	166

The other 21 relation keys lost nothing. Nothing self-loops on prod today: the 166 self-loop rows were deleted when the merges were made.

2. How the groups were merged

Mechanism	Groups	Nodes	Domains	The data that joined them	Proposed
Backfill: one CURIE hint. PrimeKG has two nodes with the same source id, so both resolve to one CURIE.	63	126	61 gene, 2 anatomy	Both nodes carry the same x_source:x_id in kg.csv, e.g. NCBI:64755 for “RUSF1” and “C16orf58”. They were already one entity in the pre-merge snapshot.	63 same
apply_canonical_entity_merges.py pass 1: Node Normalizer. Entities whose CURIEs have one preferred_curie in mods_kg_curie_cache are merged; the winner takes that CURIE.	101	237	drug	DrugBank:DB09459 → CHEBI:15671 “L-tartaric acid” and DrugBank:DB01694 → CHEBI:15671. Each node sat on its own entity in the pre-merge snapshot.	91 distinct, 4 same, 6 unclear
The same script’s pass 2: name match. An entity Node Normalizer does not know is merged into one of the same name and domain whose CURIE has a canonical prefix.	19	38	18 pathway, 1 disease	Identical names, e.g. two “Maturation of spike protein” pathways, REACTOME:R-HSA-9683686 and R-HSA-9694548; neither is in the cache.	19 distinct

How the second and third rows were dated and attributed:

• The script refuses to run without the snapshot tables mods_kg_entity_premerge, mods_kg_entity_source_premerge and mods_kg_relation_premerge, and prod has all three.
• In that snapshot the 120 pass-1 and pass-2 groups sit on 237 + 38 separate entities. The 63 one-id groups are already one entity each.
• Pass 1’s grouping key, COALESCE(preferred_curie, curie) within a domain, reproduces every drug group. Every member of a drug group has the same cached preferred_curie.
• Commit 9874abb (2026-05-13) records “actually running the merge against ~20K mergers and ~9M relations”.

This is an inference from the snapshot and the code. No run log names these groups.

The script’s first commit, 7f9180b, put an “edge-overlap safety threshold” out of scope, pending “review of the actual merges”. The jaccard_with_anchor column in the CSV is that review, per node.

3. The judgement calls: rules for Richard

Each rule decides a family of groups. Rejecting a rule flips the groups it decides, and only those. A group can rest on more than one rule: Vitamin E (J1, J2, J7), the sodium phosphates (J3, J4, J5) and glucose (J1, J5). Such a group is counted under each rule.

Rule	Proposal	Groups it decides	If rejected¹	Examples
J0	Same id, same entity. Two PrimeKG nodes with one source id (one NCBI Gene id, one UBERON id) are one entity. → same	63 (all same)	63 duplicate nodes return; no row comes back	RUSF1 and C16orf58 (NCBI Gene 64755); ovary and female gonad (UBERON 992)
J1	Stereoisomers are distinct: enantiomers, a racemate and its enantiomer, D and L forms. → distinct	9 (7 distinct, 2 unclear)	3,664 fewer rows restored	L- and D-tartaric acid; ketoconazole and levoketoconazole; norgestrel and levonorgestrel; atropine and hyoscyamine; serine and D-serine; levomenthol and racementhol
J2	An ester or other covalent prodrug is distinct from its parent, and two esters of one parent are distinct from each other. → distinct	34 (all distinct)	67,392 fewer rows restored	hydrocortisone and six esters; testosterone and four; estradiol and four; candesartan cilexetil; benazepril and benazeprilat; gabapentin enacarbil; paclitaxel poliglumex
J3	A salt is distinct from its free acid or base, and from another salt of it, when DrugBank gives each its own entry. Applies where the counter-ion is inert; a second active ingredient is evidence, not J3. → distinct	17 (all distinct)	10,714 fewer rows restored	methylene blue and methylthioninium; sodium oxybate and GHB; atracurium and its besylate; ascorbic acid and sodium ascorbate; tafamidis and Fx-1006A (its meglumine salt); monobasic and dibasic sodium phosphate
J4	A radiolabelled form is distinct from the unlabelled compound, and from the same carrier with another radionuclide. → distinct	6 (all distinct)	2,670 fewer rows restored	choline C-11; ammonia N-13; Lu-177 and Ga-68 dotatate; iobenguane sulfate I-123; sodium phosphate P-32; selenomethionine Se-75
J5	A hydrate, or DrugBank’s “unspecified form” entry, is the same substance as the named compound. → same	5 (2 same, 1 distinct, 2 unclear)	2,308 more rows restored	zinc sulfate; magnesium acetate and its tetrahydrate; both sodium phosphates’ “unspecified form” entries; dextrose and D-glucose (unclear group)
J6	Biologics from different sources are distinct: recombinant, plasma-derived or urine-derived; bovine and other heparin. → distinct	5 (4 distinct, 1 unclear)	648 fewer rows restored	recombinant and plasma-derived factor VIII; follitropin and urofollitropin; thyrotropin alfa and thyrotropin; heparin and bovine heparin
J7	A class, mixture, polymer family or element is distinct from one member, component, grade or compound of it. → distinct	11 (10 distinct, 1 unclear)	7,862 fewer rows restored	opium and morphine; vitamin D with ergocalciferol and cholecalciferol; PEG with PEG 300 and PEG 400; iron and iron dextran; neomycin and framycetin; bentonite and montmorillonite
J8	Two DrugBank entries for one product are the same: a development code and its INN, or an investigational entry and the approved one. → same	2 (all same)	738 more rows restored	CT-011 and pidilizumab; PEG-uricase and pegloticase
J9	Two source ids joined only because their names matched are split back, even when PrimeKG gives them identical edges. → distinct	6 (all distinct)	24 fewer rows restored	six Reactome SLC disease-pathway pairs: each pair shares one failed reaction under two different normal pathways, and Reactome has since renamed them apart
evidence	No rule: the members are different molecules, products, viruses or terms.	29 (all distinct)	15,574 rows ride on these groups	see below

¹ Rows restored by the proposal (106,818) minus rows restored with that rule rejected, from primekg_merge_review.py’s per_rule. Rejecting a split rule keeps its groups merged. Rejecting a merge rule (J0, J5, J8) splits its groups into single nodes. Unclear groups are left out. Vitamin E (J1, J2, J7) and the sodium phosphates (J3, J4, J5) cite several rules and flip whole, so for J1, J2, J3, J4 and J7 the figure is an upper bound.

No rule: evidence alone (29 groups, all distinct). These are:

• 2 groups of unrelated drugs that Node Normalizer conflated: guanabenz with icodextrin, and ethanol with ipecac;
• 12 groups of different molecules or materials:
  • six insulins; three low-molecular-weight heparins; three Tc-99m agents;
  • human and porcine secretin; human and ovine corticotropin-releasing hormone; factor VIII of human and of porcine sequence;
  • isosorbide and its mononitrate; pentaerythritol and its tetranitrate; adenosine and AMP;
  • two ferric pyrophosphate complexes; inositol and inositol nicotinate; almasilate and magnesium aluminium silicate;
• 2 salts whose counter-ion is a second active drug: procaine benzylpenicillin and ambroxol acefyllinate. Choline salicylate is a third, in a group whose split also needs J4;
• 12 Reactome pairs with different events: ten SARS-CoV-1/SARS-CoV-2 pairs and two disease pathways;
• the disease pair, which PrimeKG itself links parent to child.

The six unclear groups. The CSV’s partition column gives the split I lean to for each.

• Calcium phosphate and its dihydrate. DrugBank’s “Calcium Phosphate” does not say which calcium phosphate it is. The dihydrate (MW 172.09) is dicalcium phosphate.
• D-glucose, dextrose, and beta-D-glucose. J5 merges the first two, which carry the same DrugBank text. Beta-D-glucose is an anomer, and anomers interconvert in water.
• Dihydroergocryptine and dihydro-alpha-ergocryptine. DrugBank’s own text calls the alpha form “usually … the mixture”.
• Camphor and synthetic camphor. Neither entry states its stereochemistry.
• Two influenza vaccine antigen groups (A/California/7/2009 X-181, two entries; B/Brisbane/60/2008, four entries). Each group’s entries differ in antigen type or inactivation method and have identical PrimeKG edges.

A caution that cuts both ways. In 38 drug groups, PrimeKG lists drug–drug interactions between two members of the same group: 82 pairs, such as estradiol with each of its esters. The merges deleted those rows as self-loops. For a distinct group they are real assertions, and a split restores them. Three same groups have them too: CT-011 with pidilizumab, magnesium acetate with its tetrahydrate, and PEG-uricase with pegloticase. DrugBank generates interactions by drug class, so there they are a product interacting with itself, and they stay deleted. An interaction between two entries proves two DrugBank entries, not two drugs.

4. What a split restores

Rows restored, by relation. A group marked same stays merged. The sodium phosphate group splits three ways, keeping each salt with its “unspecified form” entry.

Relation	Lost now	Split the 110 distinct	+ the 6 unclear	Split all 183
drug_drug	108,174	102,670	105,138	108,174
drug_effect	888	888	888	888
pathway_protein	108	108	108	108
disease_disease	2	2	2	2
contraindication	1,566	1,566	1,566	1,566
target	364	300	364	364
indication	428	428	428	428
enzyme	438	432	438	438
transporter	346	258	336	346
off-label use	46	46	46	46
pathway_pathway	24	24	24	24
carrier	102	96	102	102
all	112,486	106,818	109,440	112,486

Per group (§7 and the CSV), the figure is marginal: what splitting that group alone restores, with every other merge in place. Some rows join two merged groups, for example an interaction between a hydrocortisone ester and a testosterone ester. Those depend on both merges, so the per-group figures add up to 111,096, not 112,486. The CSV’s group_rows_lost column is the other bound: every lost row touching the group, so a row between two merged groups counts in both.

5. Surprises

1. Node Normalizer conflates unrelated drugs. Guanabenz, an alpha-2 agonist, shares an entity with icodextrin, a glucose polymer for peritoneal dialysis. Ethanol shares one with ipecac, a plant emetic. Its cliques also join different molecules: human insulin with four analogues and beef insulin, three Tc-99m agents that share only the radionuclide, and Lu-177 dotatate (therapy) with Ga-68 dotatate (imaging).
2. SARS-CoV-1 and SARS-CoV-2 pathways are merged. Reactome builds both virus hierarchies with the same pathway names. Pass 2 merged ten such pairs by name, along with 8 disease-pathway pairs that share a name.
3. One entity’s CURIE names a third antigen. Entity 88504 carries DRUGBANK:DB14432, an A/Singapore/gp1908/2015 IVR-180 antigen. Its members are two A/California/7/2009 X-181 antigens, and PrimeKG has no node for DB14432. The merge script gave each surviving entity Node Normalizer’s preferred CURIE.
4. Name and CURIE disagree on 47 of the 101 drug entities. The name comes from the member with the most edges; the CURIE comes from Node Normalizer. Sometimes they name different substances:
   • “Choline salicylate” carries CHEBI:15354, choline;
   • “Procaine benzylpenicillin” carries CHEBI:17334, penicillin;
   • “Technetium Tc-99m exametazime” carries CHEBI:33353, the technetium atom;
   • “Levonorgestrel” carries CHEBI:231594, norgestrel.
5. Rows lost understate the harm. 16 drug groups lost no row to their own merge. They include both tartaric acids, guanabenz with icodextrin, and ketoconazole with levoketoconazole. Each still serves two drugs’ edges as one node, and a split is what fixes that.
6. A re-run would merge them again. For 228 of the 237 drug members, mods_kg_curie_cache maps the standard DRUGBANK: spelling of the id to the merged entity’s CURIE. Pass 1 groups on exactly that. A split therefore needs a guard, or the next run of the script, or a fresh backfill with Node Normalizer on, will merge them again. Pass 2 would re-merge the pathway pairs by name.
7. A PrimeKG mapping question sits behind two correct gene merges. The node PrimeKG names “COX2” (NCBI Gene 4513, MT-CO2) has disease edges to colorectal cancer, colorectal carcinoma and colorectal neoplasm, beside MELAS and cytochrome-c oxidase deficiency. The colorectal ones look like PTGS2 (COX-2) associations filed under MT-CO2 by symbol, the COX2 homonym of #752. “COX1” (MT-CO1) shows the same pattern. The merges are right, since each pair shares one id, and a split would not touch these edges. They are worth their own check, because colorectal cancer is the index disease.

6. Before any split: open points for the follow-up

These are notes for whoever implements the approved splits. None is decided here.

• Other sources’ rows on these entities were not measured. This review read PrimeKG rows only. Node Normalizer answers for an id, not for a source, so any source that met one of these DrugBank ids sits on the merged entity too. The split must route those rows as well.
• The patient overlay. Maureen’s drug objects that resolved to one of these entities need re-checking, and materialize_patient_kg() must run after the split. The overlay was out of scope here: no maple_patient row was read.
  • One check needed no patient data. Paclitaxel, one of the RGCC taxanes the mission spec names, is merged with paclitaxel poliglumex. Every poliglumex row duplicated a paclitaxel row or joined the two, so the paclitaxel node carries no foreign edge today.
• The re-merge guard (surprise 6): pass 1 and pass 2 of apply_canonical_entity_merges.py, and the cache rows behind them.
• Which member keeps the entity row. A natural choice is the member the entity’s CURIE names, or the parent compound; each other member gets its own entity under its standard CURIE. First check whether an entity with that CURIE already exists. Entity 88504’s CURIE has to change whichever member keeps it.
• Expected counts. For any approved set of partitions, primekg_merge_review.py replays the expected rows restored by relation (§8). The proposal as written gives 106,818.
• Backup, dry run, expected count, per the plan in #438’s 2026-10-05 comment.

7. The review table

Columns: Lost is the rows that splitting this group alone restores, self-loops included; the CSV breaks it down by relation. Rows after a member is its kg.csv row count. J is the Jaccard overlap of the member’s (relation, neighbour) set with that of the group’s anchor, the member whose name the entity carries.

7.1 Drugs (101 groups)

Merge key for every row: Node Normalizer’s preferred CURIE for each member’s own DrugBank: CURIE (§2, pass 1). The members’ own ids are listed.

#	Entity	Members: x_index DrugBank name (rows; J)	Node Normalizer	Lost	Verdict	Rule	Reason
4	79509 CHEBI:5134 Fluticasone propionate	14020 DB00588 Fluticasone propionate (3,722) 14038 DB08906 Fluticasone furoate (2,452; J 0.626) 14040 DB13867 Fluticasone (2,438; J 0.586)	CHEBI:5134 ‘Fluticasone’	4,572	distinct	J2	Fluticasone and two different esters of it (propionate, furoate), each its own approved DrugBank drug
5	79517 CHEBI:17650 Hydrocortisone	14028 DB00741 Hydrocortisone (3,016) 14041 DB14538 Hydrocortisone aceponate (302; J 0.098) 14042 DB14539 Hydrocortisone acetate (2,604; J 0.695) 14043 DB14540 Hydrocortisone butyrate (2,302; J 0.606) 14044 DB14541 Hydrocortisone cypionate (1,652; J 0.543) 14046 DB14543 Hydrocortisone probutate (306; J 0.099) 14047 DB14544 Hydrocortisone valerate (306; J 0.099)	CHEBI:17650 ‘Hydrocortisone’	6,932	distinct	J2	Hydrocortisone and six esters of it (aceponate, acetate, butyrate, cypionate, probutate, valerate)
6	79519 CHEBI:8378 Prednisolone	14030 DB00860 Prednisolone (2,778) 14048 DB14631 Prednisolone phosphate (2,214; J 0.461) 14050 DB15566 Prednisolone acetate (1,510; J 0.544)	CHEBI:8378 ‘Prednisolone’	2,972	distinct	J2	Prednisolone and two esters of it (phosphate, acetate)
7	79521 CHEBI:205919 Clobetasol propionate	14032 DB01013 Clobetasol propionate (2,178) 15047 DB11750 Clobetasol (2,144; J 0.938)	CHEBI:205919 ‘Clobetasol’	2,026	distinct	J2	Clobetasol propionate (the marketed ester) and the parent clobetasol
8	79543 CHEBI:18145 Vitamin E	14054 DB00163 Vitamin E (740) 14110 DB14003 alpha-Tocopherol acetate (24; J 0.008) 14524 DB14001 alpha-Tocopherol succinate (52; J 0.045) 14525 DB14002 D-alpha-Tocopherol acetate (32; J 0.043) 21946 DB14476 DL-alpha-Tocopherol (2; J 0.003)	CHEBI:18145 ‘Vitamin E’	74	distinct	J1, J2, J7	Vitamin E (DrugBank: a term for 8 tocopherols and tocotrienols), two tocopherol esters, a stereo-defined ester and the racemate
9	79639 CHEBI:17303 Morphine	14150 DB00295 Morphine (3,262) 14437 DB11130 Opium (3,230; J 0.517)	CHEBI:17303 ‘Morphine’	2,176	distinct	J7	Opium is a mixture of alkaloids; morphine is one of them
10	79697 CHEBI:17347 Testosterone	14208 DB00624 Testosterone (3,430) 14342 DB01420 Testosterone propionate (2,114; J 0.582) 14476 DB13943 Testosterone cypionate (3,160; J 0.863) 14477 DB13944 Testosterone enanthate (3,368; J 0.821) 14478 DB13946 Testosterone undecanoate (3,362; J 0.734)	CHEBI:17347 ‘Testosterone’	10,984	distinct	J2	Testosterone and four esters of it (propionate, cypionate, enanthate, undecanoate)
11	79710 CHEBI:7627 Norethisterone	14221 DB00717 Norethisterone (2,132) 15744 DB14678 Norethindrone enanthate (852; J 0.400)	CHEBI:7627 ‘Norethisterone’	808	distinct	J2	Norethisterone and its enanthate ester
12	79724 CHEBI:16469 Estradiol	14235 DB00783 Estradiol (4,458) 14479 DB13952 Estradiol acetate (3,720; J 0.762) 14480 DB13953 Estradiol benzoate (3,010; J 0.609) 14481 DB13954 Estradiol cypionate (4,158; J 0.857) 14483 DB13956 Estradiol valerate (4,154; J 0.856) 14501 DB15690 Fluoroestradiol F-18 (32; J 0.003)	CHEBI:16469 ‘Estradiol’	14,552	distinct	J2	Estradiol, four esters of it, and fluoroestradiol F-18, a different molecule (a fluorinated PET tracer)
13	79732 CHEBI:45599 Sodium lauryl sulfate	14243 DB00815 Sodium lauryl sulfate (2) 19752 DB03967 Dodecyl sulfate (2; J 0.000)	CHEBI:45599 ‘Dodecyl sulfate’	0	distinct	J3	Sodium lauryl sulfate (a salt) and the dodecyl sulfate anion
14	79758 CHEBI:6888 Methylprednisolone	14269 DB00959 Methylprednisolone (2,734) 21385 DB14643 Methylprednisolone aceponate (240; J 0.088)	CHEBI:6888 ‘Methylprednisolone’	230	distinct	J2	Methylprednisolone and its aceponate ester
15	79770 CHEBI:47519 Ketoconazole	14281 DB01026 Ketoconazole (2,802) 17623 DB05667 Levoketoconazole (4; J 0.000)	CHEBI:47519 ‘Ketoconazole’	0	distinct	J1	Levoketoconazole is one enantiomer of racemic ketoconazole, developed for a different use (cortisol synthesis inhibition)
16	79773 CHEBI:5001 Fenofibrate	14284 DB01039 Fenofibrate (2,988) 16268 DB13873 Fenofibric acid (1,804; J 0.590)	CHEBI:5001 ‘Fenofibrate’	1,754	distinct	J2	Fenofibrate is the ester prodrug of fenofibric acid; both are approved products
17	79809 CHEBI:41879 Dexamethasone	14320 DB01234 Dexamethasone (3,586) 14493 DB14649 Dexamethasone acetate (56; J 0.016) 15041 DB11487 Dexamethasone isonicotinate (2,128; J 0.585)	CHEBI:41879 ‘Dexamethasone’	2,100	distinct	J2	Dexamethasone and two esters of it (acetate, isonicotinate)
18	79913 CHEBI:43830 Methylene blue	14424 DB09241 Methylene blue (4,506) 17203 DB08167 Methylthioninium (486; J 0.107)	CHEBI:43830 ‘Methylthioninium’	482	distinct	J3	Methylene blue is methylthioninium chloride; the other entry is the cation
19	79930 CHEBI:9679 Calcium Phosphate	14441 DB11348 Calcium Phosphate (740) 14486 DB14481 Calcium phosphate dihydrate (250; J 0.320)	CHEBI:9679 ‘CID 516943’	240	unclear (lean: split)	J5	DrugBank’s ‘Calcium Phosphate’ does not say which calcium phosphate it is; the dihydrate (MW 172.09) is dicalcium phosphate dihydrate. Same salt: J5 keeps them merged. Tricalcium phosphate: different compounds, split
20	79976 CHEBI:25107 Magnesium	14487 DB14513 Magnesium (2,060) 20399 DB09407 Magnesium chloride (1,890; J 0.257)	CHEBI:25107 ‘magnesium atom’	800	distinct	J7	Magnesium (DrugBank: the element) and one specific compound of it
21	79980 CHEBI:35176 Zinc sulfate, unspecified form	14491 DB14548 Zinc sulfate, unspecified form (374) 14751 DB09322 Zinc sulfate (202; J 0.263)	CHEBI:35176 ‘Zinc sulfate’	120	same	J5	Zinc sulfate and DrugBank’s ‘unspecified form’ entry for it (hydration state not stated): one substance
22	79999 CHEBI:28304 Heparin	14510 DB01109 Heparin (1,942) 20334 DB15404 Heparin, bovine (76; J 0.039)	CHEBI:28304 ‘Heparin’	76	distinct	J6	Unfractionated heparin and bovine-derived heparin: different source tissue, a separate DrugBank entry
23	80075 CHEBI:16134 Ammonia	14585 DB11118 Ammonia (36) 14927 DB09326 Ammonia N-13 (4; J 0.000)	CHEBI:16134 ‘Ammonia’	0	distinct	J4	Ammonia and ammonia N-13, a PET perfusion tracer
24	80077 CHEBI:16236 Ethanol	14587 DB00898 Ethanol (3,226) 15256 DB13293 Ipecac (2,144; J 0.322)	CHEBI:16236 ‘Ethanol’	1,296	distinct	evidence	Unrelated substances: ethanol and ipecac (a plant emetic). Node Normalizer conflated them
25	80078 CHEBI:6062 Isosorbide mononitrate	14588 DB01020 Isosorbide mononitrate (2,382) 16719 DB09401 Isosorbide (2,066; J 0.399)	CHEBI:6062 ‘Isosorbide mononitrate’	1,246	distinct	evidence	Different molecules: isosorbide mononitrate (MW 191.14) and isosorbide (MW 146.14)
26	80079 CHEBI:30742 Polyethylene glycol 400	14589 DB11077 Polyethylene glycol 400 (1,632) 20254 DB09287 Polyethylene glycol (984; J 0.175) 20313 DB11161 Polyethylene glycol 300 (86; J 0.053)	CHEBI:30742 ‘macrogol’	480	distinct	J7	Polyethylene glycol (the polymer family) and two specific grades of it (PEG 300, PEG 400)
27	80108 CHEBI:3347 Candesartan	14605 DB00796 Candesartan cilexetil (2,116; J 0.894) 14618 DB13919 Candesartan (2,210)	CHEBI:3347 ‘Candesartan’	2,022	distinct	J2	Candesartan cilexetil is the ester prodrug; candesartan is its active form
28	80109 UMLS:C0700346 Antihemophilic factor, human recombinant	14619 DB00025 Antihemophilic factor, human recombinant (220) 14692 DB13192 Antihemophilic factor human (196; J 0.891)	UMLS:C0700346 ‘Koate-DVI’	192	distinct	J6	Recombinant factor VIII and plasma-derived factor VIII: different products
29	80127 CHEBI:25879 Pentaerythritol tetranitrate	14637 DB06154 Pentaerythritol tetranitrate (1,628) 20350 DB13526 Pentaerithrityl (888; J 0.164)	CHEBI:25879 ‘Pentaerythritol tetranitrate’	358	distinct	evidence	Different molecules: pentaerythritol tetranitrate (MW 316.14), a vasodilator, and pentaerythritol (MW 136.15)
30	80134 CHEBI:30830 Sodium oxybate	14644 DB09072 Sodium oxybate (1,688) 19279 DB01440 gamma-Hydroxybutyric acid (1,478; J 0.865)	CHEBI:30830 ‘sodium oxybate’	1,452	distinct	J3	Sodium oxybate is the sodium salt of gamma-hydroxybutyric acid
31	80144 CHEBI:15354 Choline salicylate	14653 DB00122 Choline (254; J 0.112) 14654 DB14006 Choline salicylate (2,178) 18900 DB04660 Choline alfoscerate (2; J 0.000) 20813 DB09277 Choline C 11 (1,494; J 0.666)	CHEBI:15354 ‘Choline’	1,684	distinct	J4	Four different drugs: choline; choline salicylate (salicylate is an active NSAID); choline alfoscerate (glycerophosphocholine); choline C-11 (a PET tracer)
32	80195 CHEBI:27300 Vitamin D	14560 DB00153 Ergocalciferol (520; J 0.194) 14561 DB00169 Cholecalciferol (942; J 0.238) 14705 DB11094 Vitamin D (1,180)	CHEBI:27300 ‘Ergocalciferol’	790	distinct	J7	Vitamin D (the class) and two different molecules in it, ergocalciferol (D2) and cholecalciferol (D3)
33	80208 CHEBI:27568 Selenium	14716 DB11135 Selenium (1,464) 14742 DB11127 Selenious acid (1,462; J 0.985)	CHEBI:27568 ‘Selenium’	1,430	distinct	J7	Selenium (the trace element) and selenious acid, one compound of it
34	80259 CHEBI:17234 Dextrose, unspecified form	14353 DB01914 D-glucose (34; J 0.233) 14766 DB09341 Dextrose, unspecified form (114) 15755 DB02379 Beta-D-Glucose (46; J 0.000)	CHEBI:17234 ‘D-glucose’	28	unclear → {14353, 14766} · {15755}	J1, J5	D-glucose and dextrose (same DrugBank text) are one substance under J5; beta-D-glucose is one anomer, which interconverts with the other in water, so J1 may not apply
35	80261 UNII:PUE0TO3XDR Dalteparin	14768 DB06779 Dalteparin (1,938) 15890 DB06822 Tinzaparin (1,862; J 0.953) 16669 DB09258 Bemiparin (1,856; J 0.952)	UNII:PUE0TO3XDR ‘Dalteparin’	3,634	distinct	evidence	Three different low-molecular-weight heparins, each separately made and approved
36	80285 CHEBI:45863 Paclitaxel	14792 DB01229 Paclitaxel (3,484) 21238 DB13809 Paclitaxel poliglumex (118; J 0.033)	CHEBI:45863 ‘Paclitaxel’	118	distinct	J2	Paclitaxel and paclitaxel poliglumex, a polymer-drug conjugate prodrug
37	80292 CHEBI:5931 Insulin human	14435 DB09564 Insulin degludec (1,466; J 0.891) 14799 DB00030 Insulin human (1,636) 15283 DB00047 Insulin glargine (1,548; J 0.925) 15331 DB01306 Insulin aspart (1,292; J 0.755) 15332 DB01309 Insulin glulisine (1,468; J 0.895) 15378 DB09456 Insulin beef (1,612; J 0.983)	CHEBI:5931 ‘insulin human’	7,064	distinct	evidence	Six different insulins: human insulin, four analogues with different sequences and kinetics, and beef insulin
38	80297 CHEBI:15671 L-tartaric acid	14804 DB09459 L-tartaric acid (80) 18840 DB01694 D-tartaric acid (10; J 0.000)	CHEBI:15671 ‘L-tartaric acid’	0	distinct	J1	L- and D-tartaric acid are enantiomers (known example)
39	80309 CHEBI:4735 Edetic acid	14756 DB14600 Edetate disodium anhydrous (12; J 0.006) 14816 DB00974 Edetic acid (1,340)	CHEBI:4735 ‘Edetic acid’	8	distinct	J3	Edetate disodium (a salt, used for hypercalcaemia) and edetic acid
40	80328 CHEBI:3011 Benazepril	14835 DB00542 Benazepril (1,968) 21149 DB14125 Benazeprilat (1,612; J 0.814)	CHEBI:3011 ‘Benazepril’	1,600	distinct	J2	Benazepril is the ester prodrug; benazeprilat is its active metabolite
41	80388 CHEBI:82717 Tedizolid phosphate	14895 DB09042 Tedizolid phosphate (780) 14896 DB14569 Tedizolid (42; J 0.054)	CHEBI:82717 ‘Tedizolid’	42	distinct	J2	Tedizolid phosphate is the prodrug of tedizolid
42	80411 CHEBI:16335 Adenosine	14632 DB00131 Adenosine phosphate (42; J 0.000) 14918 DB00640 Adenosine (1,520)	CHEBI:16335 ‘Adenosine’	0	distinct	evidence	Different molecules: adenosine monophosphate (a nucleotide) and adenosine (the nucleoside)
43	80436 CHEBI:231594 Levonorgestrel	14826 DB09389 Norgestrel (924; J 0.515) 14943 DB00367 Levonorgestrel (1,788)	CHEBI:231594 ‘Norgestrel’	878	distinct	J1	Norgestrel is the racemate; levonorgestrel is its active enantiomer
44	80442 CHEBI:17698 Chloramphenicol	14392 DB07565 Chloramphenicol succinate (38; J 0.019) 14949 DB00446 Chloramphenicol (1,764)	CHEBI:17698 ‘Chloramphenicol’	34	distinct	J2	Chloramphenicol succinate is the ester prodrug of chloramphenicol
45	80496 CHEBI:16962 Cortisone acetate	15003 DB01380 Cortisone acetate (2,864) 15071 DB14681 Cortisone (1,926; J 0.637)	CHEBI:16962 ‘Cortisone’	1,800	distinct	J2	Cortisone acetate and the parent cortisone
46	80519 CHEBI:17252 Hydroxyprogesterone caproate	15026 DB06789 Hydroxyprogesterone caproate (1,426) 15740 DB14570 Hydroxyprogesterone (266; J 0.154)	CHEBI:17252 ‘Hydroxyprogesterone’	222	distinct	J2	Hydroxyprogesterone caproate (the drug) and 17-hydroxyprogesterone (the endogenous steroid)
47	80557 CHEBI:6970 Mometasone furoate	14970 DB00764 Mometasone (2,250; J 0.873) 15064 DB14512 Mometasone furoate (2,350)	CHEBI:6970 ‘Mometasone’	2,078	distinct	J2	Mometasone and its furoate ester (the marketed form)
48	80620 CHEBI:8028 Perphenazine	15127 DB00850 Perphenazine (2,770) 21386 DB14651 Perphenazine enanthate (1,782; J 0.641)	CHEBI:8028 ‘Perphenazine’	1,762	distinct	J2	Perphenazine and its enanthate depot ester
49	80720 CHEBI:85155 Netupitant	15227 DB09048 Netupitant (1,860) 15731 DB14019 Fosnetupitant (414; J 0.221)	CHEBI:85155 ‘Netupitant’	414	distinct	J2	Fosnetupitant is the intravenous prodrug of netupitant
50	80753 CHEBI:4464 Dexchlorpheniramine	14575 DB09555 Dexchlorpheniramine maleate (1,484; J 0.777) 15260 DB13679 Dexchlorpheniramine (1,682)	CHEBI:4464 ‘Dexchlorpheniramine’	1,382	distinct	J3	Dexchlorpheniramine maleate (the salt) and dexchlorpheniramine
51	80797 CHEBI:92998 Guanabenz	15304 DB00629 Guanabenz (2,040) 15754 DB00702 Icodextrin (2; J 0.000)	CHEBI:92998 ‘Guanabenz’	0	distinct	evidence	Unrelated drugs: guanabenz (an alpha-2 agonist) and icodextrin (a glucose polymer for peritoneal dialysis). Node Normalizer conflated them
52	80960 CHEBI:50785 Ethynodiol diacetate	15467 DB00823 Ethynodiol diacetate (1,172) 15724 DB13866 Etynodiol (830; J 0.708)	CHEBI:50785 ‘Etynodiol’	786	distinct	J2	Ethynodiol diacetate (the marketed ester) and etynodiol (never marketed, per DrugBank)
53	80961 CHEBI:25187 Levomenthol	15468 DB00825 Levomenthol (1,324) 20335 DB14123 Racementhol (122; J 0.000)	CHEBI:25187 ‘Menthol’	0	distinct	J1	Levomenthol is one enantiomer; racementhol is the racemate
54	80980 CHEBI:16973 Desoxycorticosterone pivalate	15487 DB01134 Desoxycorticosterone pivalate (18) 15581 DB06780 Desoxycorticosterone acetate (18; J 1.000)	CHEBI:16973 ‘desoxycortone’	18	distinct	J2	Two different esters of desoxycorticosterone (pivalate, acetate)
55	81072 CHEBI:135590 Ambroxol	15579 DB06742 Ambroxol (18) 15897 DB13141 Ambroxol acefyllinate (8; J 0.083)	CHEBI:135590 ‘Ambroxol’	2	distinct	evidence	Ambroxol acefyllinate adds a second active ingredient (acefylline, a xanthine)
56	81144 CHEBI:34831 Melengestrol	15070 DB14659 Melengestrol acetate (1,704; J 0.574) 15651 DB11529 Melengestrol (1,860)	CHEBI:34831 ‘melengestrol’	1,240	distinct	J2	Melengestrol and its acetate ester
57	81201 CHEBI:59919 Dihydroergocryptine	15641 DB11274 Dihydro-alpha-ergocryptine (24; J 0.008) 15708 DB13385 Dihydroergocryptine (2,104)	CHEBI:59919 ‘Dihydro-alpha-ergocryptine’	16	unclear (lean: split)	J7	DrugBank’s own text says alpha-dihydroergocryptine ‘is usually referred to the mixture’ of alpha and beta; mixture vs component (J7) or one drug under two names
58	81222 CHEBI:135564 Nomegestrol acetate	15658 DB11636 Nomegestrol (830; J 0.906) 15729 DB13981 Nomegestrol acetate (916)	CHEBI:135564 ‘Nomegestrol acetate’	786	distinct	J2	Nomegestrol and its acetate ester
59	81234 CHEBI:174358 Eslicarbazepine	15620 DB09119 Eslicarbazepine acetate (2,100; J 0.971) 15741 DB14575 Eslicarbazepine (2,146)	CHEBI:174358 ‘Eslicarbazepine’	2,058	distinct	J2	Eslicarbazepine acetate is the prodrug of eslicarbazepine
60	81264 CHEBI:102484 Sulfisoxazole	15771 DB00263 Sulfisoxazole (2,030) 15803 DB14033 Acetyl sulfisoxazole (624; J 0.306)	CHEBI:102484 ‘Sulfisoxazole’	598	distinct	J2	Acetyl sulfisoxazole is an ester of sulfisoxazole
61	81310 CHEBI:2469 Adefovir dipivoxil	15817 DB00718 Adefovir dipivoxil (1,618) 20851 DB13868 Adefovir (1,410; J 0.869)	CHEBI:2469 ‘Adefovir’	1,388	distinct	J2	Adefovir dipivoxil is the diester prodrug of adefovir
62	81534 CHEBI:28939 Acetylcysteine	16036 DB06151 Acetylcysteine (222) 20971 DB14479 Acetylcysteine zinc (120; J 0.000)	CHEBI:28939 ‘Acetylcysteine’	0	distinct	J3	Acetylcysteine and its zinc salt
63	81559 CHEBI:92712 Atropine	16057 DB00424 Hyoscyamine (2,910; J 0.773) 16061 DB00572 Atropine (3,216)	CHEBI:92712 ‘ENDO-ATROPINE’	2,646	distinct	J1	Hyoscyamine is the L-enantiomer; atropine is the racemate
64	81677 CHEMBL.COMPOUND:CHEMBL2108455 Turoctocog alfa	16179 DB09109 Turoctocog alfa (200) 18412 DB11606 Susoctocog alfa (194; J 0.931)	CHEMBL.COMPOUND:CHEMBL2108455 ‘octocog alfa’	186	distinct	evidence	Different factor VIII molecules: human sequence (turoctocog) and porcine sequence (susoctocog)
65	81715 CHEBI:17115 Serine	16217 DB00133 Serine (98) 17034 DB03929 D-Serine (10; J 0.019)	CHEBI:17115 ‘Serine’	2	distinct	J1	L-serine and D-serine are enantiomers
66	81792 CHEBI:132767 Ferric pyrophosphate	16294 DB09147 Ferric pyrophosphate (318) 16296 DB13995 Ferric pyrophosphate citrate (148; J 0.447)	CHEBI:132767 ‘Ferric pyrophosphate’	144	distinct	evidence	Different iron complexes: ferric pyrophosphate (insoluble) and ferric pyrophosphate citrate (soluble, per DrugBank)
67	81840 CHEBI:18248 Iron	14114 DB00893 Iron Dextran (284; J 0.810) 16341 DB01592 Iron (324)	CHEBI:18248 ‘Iron’	266	distinct	J7	Iron (the element) and iron dextran (one specific complex)
68	81935 CHEBI:34584 Boldenone	16436 DB01541 Boldenone (294) 16463 DB14639 Boldenone undecylenate (2; J 0.007)	CHEBI:34584 ‘Boldenone’	2	distinct	J2	Boldenone and its undecylenate ester
69	82393 CHEBI:232390 Secretin human	16892 DB09532 Secretin human (390) 21400 DB09527 Secretin porcine (326; J 0.836)	CHEBI:232390 ‘CID 102090385’	318	distinct	evidence	Different peptides: human and porcine secretin (MW 3039.46 vs 3055.46)
70	82524 CHEBI:78538 Tafamidis	14081 DB05352 Fx-1006A (2; J 0.000) 17023 DB11644 Tafamidis (350)	CHEBI:78538 ‘Fx-1006A’	0	distinct	J3	Fx-1006A is tafamidis meglumine (DrugBank’s tafamidis text says so), the salt; tafamidis is the free acid
71	82605 CHEBI:81569 Follitropin	17104 DB00066 Follitropin (376) 17105 DB00094 Urofollitropin (180; J 0.004)	CHEBI:81569 ‘Follitropin’	2	distinct	J6	Recombinant follitropin and urine-derived urofollitropin
72	82785 UNII:MRL3739G66 Lutetium Lu 177 dotatate	17282 DB13985 Lutetium Lu 177 dotatate (24) 19675 DB13925 Dotatate gallium Ga-68 (12; J 0.500)	UNII:MRL3739G66 ‘Dotatate’	12	distinct	J4	One peptide with two radionuclides: Lu-177 dotatate (therapy) and Ga-68 dotatate (imaging)
73	82807 CHEBI:15553 Dinoprost tromethamine	17302 DB01160 Dinoprost tromethamine (228) 19699 DB12789 Dinoprost (98; J 0.006)	CHEBI:15553 ‘Dinoprost’	2	distinct	J3	Dinoprost tromethamine (the salt) and dinoprost
74	82837 CHEBI:81567 Thyrotropin alfa	17332 DB00024 Thyrotropin alfa (970) 20956 DB15263 Thyrotropin (374; J 0.386)	CHEBI:81567 ‘Thyrotropin’	372	distinct	J6	Recombinant thyrotropin alfa and endogenous thyrotropin
75	82852 CHEBI:7507 Neomycin	17347 DB00994 Neomycin (1,930) 18101 DB00452 Framycetin (1,668; J 0.858)	CHEBI:7507 ‘Framycetin’	1,626	distinct	J7	Neomycin is a complex of components; framycetin is one of them (neomycin B)
76	82901 CHEBI:6900 Metocurine	17396 DB01336 Metocurine (2,378) 17645 DB00416 Metocurine iodide (2,178; J 0.914)	CHEBI:6900 ‘Metocurine’	2,122	distinct	J3	Metocurine and its iodide salt
77	83066 CHEBI:42797 Gabapentin enacarbil	14709 DB00996 Gabapentin (2,220; J 0.456) 17559 DB08872 Gabapentin enacarbil (2,740)	CHEBI:42797 ‘Gabapentin’	1,536	distinct	J2	Gabapentin enacarbil is a prodrug of gabapentin, approved for different indications
78	84213 CHEBI:75275 Tetradecyl hydrogen sulfate (ester)	17060 DB00464 Sodium tetradecyl sulfate (32; J 0.000) 18702 DB11328 Tetradecyl hydrogen sulfate (ester) (1,474)	CHEBI:75275 ‘Tetradecyl hydrogen sulfate (ester)‘	0	distinct	J3	Sodium tetradecyl sulfate (the salt) and tetradecyl hydrogen sulfate (the acid)
79	85123 UMLS:C2346821 CT-011	19602 DB05916 CT-011 (668) 21194 DB15383 Pidilizumab (666; J 0.991)	UMLS:C2346821 ‘pidilizumab’	658	same	J8	CT-011 is the development code for pidilizumab, one anti-PD-1 antibody
80	85689 CHEBI:62964 Magnesium acetate tetrahydrate	17633 DB13996 Magnesium acetate (554; J 0.965) 20145 DB09409 Magnesium acetate tetrahydrate (566)	CHEBI:62964 ‘CID 517043’	544	same	J5	Magnesium acetate and its tetrahydrate: one salt
81	85779 CHEBI:30769 Sodium citrate	16350 DB04272 Citric acid (246; J 0.071) 20218 DB09154 Sodium citrate (1,746)	CHEBI:30769 ‘Citric acid’	132	distinct	J3	Citric acid and sodium citrate, its sodium salt
82	85782 CHEBI:92769 Iobenguane sulfate I-123	20220 DB06704 Iobenguane (276; J 0.038) 20221 DB09546 Iobenguane sulfate I-123 (1,474)	CHEBI:92769 ‘Iobenguane’	64	distinct	J4	Iobenguane and the I-123 iobenguane sulfate radiopharmaceutical
83	85856 CHEBI:34683 Sodium phosphate, monobasic	20290 DB09449 Sodium phosphate, monobasic (1,738) 20291 DB14502 Sodium phosphate, dibasic (870; J 0.495) 20292 DB14503 Sodium phosphate, monobasic, unspecified form (832; J 0.470) 20542 DB09370 Sodium phosphate P 32 (6; J 0.000) 21026 DB14504 Sodium phosphate, dibasic, unspecified form (820; J 0.465)	CHEBI:34683 ‘CID 58592228’	2,482	distinct → {20290, 20292} · {20291, 21026} · {20542}	J3, J4, J5	Two different salts (monobasic, dibasic), each with its ‘unspecified form’ entry, and the P-32 radiotherapeutic
84	85860 CHEBI:17268 Inositol	15885 DB08949 Inositol nicotinate (70; J 0.039) 20293 DB13178 Inositol (1,520)	CHEBI:17268 ‘Inositol’	60	distinct	evidence	Inositol nicotinate is a niacin ester that delivers nicotinic acid; inositol is a different drug
85	85891 CHEBI:135361 Methscopolamine	16058 DB00462 Methscopolamine bromide (1,922; J 0.899) 20321 DB11315 Methscopolamine (2,120)	CHEBI:135361 ‘Methscopolamine’	1,892	distinct	J3	Methscopolamine and its bromide salt
86	85926 CHEBI:36773 Synthetic camphor	18372 DB01744 Camphor (8; J 0.000) 20338 DB14156 Synthetic camphor (68)	CHEBI:36773 ‘camphor’	0	unclear (lean: split)	J1	Synthetic camphor is the racemate and natural camphor the (+)-enantiomer, but neither DrugBank entry states its stereochemistry
87	85941 CHEBI:53248 Povidone	20352 DB11061 Povidone (22) 20353 DB14478 Povidone K30 (22; J 1.000)	CHEBI:53248 ‘Povidone’	22	distinct	J7	Povidone (the polymer) and povidone K30, one grade of it
88	86030 UNII:6M3P64V0NC Almasilate	20403 DB13595 Almasilate (1,964) 21273 DB11230 Magnesium Aluminum Silicate (182; J 0.050)	UNII:6M3P64V0NC ‘Almasilate’	100	distinct	evidence	Different materials: almasilate, a synthetic crystalline antacid, and magnesium aluminium silicate, a natural clay excipient
89	86036 CHEBI:17334 Procaine benzylpenicillin	14288 DB01053 Benzylpenicillin (1,232; J 0.271) 20409 DB09320 Procaine benzylpenicillin (1,806)	CHEBI:17334 ‘penicillin’	620	distinct	evidence	Procaine benzylpenicillin adds procaine, a second active ingredient (a depot formulation)
90	86069 CHEBI:7466 Nandrolone	14825 DB08804 Nandrolone decanoate (402; J 0.762) 16433 DB00984 Nandrolone phenpropionate (374; J 0.809) 20432 DB13169 Nandrolone (458)	CHEBI:7466 ‘Nandrolone’	678	distinct	J2	Nandrolone and two esters of it (decanoate, phenpropionate)
91	86079 CHEBI:78320 Ammonium lactate	20217 DB04398 Lactic acid (28; J 0.000) 20437 DB06768 Ammonium lactate (80)	CHEBI:78320 ‘Lactic acid’	0	distinct	J3	Lactic acid and ammonium lactate, its salt
92	86199 CHEBI:2914 Atracurium	17648 DB00732 Atracurium besylate (2,178; J 0.971) 20460 DB13295 Atracurium (2,236)	CHEBI:2914 ‘Atracurium’	2,120	distinct	J3	Atracurium and its besylate salt
93	86316 UMLS:C2350656 Pegloticase	20488 DB09208 Pegloticase (114) 21528 DB05321 PEG-uricase (72; J 0.603)	UMLS:C2350656 ‘pegloticase’	70	same	J8	PEG-uricase (a PEG conjugate of recombinant porcine uricase, in phase III in 2006) is pegloticase before approval
94	87902 CHEBI:34838 Drostanolone propionate	14526 DB00858 Drostanolone (6; J 0.000) 20630 DB14655 Drostanolone propionate (192)	CHEBI:34838 ‘Drostanolone’	0	distinct	J2	Drostanolone and its propionate ester
95	88287 CHEBI:140407 Ceftobiprole	20665 DB04918 Ceftobiprole (1,548) 21610 DB14733 Ceftobiprole medocaril (34; J 0.022)	CHEBI:140407 ‘Ceftobiprole’	34	distinct	J2	Ceftobiprole medocaril is the prodrug of ceftobiprole
96	88429 CHEBI:33353 Technetium Tc-99m exametazime	16329 DB00113 Technetium Tc-99m arcitumomab (660; J 0.011) 20181 DB09266 Technetium Tc-99m tilmanocept (58; J 0.004) 20807 DB09163 Technetium Tc-99m exametazime (1,472)	CHEBI:33353 ‘technetium atom’	24	distinct	evidence	Three different radiopharmaceuticals that share only the Tc-99m label: an anti-CEA antibody fragment, tilmanocept, exametazime
97	88504 DRUGBANK:DB14432 Influenza A virus A/California/7/2009 X-181 (H1N1) antigen (propiolactone inactivated)	20882 DB10600 Influenza A virus A/California/7/2009 X-181 (H1N1) antigen (propiolactone inactivated) (584) 20893 DB11041 Influenza A virus A/California/7/2009 X-181 (H1N1) hemagglutinin antigen (propiolactone inactivated) (584; J 1.000)	DRUGBANK:DB14432 ‘Influenza A virus A/Singapore/gp1908/2015 IVR-180 (H1N1) hemagglutinin antigen (formaldehyde inactivated)‘	560	unclear (lean: split)	none	Two DrugBank vaccine components for one strain (A/California/7/2009 X-181), ‘antigen’ and ‘hemagglutinin antigen’, both propiolactone-inactivated; identical PrimeKG edges
98	88506 UMLS:C4033637 Influenza B virus B/Brisbane/60/2008 antigen (formaldehyde inactivated)	20884 DB10794 Influenza B virus B/Brisbane/60/2008 antigen (formaldehyde inactivated) (584) 20892 DB11040 Influenza B virus B/Brisbane/60/2008 hemagglutinin antigen (propiolactone inactivated) (584; J 1.000) 20894 DB11044 Influenza B virus B/Brisbane/60/2008 antigen (propiolactone inactivated) (584; J 1.000) 20944 DB14385 Influenza B virus B/Brisbane/60/2008 hemagglutinin antigen (formaldehyde inactivated) (584; J 1.000)	UMLS:C4033637 ‘influenza B virus B/Brisbane/60/2008 antigen’	1,680	unclear (lean: split)	J6	Four DrugBank vaccine components for one strain (B/Brisbane/60/2008): antigen or hemagglutinin antigen, formaldehyde- or propiolactone-inactivated; identical PrimeKG edges
99	88589 CHEBI:133354 Bentonite	20967 DB11306 Bentonite (54) 20969 DB13654 Montmorillonite (54; J 1.000)	CHEBI:133354 ‘Bentonite’	54	distinct	J7	Bentonite is a clay; montmorillonite is its main mineral
100	88647 CHEBI:22652 Sodium ascorbate	14128 DB00126 Ascorbic acid (320; J 0.108) 21025 DB14482 Sodium ascorbate (934)	CHEBI:22652 ‘Ascorbic acid’	114	distinct	J3	Ascorbic acid and sodium ascorbate, its salt
101	89004 CHEBI:28842 Magnesium stearate	14355 DB03193 Stearic acid (8; J 0.000) 21382 DB14077 Magnesium stearate (182)	CHEBI:28842 ‘Stearic acid’	0	distinct	J3	Stearic acid and magnesium stearate, its salt
102	89006 CHEBI:50848 Loteprednol	14754 DB14596 Loteprednol etabonate (46; J 0.000) 21384 DB00873 Loteprednol (374)	CHEBI:50848 ‘Loteprednol’	0	distinct	J2	Loteprednol etabonate (the marketed ester) and loteprednol
103	89236 CHEBI:136037 Corticorelin ovine triflutate	14642 DB05394 Corticorelin (20; J 0.000) 21614 DB09067 Corticorelin ovine triflutate (136)	CHEBI:136037 ‘Corticorelin’	0	distinct	evidence	Different peptides: corticorelin (human CRF) and corticorelin ovine triflutate (the ovine sequence)
104	89486 CHEBI:30021 Selenomethionine Se-75	21864 DB09400 Selenomethionine Se-75 (2) 21866 DB11142 Selenomethionine (2; J 1.000)	CHEBI:30021 ‘Selenomethionine’	2	distinct	J4	Selenomethionine and its Se-75 radiolabelled form

7.2 Pathways (18 groups)

Merge key for every row: identical names, joined by pass 2. Reactome records were read on 2026-10-05.

#	Entity	Members: x_index Reactome id name (rows; J)	Lost	Verdict	Rule	Reason
166	169000 REACTOME:R-HSA-3656244 Defective B4GALT1 causes B4GALT1-CDG (CDG-2d)	128266 R-HSA-3656244 Defective B4GALT1 causes B4GALT1-CDG (CDG-2d) (18) 128330 R-HSA-4793953 Defective B4GALT1 causes B4GALT1-CDG (CDG-2d) (4; J 0.100)	2	distinct	evidence	Different Reactome pathways that shared a name: R-HSA-3656244: disease congenital disorder of glycosylation type II; normal pathway Keratan sulfate/keratin metabolism; 3 events; vs R-HSA-4793953: disease congenital disorder of glycosylation type II; normal pathway Asparagine N-linked glycosylation; event R-HSA-4793956; PrimeKG proteins 8 vs 1
167	169493 REACTOME:R-HSA-5683678 Defective ABCA3 causes SMDP3	128636 R-HSA-5683678 Defective ABCA3 causes SMDP3 (4) 128776 R-HSA-5688399 Defective ABCA3 causes SMDP3 (4; J 0.333)	2	distinct	evidence	Different Reactome pathways that shared a name: R-HSA-5683678: disease newborn respiratory distress syndrome; normal pathway ABC-family protein mediated transport; event R-HSA-5683672; vs R-HSA-5688399: disease interstitial lung disease, newborn respiratory distress syndrome; normal pathway Surfactant metabolism; event R-HSA-5688397; different PrimeKG parents
168	169498 REACTOME:R-HSA-5619037 Defective SLC35A1 causes congenital disorder of glycosylation 2F (CDG2F)	128641 R-HSA-5619037 Defective SLC35A1 causes congenital disorder of glycosylation 2F (CDG2F) (4) 128707 R-HSA-5663020 Defective SLC35A1 causes congenital disorder of glycosylation 2F (CDG2F) (4; J 1.000)	4	distinct	J9	Joined only by name; in PrimeKG both have the same parent and protein. R-HSA-5619037: disease congenital disorder of glycosylation type II; normal pathway Transport of vitamins, nucleosides, and related molecules; event R-HSA-5651942; vs R-HSA-5663020: disease congenital disorder of glycosylation type II; normal pathway Sialic acid metabolism; event R-HSA-5651942; Reactome now names them apart
169	169504 REACTOME:R-HSA-5619044 Defective SLC6A19 causes Hartnup disorder (HND)	128647 R-HSA-5619044 Defective SLC6A19 causes Hartnup disorder (HND) (4) 128702 R-HSA-5659735 Defective SLC6A19 causes Hartnup disorder (HND) (4; J 1.000)	4	distinct	J9	Joined only by name; in PrimeKG both have the same parent and protein. R-HSA-5619044: disease Hartnup disease; normal pathway SLC-mediated transport of neurotransmitters; event R-HSA-5659734; vs R-HSA-5659735: disease Hartnup disease; normal pathway SLC-mediated transport of amino acids; event R-HSA-5659734; Reactome now names them apart
170	169531 REACTOME:R-HSA-5619079 Defective SLC6A18 may confer susceptibility to iminoglycinuria and/or hyperglycinuria	128674 R-HSA-5619079 Defective SLC6A18 may confer susceptibility to iminoglycinuria and/or hyperglycinuria (4) 128701 R-HSA-5659729 Defective SLC6A18 may confer susceptibility to iminoglycinuria and/or hyperglycinuria (4; J 1.000)	4	distinct	J9	Joined only by name; in PrimeKG both have the same parent and protein. R-HSA-5619079: not in Reactome now (HTTP Error 404: Not Found); vs R-HSA-5659729: not in Reactome now (HTTP Error 404: Not Found)
171	169532 REACTOME:R-HSA-5619081 Defective SLC6A3 causes Parkinsonism-dystonia infantile (PKDYS)	128675 R-HSA-5619081 Defective SLC6A3 causes Parkinsonism-dystonia infantile (PKDYS) (4) 128704 R-HSA-5660724 Defective SLC6A3 causes Parkinsonism-dystonia infantile (PKDYS) (4; J 1.000)	4	distinct	J9	Joined only by name; in PrimeKG both have the same parent and protein. R-HSA-5619081: disease neurodegenerative disease; normal pathway Neurotransmitter clearance; event R-HSA-5660706; vs R-HSA-5660724: disease neurodegenerative disease; normal pathway SLC-mediated transport of neurotransmitters; event R-HSA-5660706; Reactome now names them apart
172	169545 REACTOME:R-HSA-5619101 Variant SLC6A20 contributes towards hyperglycinuria (HG) and iminoglycinuria (IG)	128688 R-HSA-5619101 Variant SLC6A20 contributes towards hyperglycinuria (HG) and iminoglycinuria (IG) (4) 128703 R-HSA-5660686 Variant SLC6A20 contributes towards hyperglycinuria (HG) and iminoglycinuria (IG) (4; J 1.000)	4	distinct	J9	Joined only by name; in PrimeKG both have the same parent and protein. R-HSA-5619101: disease amino acid metabolic disorder; normal pathway SLC-mediated transport of neurotransmitters; event R-HSA-5660694; vs R-HSA-5660686: disease amino acid metabolic disorder; normal pathway SLC-mediated transport of amino acids; event R-HSA-5660694; Reactome now names them apart
173	169553 REACTOME:R-HSA-5619114 Defective SLC5A7 causes distal hereditary motor neuronopathy 7A (HMN7A)	128696 R-HSA-5619114 Defective SLC5A7 causes distal hereditary motor neuronopathy 7A (HMN7A) (4) 128700 R-HSA-5658471 Defective SLC5A7 causes distal hereditary motor neuronopathy 7A (HMN7A) (4; J 1.000)	4	distinct	J9	Joined only by name; in PrimeKG both have the same parent and protein. R-HSA-5619114: disease motor peripheral neuropathy; normal pathway Neurotransmitter release cycle; event R-HSA-5658483; vs R-HSA-5658471: disease motor peripheral neuropathy; normal pathway SLC-mediated transport of organic anions; event R-HSA-5658483; Reactome now names them apart
174	170344 REACTOME:R-HSA-9679504 Translation of Replicase and Assembly of the Replication Transcription Complex	63050 R-HSA-9679504 Translation of Replicase and Assembly of the Replication Transcription Complex (34) 63058 R-HSA-9694676 Translation of Replicase and Assembly of the Replication Transcription Complex (34; J 0.700)	30	distinct	evidence	The same name in two virus hierarchies, SARS-CoV-1 and SARS-CoV-2 (PrimeKG parents); R-HSA-9679504: disease severe acute respiratory syndrome; 8 events; vs R-HSA-9694676: disease COVID-19; 7 events; 0 event(s) shared
175	170345 REACTOME:R-HSA-9679509 Virion Assembly and Release	129307 R-HSA-9679509 Virion Assembly and Release (2) 129322 R-HSA-9694322 Virion Assembly and Release (2; J 0.000)	0	distinct	evidence	The same name in two virus hierarchies, SARS-CoV-1 and SARS-CoV-2 (PrimeKG parents); R-HSA-9679509: disease severe acute respiratory syndrome; 8 events; vs R-HSA-9694322: disease COVID-19; 8 events; 0 event(s) shared
176	170347 REACTOME:R-HSA-9683701 Translation of structural proteins	63054 R-HSA-9683701 Translation of structural proteins (12) 63057 R-HSA-9694635 Translation of structural proteins (12; J 0.000)	10	distinct	evidence	The same name in two virus hierarchies, SARS-CoV-1 and SARS-CoV-2 (PrimeKG parents); R-HSA-9683701: disease severe acute respiratory syndrome; 10 events; vs R-HSA-9694635: disease COVID-19; 10 events; 0 event(s) shared
177	170349 REACTOME:R-HSA-9684325 Maturation of replicase proteins	129309 R-HSA-9684325 Maturation of replicase proteins (2) 129330 R-HSA-9694301 Maturation of replicase proteins (2; J 0.000)	2	distinct	evidence	The same name in two virus hierarchies, SARS-CoV-1 and SARS-CoV-2 (PrimeKG parents); R-HSA-9684325: disease severe acute respiratory syndrome; 14 events; vs R-HSA-9694301: disease COVID-19; 18 events; 0 event(s) shared
178	170359 REACTOME:R-HSA-9683610 Maturation of nucleoprotein	129316 R-HSA-9683610 Maturation of nucleoprotein (22) 129327 R-HSA-9694631 Maturation of nucleoprotein (22; J 0.833)	22	distinct	evidence	The same name in two virus hierarchies, SARS-CoV-1 and SARS-CoV-2 (PrimeKG parents); R-HSA-9683610: disease severe acute respiratory syndrome; 9 events; vs R-HSA-9694631: disease COVID-19; 14 events; 1 event(s) shared
179	170360 REACTOME:R-HSA-9683612 Maturation of protein M	129317 R-HSA-9683612 Maturation of protein M (2) 129326 R-HSA-9694594 Maturation of protein M (2; J 0.000)	2	distinct	evidence	The same name in two virus hierarchies, SARS-CoV-1 and SARS-CoV-2 (PrimeKG parents); R-HSA-9683612: disease severe acute respiratory syndrome; 3 events; vs R-HSA-9694594: disease COVID-19; 3 events; 0 event(s) shared
180	170361 REACTOME:R-HSA-9683673 Maturation of protein 3a	129318 R-HSA-9683673 Maturation of protein 3a (20) 129328 R-HSA-9694719 Maturation of protein 3a (20; J 0.818)	20	distinct	evidence	The same name in two virus hierarchies, SARS-CoV-1 and SARS-CoV-2 (PrimeKG parents); R-HSA-9683673: disease severe acute respiratory syndrome; 7 events; vs R-HSA-9694719: disease COVID-19; 7 events; 0 event(s) shared
181	170362 REACTOME:R-HSA-9683683 Maturation of protein E	129319 R-HSA-9683683 Maturation of protein E (10) 129324 R-HSA-9694493 Maturation of protein E (10; J 0.667)	10	distinct	evidence	The same name in two virus hierarchies, SARS-CoV-1 and SARS-CoV-2 (PrimeKG parents); R-HSA-9683683: disease severe acute respiratory syndrome; 5 events; vs R-HSA-9694493: disease COVID-19; 5 events; 0 event(s) shared
182	170366 REACTOME:R-HSA-9694614 Attachment and Entry	129306 R-HSA-9678110 Attachment and Entry (10) 129323 R-HSA-9694614 Attachment and Entry (12; J 0.571)	8	distinct	evidence	The same name in two virus hierarchies, SARS-CoV-1 and SARS-CoV-2 (PrimeKG parents); R-HSA-9678110: disease severe acute respiratory syndrome; 11 events; vs R-HSA-9694614: disease COVID-19; 11 events; 2 event(s) shared
183	170371 REACTOME:R-HSA-9694548 Maturation of spike protein	129320 R-HSA-9683686 Maturation of spike protein (12) 129325 R-HSA-9694548 Maturation of spike protein (60; J 0.161)	12	distinct	evidence	The same name in two virus hierarchies, SARS-CoV-1 and SARS-CoV-2 (PrimeKG parents); R-HSA-9683686: disease severe acute respiratory syndrome; 8 events; vs R-HSA-9694548: disease COVID-19; 12 events; 1 event(s) shared

7.3 Disease (1 group)

#	Entity	Members: x_index PrimeKG id name (rows; J)	Lost	Verdict	Rule	Reason
3	117498 MONDO:31178 monosomy 7 myelodysplasia and leukemia syndrome	29018 MONDO_grouped:9646_30801 Monosomy 7 myelodysplasia and leukemia syndrome (24) 38696 MONDO:31178 monosomy 7 myelodysplasia and leukemia syndrome (4; J 0.000)	2	distinct	evidence	MONDO:0031178 is the parent syndrome; the other node is PrimeKG’s grouping of its subtypes 1 and 2 (MONDO:0009646, MONDO:0030801). PrimeKG links the two with a disease_disease edge, which the merge deleted as a self-loop

7.4 Genes (61 groups) and anatomy (2 groups)

Merge key for every row: both nodes carry one PrimeKG source id, so both resolved to one CURIE at backfill. Verdict: same (J0) for all 63. No row was lost to these merges. In each gene pair, one node carries only disease and phenotype edges and the other carries none of those, so the merge joins two halves of one gene’s record. Both anatomy pairs are one UBERON term under two labels, and there the second label carries most of the tissue-expression edges.

#	Entity	PrimeKG id	Name the entity carries (rows)	Other name (rows)
1	170810 UBERON:468	UBERON:468	63373 multicellular organism (28)	129373 multi-cellular organism (34,710)
2	170926 UBERON:992	UBERON:992	63461 ovary (4)	129374 female gonad (31,684)
105	61477 NCBIGene:64755	NCBI:64755	1419 RUSF1 (168)	34126 C16orf58 (12)
106	61592 NCBIGene:55112	NCBI:55112	7728 DYNC2I1 (102)	34948 WDR60 (16)
107	61769 NCBIGene:57535	NCBI:57535	58296 ELAPOR1 (34)	34122 KIAA1324 (12)
108	62840 NCBIGene:55262	NCBI:55262	8127 TRAPPC14 (62)	35122 MAP11 (2)
109	62987 NCBIGene:6901	NCBI:6901	3236 TAFAZZIN (198)	34675 TAZ (8)
110	63175 NCBIGene:63932	NCBI:63932	1027 STEEP1 (38)	34863 CXorf56 (4)
111	63856 NCBI:4540	NCBI:4540	1368 MT-ND5 (340)	35060 ND5 (10)
112	64134 NCBI:4509	NCBI:4509	5082 MT-ATP8 (146)	35227 ATP8 (4)
113	66178 NCBIGene:115948	NCBI:115948	9854 ODAD3 (88)	35195 CCDC151 (2)
114	66208 NCBIGene:23247	NCBI:23247	11990 KATNIP (24)	35011 KIAA0556 (2)
115	66391 NCBIGene:83538	NCBI:83538	4703 ODAD4 (68)	35191 TTC25 (2)
116	66924 NCBI:4538	NCBI:4538	3101 MT-ND4 (328)	33773 ND4 (12)
117	66984 NCBI:4536	NCBI:4536	3142 MT-ND2 (154)	35239 ND2 (10)
118	67641 NCBIGene:139212	NCBI:139212	10009 DNAAF6 (54)	35196 PIH1D3 (4)
119	67673 NCBIGene:255758	NCBI:255758	13237 DYNLT2B (70)	34952 TCTEX1D2 (16)
120	67795 NCBIGene:4512	NCBI:4512	3606 MT-CO1 (202)	34395 COX1 (22)
121	68126 NCBIGene:4514	NCBI:4514	7358 MT-CO3 (136)	35065 COX3 (12)
122	68589 NCBIGene:55259	NCBI:55259	11159 DNAI7 (26)	33947 CASC1 (4)
123	68934 NCBIGene:113201	NCBI:113201	9792 GOLM2 (48)	34042 CASC4 (2)
124	69034 NCBIGene:2538	NCBI:2538	4377 G6PC1 (80)	34095 G6PC (22)
125	69343 NCBIGene:144577	NCBI:144577	13763 KICS2 (36)	34708 C12orf66 (2)
126	71116 NCBI:4541	NCBI:4541	5789 MT-ND6 (134)	35061 ND6 (12)
127	71142 NCBIGene:4513	NCBI:4513	5811 MT-CO2 (196)	33906 COX2 (20)
128	71591 NCBIGene:23639	NCBI:23639	10538 DNAAF11 (78)	35187 LRRC6 (2)
129	72286 NCBIGene:4535	NCBI:4535	6637 MT-ND1 (254)	34591 ND1 (16)
130	72497 NCBI:4539	NCBI:4539	11604 MT-ND4L (118)	34397 ND4L (8)
131	72856 NCBIGene:54928	NCBI:54928	7103 BPNT2 (90)	35319 IMPAD1 (2)
132	73246 NCBIGene:89891	NCBI:89891	8366 DYNC2I2 (116)	35089 WDR34 (12)
133	73250 NCBIGene:4508	NCBI:4508	7434 MT-ATP6 (140)	34700 ATP6 (18)
134	73334 NCBI:4537	NCBI:4537	7504 MT-ND3 (134)	35059 ND3 (12)
135	73604 NCBIGene:93233	NCBI:93233	8276 ODAD1 (34)	35194 CCDC114 (2)
136	73620 NCBIGene:30834	NCBI:30834	13436 POLR1H (78)	34671 ZNRD1 (10)
137	74479 NCBIGene:57700	NCBI:57700	11243 FHIP2A (14)	22104 FAM160B1 (4)
138	74852 NCBIGene:84529	NCBI:84529	56820 CDIN1 (18)	35291 C15orf41 (6)
139	75550 NCBIGene:4519	NCBI:4519	9427 MT-CYB (174)	34396 CYTB (12)
140	75581 NCBIGene:728215	NCBI:728215	59893 NALF1 (18)	35413 FAM155A (2)
141	76168 NCBIGene:137392	NCBI:137392	13548 CIBAR1 (38)	34545 FAM92A (12)
142	77341 NCBIGene:91574	NCBI:91574	56871 MTRFR (22)	35256 C12orf65 (6)
143	77429 NCBIGene:157657	NCBI:157657	57109 CFAP418 (26)	34503 C8orf37 (6)
144	78308 NCBI:132332	NCBI:132332	83713 SMIM43 (2)	34750 TMEM155 (4)
145	78431 NCBIGene:55130	NCBI:55130	56658 ODAD2 (22)	35189 ARMC4 (2)
146	78965 NCBI:65998	NCBI:65998	61463 ZFTA (4)	35205 C11orf95 (8)
147	89652 NCBIGene:4549	NCBI:4549	22053 RNR1 (10)	78273 MT-RNR1 (86)
148	89851 NCBIGene:4567	NCBI:4567	22109 TRNL1 (20)	78249 MT-TL1 (78)
149	111295 NCBI:339970	NCBI:339970	34784 GCOM2 (30)	60796 POLR2MP1 (10)
150	111835 NCBIGene:4570	NCBI:4570	35150 TRNN (8)	78258 MT-TN (78)
151	111836 NCBIGene:4574	NCBI:4574	35151 TRNS1 (16)	78261 MT-TS1 (78)
152	111840 NCBI:339145	NCBI:339145	35156 FAM92B (10)	57339 CIBAR2 (10)
153	112051 NCBIGene:4575	NCBI:4575	35220 TRNS2 (6)	78268 MT-TS2 (70)
154	112118 NCBIGene:4578	NCBI:4578	35228 TRNW (6)	78256 MT-TW (56)
155	112119 NCBIGene:4511	NCBI:4511	35230 TRNC (2)	78259 MT-TC (78)
156	112120 NCBIGene:4558	NCBI:4558	35231 TRNF (4)	78250 MT-TF (80)
157	112121 NCBI:4564	NCBI:4564	35232 TRNH (4)	78266 MT-TH (78)
158	112122 NCBIGene:4566	NCBI:4566	35233 TRNK (8)	78263 MT-TK (78)
159	112123 NCBIGene:4572	NCBI:4572	35234 TRNQ (4)	78254 MT-TQ (76)
160	112124 NCBIGene:4577	NCBI:4577	35235 TRNV (4)	78251 MT-TV (78)
161	112125 NCBIGene:4565	NCBI:4565	35236 TRNI (2)	78253 MT-TI (70)
162	112126 NCBIGene:4571	NCBI:4571	35237 TRNP (2)	78272 MT-TP (78)
163	112127 NCBI:4568	NCBI:4568	35238 TRNL2 (6)	78269 MT-TL2 (56)
164	112206 NCBI:4556	NCBI:4556	35255 TRNE (6)	78270 MT-TE (78)
165	112209 NCBI:4576	NCBI:4576	35257 TRNT (4)	78271 MT-TT (78)

8. Method

8.1 Data

• Prod, read-only. The snapshot ran in the backend container, against the production database. Every session opened with SET SESSION CHARACTERISTICS AS TRANSACTION READ ONLY, and the script asserts transaction_read_only = on.
  • Tables read: mods_kg_entity, mods_kg_entity_source, mods_kg_relation, mods_kg_curie_cache, mods_kg_entity_premerge, mods_kg_entity_source_premerge.
  • Every relation and source query filters to source_id = 'prime_kg'. Every query is an aggregate, or reads the 183 merged entities and their 401 nodes.
  • Nothing patient-side was read: no maple_patient row, no mods_patient, no document or EHR table.
  • The snapshot was taken on 2026-10-05 at 15:06 PDT and came to 243 KB.
• PrimeKG kg.csv.
  • Path: C:\Maple\worker\data\kg\prime_kg\kg.csv, the file #696’s re-ingest used. 981,751,236 bytes, file date 2026-04-07, 8,100,498 rows.
  • MD5 aac8191d4fbc5bf09cdf8c3c78b4e75f, which matches Harvard Dataverse datafile 6180620 (kg.csv, dataset doi:10.7910/DVN/IXA7BM, version 2.1, released 2022-05-02).
  • SHA-256 0a5a010fcd88dcec572a573d98c26c1c244a11567092b6161d3ded611cfb5c69.
  • The bundle copy at C:\Maple\docs\References\PrimeKG\Data\kg.csv (downloaded 2026-04-09) is byte-identical.
• Evidence for the verdicts, cited in the reason and basis columns:
  • DrugBank descriptions, groups and molecular weights from PrimeKG’s drug_features.csv;
  • MONDO names and BERT groups from disease_features.csv;
  • pathway parents and proteins from kg.csv;
  • Reactome Content Service records (GET https://reactome.org/ContentService/data/query/{stId}, read 2026-10-05).
  • The two feature files are in the same bundle directory: drug_features.csv SHA-256 21ca9cc6…5de73, disease_features.csv SHA-256 163bbcf9…50d50.

8.2 Reproduce

Save §8.4 and §8.5 as files, then, from Git Bash on the host:

# 1. Prod snapshot, read-only (about 30 s, about 243 KB of egress)
docker exec -i BACKEND python - < primekg_merge_snapshot.py > primekg_merge_snapshot.json
 
# 2. Replay and table (local only, about 40 s). Feeding the committed CSV back in
#    carries its proposal columns over and recomputes every measured column.
python primekg_merge_review.py --snapshot primekg_merge_snapshot.json \
    --kg-csv C:/Maple/worker/data/kg/prime_kg/kg.csv \
    --proposal docs/kg_primekg_merge_review.csv \
    --out-csv kg_primekg_merge_review.csv --out-json review_totals.json

On 2026-10-05, step 2 run against the 15:06 PDT snapshot reproduced the committed CSV byte for byte. The script writes LF line endings, and a Windows checkout with core.autocrlf on holds CRLF, so compare with diff --strip-trailing-cr. Its console line reports two checks, and both held:

• the stored relations match the replay on all 33 relation keys;
• they match on all 183 entities, relation by relation.

A later snapshot gives the same table until the store changes.

The proposal columns are written by hand: verdict, partition, rules, reason and basis. The script only carries them over, by group key. The key is the group’s PrimeKG x_index values joined by +, which do not change between environments. A partition lists the proposed nodes after a split: blocks separated by |, members of a block joined by +. To test another proposal, edit those columns and re-run step 2. The JSON’s scenarios then gives the rows each split restores, by relation.

8.3 Column dictionary (CSV)

Column	Meaning
group, group_key	Group number on this page; the members’ x_index values joined by +.
entity_id, entity_curie, entity_name, domain	The canonical entity today.
merge_step, merge_key	Which mechanism joined the group (§2), and the data that shows it.
x_index, primekg_id, primekg_type, primekg_name	The PrimeKG node, as in kg.csv.
premerge_curie, nn_preferred	The node’s entity CURIE in the pre-merge snapshot, and Node Normalizer’s cached answer for it.
anchor	yes for the member whose name the entity carries.
rows, rows_by_relation	kg.csv rows touching the node.
shared_with_anchor, jaccard_with_anchor	Overlap of the node’s (relation, neighbour) set with the anchor’s.
group_rows_lost, group_lost_by_relation	Rows touching the group that the merges collapsed. A row between two merged groups counts in both.
group_rows_restored_if_only_this_split	Rows that splitting this group alone restores.
group_self_loop_rows	Rows between two members, deleted as self-loops.
verdict, partition, rules, reason, basis	The proposal. basis is per node.

8.4 primekg_merge_snapshot.py

"""#438 item 2: read-only snapshot of the PrimeKG entity merges on the canonical KG.
 
Reads only KG tables, only PrimeKG rows (source_id = 'prime_kg'), and only
aggregates or the merged groups themselves. Every transaction is read-only.
Writes one JSON document to stdout.
 
    docker exec -i BACKEND python - < primekg_merge_snapshot.py > primekg_merge_snapshot.json
"""
import datetime
import json
import os
import sys
 
from sqlalchemy import create_engine, text
 
# PrimeKG's x_source label -> standard CURIE prefix, and the zero-padded
# prefixes: the rules of connectors/prime_kg_connector.primekg_curie (#53, #704),
# inlined so the snapshot does not depend on the running container's code.
_PREFIX = {"NCBI": "NCBIGene", "DrugBank": "DRUGBANK", "HPO": "HP"}
_PAD = {"HP": 7, "MONDO": 7, "UBERON": 7, "GO": 7}
 
 
def hint_spellings(src, oid):
    """Every spelling of a node's CURIE the store has used: PrimeKG's own label
    (before #53), the standard prefix (#53), padded and unpadded (#704)."""
    src, oid = (src or "").strip(), (oid or "").strip()
    if not src or not oid:
        return set()
    out = {f"{src}:{oid}"}
    p = _PREFIX.get(src, src)
    out.add(f"{p}:{oid}")
    if _PAD.get(p) and oid.isdigit():
        out.add(f"{p}:{oid.zfill(_PAD[p])}")
        out.add(f"{p}:{int(oid)}")
    return out
 
 
GROUPS = """
    SELECT entity_id FROM mods_kg_entity_source WHERE source_id = 'prime_kg'
    GROUP BY entity_id HAVING COUNT(DISTINCT external_id) > 1
"""
 
engine = create_engine(os.environ["DATABASE_URL"])
out = {"taken_at": datetime.datetime.now(datetime.timezone.utc).isoformat()}
with engine.connect() as conn:
    conn.execute(text("SET SESSION CHARACTERISTICS AS TRANSACTION READ ONLY"))
    conn.commit()
    conn.execute(text("SET statement_timeout = '600s'"))
    assert conn.execute(text("SHOW transaction_read_only")).scalar() == "on"
 
    r = conn.execute(text("""
        SELECT COUNT(*), COUNT(DISTINCT external_id), COUNT(DISTINCT entity_id)
        FROM mods_kg_entity_source WHERE source_id = 'prime_kg'
    """)).one()
    out["prime_kg_source_rows"] = {"rows": r[0], "external_ids": r[1], "entities": r[2]}
    out["external_ids_on_two_entities"] = conn.execute(text("""
        SELECT COUNT(*) FROM (SELECT external_id FROM mods_kg_entity_source
            WHERE source_id = 'prime_kg' GROUP BY external_id
            HAVING COUNT(DISTINCT entity_id) > 1) t
    """)).scalar()
 
    ents = conn.execute(text(f"""
        WITH g AS ({GROUPS})
        SELECT e.id, e.curie, e.name, e.domain, e.is_synthetic_curie
        FROM mods_kg_entity e JOIN g ON g.entity_id = e.id ORDER BY e.id
    """)).fetchall()
    ids = [e.id for e in ents]
    out["entities"] = [{"entity_id": e.id, "curie": e.curie, "name": e.name,
                        "domain": e.domain, "is_synthetic_curie": e.is_synthetic_curie}
                       for e in ents]
 
    mem = conn.execute(text("""
        SELECT entity_id, external_id, source_name, source_entity_type, source_metadata
        FROM mods_kg_entity_source
        WHERE source_id = 'prime_kg' AND entity_id = ANY(:ids)
        ORDER BY entity_id, external_id::bigint
    """), {"ids": ids}).fetchall()
    out["members"] = [{"entity_id": m.entity_id, "external_id": m.external_id,
                       "source_name": m.source_name, "source_entity_type": m.source_entity_type,
                       "original_source": (m.source_metadata or {}).get("original_source"),
                       "original_id": (m.source_metadata or {}).get("original_id")}
                      for m in mem]
 
    # Where each member sat in the pre-merge snapshot that
    # scripts/apply_canonical_entity_merges.py requires before it runs.
    pre = conn.execute(text("""
        SELECT es.external_id, es.entity_id, e.curie
        FROM mods_kg_entity_source_premerge es
        LEFT JOIN mods_kg_entity_premerge e ON e.id = es.entity_id
        WHERE es.source_id = 'prime_kg' AND es.external_id = ANY(:x)
    """), {"x": [m.external_id for m in mem]}).fetchall()
    out["premerge"] = [{"external_id": p.external_id, "entity_id": p.entity_id,
                        "curie": p.curie} for p in pre]
 
    # What Node Normalizer answered for each spelling (mods_kg_curie_cache).
    queries = set(e.curie for e in ents) | set(p.curie for p in pre if p.curie)
    for m in out["members"]:
        queries |= hint_spellings(m["original_source"], m["original_id"])
    cache = conn.execute(text("""
        SELECT query_curie, preferred_curie, preferred_name
        FROM mods_kg_curie_cache WHERE query_curie = ANY(:q)
    """), {"q": sorted(queries)}).fetchall()
    out["cache"] = [{"query_curie": c.query_curie, "preferred_curie": c.preferred_curie,
                     "preferred_name": c.preferred_name} for c in cache]
 
    # Stored PrimeKG relations, for the reconciliation against kg.csv.
    out["stored_by_source_predicate"] = {sp: n for sp, n in conn.execute(text("""
        SELECT source_predicate, COUNT(*) FROM mods_kg_relation
        WHERE source_id = 'prime_kg' GROUP BY 1
    """)).fetchall()}
    out["stored_self_loops"] = conn.execute(text("""
        SELECT COUNT(*) FROM mods_kg_relation
        WHERE source_id = 'prime_kg' AND subject_id = object_id
    """)).scalar()
    # Stored PrimeKG relations touching each merged entity, for a per-group check.
    out["stored_touching_entity"] = [[e, sp, n] for e, sp, n in conn.execute(text(f"""
        WITH g AS ({GROUPS})
        SELECT g.entity_id, r.source_predicate, COUNT(*)
        FROM g JOIN mods_kg_relation r
          ON r.subject_id = g.entity_id OR r.object_id = g.entity_id
        WHERE r.source_id = 'prime_kg'
        GROUP BY 1, 2 ORDER BY 1, 2
    """)).fetchall()]
    conn.rollback()
 
json.dump(out, sys.stdout, indent=1, default=str)

8.5 primekg_merge_review.py

"""#438 item 2: the review table for PrimeKG nodes merged onto one canonical entity.
 
Replays PrimeKG's kg.csv under the merges recorded on prod, and measures what
each merge costs. Local only: reads two or three files, writes two, touches no
database.
 
Inputs
  --snapshot  JSON from primekg_merge_snapshot.py (read-only on prod)
  --kg-csv    PrimeKG kg.csv (Harvard Dataverse file 6180620)
  --proposal  optional: the review CSV (or a JSON of the same fields). Its
              hand-written columns (verdict, partition, rules, reason, basis)
              are carried over by group key; every other column is recomputed.
Outputs
  --out-csv   one row per merged PrimeKG node
  --out-json  the totals the review doc quotes
 
    python primekg_merge_review.py --snapshot primekg_merge_snapshot.json \
        --kg-csv C:/Maple/worker/data/kg/prime_kg/kg.csv \
        --proposal docs/kg_primekg_merge_review.csv \
        --out-csv docs/kg_primekg_merge_review.csv --out-json review_totals.json
 
Units: rows are directed kg.csv rows, the unit mods_kg_relation stores
(PrimeKG writes every edge in both directions). A drug_protein row's relation
key is its role (display_relation), as stored since #696.
"""
import argparse
import csv
import hashlib
import json
from collections import Counter, defaultdict
 
import numpy as np
 
SHIFT = 1 << 21        # node ids < 2**21 (PrimeKG x_index < 129,375; entity block ids offset)
ENT_OFFSET = 1_000_000  # merged entity -> 1e6 + entity_id (< 2**21)
BLOCK_OFFSET = 1_500_000  # a kept-together block of a split group -> 1.5e6 + its smallest x_index
PROPOSAL_COLS = ["verdict", "partition", "rules", "reason"]
 
 
def sha256(path):
    h = hashlib.sha256()
    with open(path, "rb") as f:
        for chunk in iter(lambda: f.read(1 << 20), b""):
            h.update(chunk)
    return h.hexdigest()
 
 
def read_kg(path, members):
    """kg.csv as int arrays (x, y, relation key code), plus node info for members."""
    codes, xs, ys, rs, info = {}, [], [], [], {}
    with open(path, encoding="utf-8", newline="") as f:
        rd = csv.reader(f)
        ix = {h: i for i, h in enumerate(next(rd))}
        for row in rd:
            rel = row[ix["relation"]]
            key = row[ix["display_relation"]] if rel == "drug_protein" else rel
            c = codes.setdefault(key, len(codes))
            x, y = int(row[ix["x_index"]]), int(row[ix["y_index"]])
            xs.append(x); ys.append(y); rs.append(c)
            if x in members and x not in info:
                info[x] = {k: row[ix["x_" + k]] for k in ("id", "type", "name", "source")}
    names = [None] * len(codes)
    for k, v in codes.items():
        names[v] = k
    return np.array(xs, np.int64), np.array(ys, np.int64), np.array(rs, np.int64), names, info
 
 
def keys(gx, gy, r):
    return (gx * SHIFT + gy) * 64 + r
 
 
def kept_by_rel(gx, gy, r, nrel):
    """Distinct stored keys per relation code, self-loops dropped."""
    m = gx != gy
    u = np.unique(keys(gx[m], gy[m], r[m]))
    return np.bincount(u % 64, minlength=nrel)
 
 
def fmt(counter):
    return "; ".join(f"{k} {v:,}" for k, v in sorted(counter.items(), key=lambda kv: (-kv[1], kv[0])) if v)
 
 
def load_proposal(path):
    if not path:
        return {}, {}
    if path.endswith(".json"):
        p = json.load(open(path, encoding="utf-8"))
        return ({g: {c: v.get(c, "") for c in PROPOSAL_COLS} for g, v in p.items()},
                {int(x): b for v in p.values() for x, b in v.get("basis", {}).items()})
    groups, basis = {}, {}
    with open(path, encoding="utf-8", newline="") as f:
        for row in csv.DictReader(f):
            groups[row["group_key"]] = {c: row.get(c, "") for c in PROPOSAL_COLS}
            basis[int(row["x_index"])] = row.get("basis", "")
    return groups, basis
 
 
def parse_partition(text_, xs):
    """'a+b | c' -> [[a, b], [c]]; empty -> one block (the merge as it stands)."""
    if not text_.strip():
        return [sorted(xs)]
    return [[int(v) for v in blk.split("+")] for blk in text_.split("|")]
 
 
def main():
    ap = argparse.ArgumentParser(description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter)
    ap.add_argument("--snapshot", required=True)
    ap.add_argument("--kg-csv", required=True)
    ap.add_argument("--proposal")
    ap.add_argument("--out-csv", required=True)
    ap.add_argument("--out-json", required=True)
    a = ap.parse_args()
 
    snap = json.load(open(a.snapshot, encoding="utf-8"))
    ents = {e["entity_id"]: e for e in snap["entities"]}
    mem = {int(m["external_id"]): m for m in snap["members"]}
    pre = {int(p["external_id"]): p for p in snap["premerge"]}
    cache = {c["query_curie"]: c for c in snap["cache"]}
    groups = defaultdict(list)
    for x, m in mem.items():
        groups[m["entity_id"]].append(x)
    for g in groups.values():
        g.sort()
    proposal, basis = load_proposal(a.proposal)
 
    x, y, r, rel_names, info = read_kg(a.kg_csv, mem)
    nrel = len(rel_names)
    nmax = int(max(x.max(), y.max())) + 1
    ident = np.arange(nmax, dtype=np.int64)
    cur = ident.copy()
    for xi, m in mem.items():
        cur[xi] = ENT_OFFSET + m["entity_id"]
 
    # ── Reconciliation: kg.csv replayed under the merges vs what prod stores ──
    raw = np.bincount(r, minlength=nrel)
    assert len(np.unique(keys(x, y, r))) == len(r), "kg.csv has duplicate (x, y, relation) rows"
    assert not (x == y).any(), "kg.csv has native self-loops"
    replay = kept_by_rel(cur[x], cur[y], r, nrel)
    loops = np.bincount(r[(cur[x] == cur[y])], minlength=nrel)
    stored = snap["stored_by_source_predicate"]
    recon = [{"relation": rel_names[c], "kg_csv_rows": int(raw[c]), "replayed": int(replay[c]),
              "stored_on_prod": int(stored.get(rel_names[c], 0)),
              "lost_to_merges": int(raw[c] - replay[c]), "of_which_self_loops": int(loops[c])}
             for c in np.argsort(-raw)]
    mismatch = [row for row in recon if row["replayed"] != row["stored_on_prod"]]
 
    # ── Per-node neighbours and row counts (rows touching a member) ──────────
    touch = np.isin(x, list(mem)) | np.isin(y, list(mem))
    ti = np.nonzero(touch)[0]
    tx, ty, tr = x[ti], y[ti], r[ti]
    rows_by = defaultdict(Counter)
    nbrs = defaultdict(set)
    for a_, b_, c_ in zip(tx.tolist(), ty.tolist(), tr.tolist()):
        if a_ in mem:
            rows_by[a_][rel_names[c_]] += 1
            nbrs[a_].add((c_, b_))
        if b_ in mem:
            rows_by[b_][rel_names[c_]] += 1
 
    def restored(mapping, sel):
        """Rows restored, per relation, when `mapping` replaces the current one on rows `sel`."""
        before = kept_by_rel(cur[tx[sel]], cur[ty[sel]], tr[sel], nrel)
        after = kept_by_rel(mapping[tx[sel]], mapping[ty[sel]], tr[sel], nrel)
        return after - before
 
    def split_map(mapping, xs, partition):
        for blk in partition:
            for v in blk:
                mapping[v] = v if len(blk) == 1 else BLOCK_OFFSET + min(blk)
        return mapping
 
    stored_touch = defaultdict(dict)
    for eid, sp, n in snap["stored_touching_entity"]:
        stored_touch[eid][sp] = n
    touch_mismatch = []
    out_rows, gsum = [], []
    order = sorted(groups, key=lambda e: (ents[e]["domain"], ents[e]["entity_id"]))
    for gi, eid in enumerate(order, 1):
        e, xs = ents[eid], groups[eid]
        gkey = "+".join(str(v) for v in xs)
        pre_ents = {pre[v]["entity_id"] for v in xs}
        prefs = [cache.get(pre[v]["curie"] or "", {}) for v in xs]
        pref_set = {p.get("preferred_curie") for p in prefs}
        if len(pre_ents) == 1:
            step = "backfill: one CURIE hint for all members"
            src_ids = sorted({f"{info[v]['source']}:{info[v]['id']}" for v in xs})
            key_txt = f"every member carries the same PrimeKG id {', '.join(src_ids)}"
        elif None not in pref_set and len(pref_set) == 1:
            step = "apply_canonical_entity_merges pass 1: Node Normalizer"
            p0 = prefs[0]
            key_txt = (f"Node Normalizer gives {', '.join(pre[v]['curie'] for v in xs)} "
                       f"one preferred CURIE, {p0['preferred_curie']} ('{p0.get('preferred_name')}')")
        else:
            step = "apply_canonical_entity_merges pass 2: name match"
            key_txt = (f"same name, same domain; Node Normalizer knows none of "
                       f"{', '.join(pre[v]['curie'] for v in xs)}")
        anchor = next((v for v in xs if info[v]["name"].lower() == e["name"].lower()), xs[0])
        sel = np.isin(tx, xs) | np.isin(ty, xs)
        # Rows touching the group that the merges collapsed (a row between two
        # merged groups counts in both), and rows that splitting this group
        # alone, with every other merge left in place, would restore.
        kept = kept_by_rel(cur[tx[sel]], cur[ty[sel]], tr[sel], nrel)
        g_lost = np.bincount(tr[sel], minlength=nrel) - kept
        g_restore = restored(split_map(cur.copy(), xs, [[v] for v in xs]), sel)
        g_loops = int(((cur[tx[sel]] == cur[ty[sel]])).sum())
        prod_touch = stored_touch.get(eid, {})
        if {rel_names[c]: int(n) for c, n in enumerate(kept) if n} != prod_touch:
            touch_mismatch.append(eid)
        prop = proposal.get(gkey, {c: "" for c in PROPOSAL_COLS})
        gsum.append({"group": gi, "group_key": gkey, "entity_id": eid, "domain": e["domain"],
                     "step": step, "lost_touching": {rel_names[c]: int(n) for c, n in enumerate(g_lost) if n},
                     "restored_if_only_this_split": {rel_names[c]: int(n) for c, n in enumerate(g_restore) if n},
                     "self_loop_rows": g_loops, **prop})
        for v in xs:
            inter = len(nbrs[v] & nbrs[anchor]); union = len(nbrs[v] | nbrs[anchor])
            out_rows.append({
                "group": gi, "group_key": gkey, "entity_id": eid, "entity_curie": e["curie"],
                "entity_name": e["name"], "domain": e["domain"], "merge_step": step,
                "merge_key": key_txt, "x_index": v,
                "primekg_id": f"{info[v]['source']}:{info[v]['id']}", "primekg_type": info[v]["type"],
                "primekg_name": info[v]["name"], "premerge_curie": pre[v]["curie"],
                "nn_preferred": cache.get(pre[v]["curie"] or "", {}).get("preferred_curie") or "",
                "anchor": "yes" if v == anchor else "",
                "rows": sum(rows_by[v].values()), "rows_by_relation": fmt(rows_by[v]),
                "shared_with_anchor": "" if v == anchor else inter,
                "jaccard_with_anchor": "" if v == anchor else (f"{inter / union:.3f}" if union else ""),
                "group_rows_lost": int(g_lost.sum()),
                "group_lost_by_relation": fmt({rel_names[c]: int(n) for c, n in enumerate(g_lost)}),
                "group_rows_restored_if_only_this_split": int(g_restore.sum()),
                "group_self_loop_rows": g_loops,
                **prop, "basis": basis.get(v, ""),
            })
 
    # ── Scenarios: what a split restores, by relation ─────────────────────────
    allm = np.ones(len(tx), bool)
    all_split = restored(ident, allm)
    scen = {"split_all_183": {rel_names[c]: int(n) for c, n in enumerate(all_split) if n}}
    def cites(p, rule):
        return rule in [t.strip() for t in p.get("rules", "").split(",")]
 
    # A rule that some `same` group cites keeps nodes together (a merge rule);
    # any other rule separates them.
    merge_rules = {t.strip() for p in proposal.values() if p.get("verdict") == "same"
                   for t in p.get("rules", "").split(",") if t.strip()}
 
    def proposal_restores(split_verdicts, flip_rule=None):
        """Rows restored, per relation, by splitting the groups whose verdict is in
        `split_verdicts` (a distinct group along its partition). With `flip_rule`,
        the groups citing that rule are treated as if it were rejected: for a merge
        rule they split into single nodes, for any other rule they stay merged.
        A group citing several rules flips whole, so for those three groups the
        figure is an upper bound. Unclear groups are never flipped."""
        mp = cur.copy()
        for eid in order:
            xs = groups[eid]
            p = proposal.get("+".join(map(str, xs)), {})
            v = p.get("verdict")
            if flip_rule and cites(p, flip_rule) and v != "unclear":
                v = "split-whole" if flip_rule in merge_rules else "same"
            if v == "split-whole":
                split_map(mp, xs, [[x_] for x_ in xs])
            elif v == "distinct" and v in split_verdicts:
                split_map(mp, xs, parse_partition(p.get("partition", ""), xs))
            elif v in split_verdicts:
                split_map(mp, xs, [[x_] for x_ in xs])
        return restored(mp, allm)
 
    per_rule = {}
    if proposal:
        for name, verdicts in (("split_distinct", {"distinct"}), ("split_distinct_and_unclear", {"distinct", "unclear"})):
            res = proposal_restores(verdicts)
            scen[name] = {rel_names[c]: int(n) for c, n in enumerate(res) if n}
        base = sum(scen["split_distinct"].values())
        rules_seen = sorted({t.strip() for p in proposal.values() for t in p.get("rules", "").split(",") if t.strip()})
        for rule in rules_seen:
            flipped = int(proposal_restores({"distinct"}, flip_rule=rule).sum())
            per_rule[rule] = {"groups": sum(1 for p in proposal.values() if cites(p, rule)),
                              "kind": "merge" if rule in merge_rules else "split",
                              "rows_restored_if_rejected": flipped, "change_vs_proposal": flipped - base}
 
    cols = ["group", "group_key", "entity_id", "entity_curie", "entity_name", "domain", "merge_step",
            "merge_key", "x_index", "primekg_id", "primekg_type", "primekg_name", "premerge_curie",
            "nn_preferred", "anchor", "rows", "rows_by_relation", "shared_with_anchor",
            "jaccard_with_anchor", "group_rows_lost", "group_lost_by_relation",
            "group_rows_restored_if_only_this_split", "group_self_loop_rows", *PROPOSAL_COLS, "basis"]
    with open(a.out_csv, "w", encoding="utf-8", newline="") as f:
        w = csv.DictWriter(f, fieldnames=cols, lineterminator="\n")
        w.writeheader()
        w.writerows(out_rows)
    totals = {
        "kg_csv": {"path": a.kg_csv, "sha256": sha256(a.kg_csv), "rows": int(len(r))},
        "snapshot_taken_at": snap["taken_at"],
        "prime_kg_source_rows": snap["prime_kg_source_rows"],
        "groups": len(groups), "member_nodes": len(mem), "surplus_nodes": len(mem) - len(groups),
        "by_domain": {d: {"groups": sum(1 for e in groups if ents[e]["domain"] == d),
                          "nodes": sum(len(g) for e, g in groups.items() if ents[e]["domain"] == d)}
                      for d in sorted({e["domain"] for e in ents.values()})},
        "by_step": dict(Counter(g["step"] for g in gsum)),
        "reconciliation": recon, "reconciliation_mismatches": mismatch,
        "replayed_total": int(replay.sum()), "stored_total": sum(stored.values()),
        "stored_self_loops": snap["stored_self_loops"],
        "groups_whose_stored_rows_differ_from_replay": touch_mismatch,
        "scenarios": scen, "per_rule": per_rule, "groups_detail": gsum,
    }
    json.dump(totals, open(a.out_json, "w", encoding="utf-8"), indent=1)
    print(f"{len(groups)} groups, {len(mem)} nodes; replayed {int(replay.sum()):,} rows vs "
          f"{sum(stored.values()):,} stored; mismatched relations: {len(mismatch)}; "
          f"groups whose stored rows differ from the replay: {len(touch_mismatch)}")
    for k, v in scen.items():
        print(f"{k}: {sum(v.values()):,} rows  {fmt(Counter(v))}")
 
 
if __name__ == "__main__":
    main()