incorrect subclass-of edge in KG2.9.0pre

[saramsey]

This issue was brought to our attention by Kaiwen He of Team Unsecret Agent.

"UMLS:C2267095" ("M [Preparations]") -"biolink:subclass_of"-> "UMLS:C0025598"
 ("metformin")
"UMLS:C0025424" ("Mercury") -"biolink:subclass_of"->  "UMLS:C2267095" ("M [Preparations]")

This looks to be coming from UMLS MEDRT:

{"domain_range_exclusion": false, "id": "UMLS:C2267095---UMLS:has_parent---None---None---None---UMLS:C0025598---umls_source:MED-RT", "negated": false, "object": "UMLS:C0025598", "predicate": "biolink:subclass_of", "predicate_label": "has_parent", "primary_knowledge_source": "infores:medrt-umls", "publications": [], "publications_info": {}, "qualified_object_aspect": null, "qualified_object_direction": null, "qualified_predicate": null, "relation_label": "has_parent", "source_predicate": "UMLS:has_parent", "subject": "UMLS:C2267095", "update_date": "2023"}

{"domain_range_exclusion": false, "id": "UMLS:C2267095---UMLS:parent_of---None---None---None---UMLS:C0025424---umls_source:MED-RT", "negated": false, "object": "UMLS:C2267095", "predicate": "biolink:subclass_of", "predicate_label": "parent_of", "primary_knowledge_source": "infores:medrt-umls", "publications": [], "publications_info": {}, "qualified_object_aspect": null, "qualified_object_direction": null, "qualified_predicate": null, "relation_label": "INVERTED:parent_of", "source_predicate": "UMLS:parent_of", "subject": "UMLS:C0025424", "update_date": "2023"}

[saramsey]

this bug seems to be an issue in UMLS source MED-RT, which is providing edges with relation_label of has_parent and direction of N, and with the drug class and the drug swapped compared to what they should be; I think MED-RT should be using a Y of the relation is really has_parent; though the relation label is CHD so it may well be that they should be using relation label parent_of and keeping the direction N.

In any event, although I'd prefer to patch this in the extract script, in the interest of time, I'm putting a patch in the umls_util.py module

[saramsey]

@ecwood can take a look when they are back in Summer 2025

[saramsey]

    "relations": {
      "MED-RT": {
        "CHD,has_parent,N": [
          "C0025598"
        ],

note, C0025598 is metformin, the drug (not the drug class)

[saramsey]

Note, I have not actually tested this code yet. Need to do that.

[saramsey]

I tested this on kg2102build.rtx.ai in /home/ubuntu/issue-377 and verified that it works. With the above commit, the new block of JSON for the edge is:

{
  "domain_range_exclusion": false,
  "id": "UMLS:C0025598---UMLS:has_parent---None---None---None---UMLS:C2267095---umls_source:MED-RT",
  "negated": false,
  "object": "UMLS:C2267095",
  "predicate": null,
  "primary_knowledge_source": "umls_source:MED-RT",
  "publications": [],
  "publications_info": {},
  "qualified_object_aspect": null,
  "qualified_object_direction": null,
  "qualified_predicate": null,
  "relation_label": "has_parent",
  "source_predicate": "UMLS:has_parent",
  "subject": "UMLS:C0025598",
  "update_date": "2023"
}

[saramsey]

On kg2102build.rtx.ai, the old file (/home/ubuntu/kg2-build/kg2-umls-edges.jsonl) has:

grep 'MED-RT' /home/ubuntu/kg2-build/kg2-umls-edges.jsonl | grep has_parent | grep C0025598 | jq . | less
{
  "domain_range_exclusion": false,
  "id": "UMLS:C2267095---UMLS:has_parent---None---None---None---UMLS:C0025598---umls_source:MED-RT
",
  "negated": false,
  "object": "UMLS:C0025598",
  "predicate": null,
  "primary_knowledge_source": "umls_source:MED-RT",
  "publications": [],
  "publications_info": {},
  "qualified_object_aspect": null,
  "qualified_object_direction": null,
  "qualified_predicate": null,
  "relation_label": "has_parent",
  "source_predicate": "UMLS:has_parent",
  "subject": "UMLS:C2267095",
  "update_date": "2023"
}

[saramsey]

It looks like this bugfix works. But will not close until we can rerun the cypher in the next build.

match (n {id: 'UMLS:C2267095'})-[r]-(m {id: 'UMLS:C0025598'}) return n,m

should show two subclass_of edges, both pointing from metformin to the drug class M [preparations]

[saramsey]

I believe that this issue only affects MED-RT UMLS edges of relation label has_parent. Thus, it does not affect all UMLS edges or even all MED-RT UMLS edges. Thus this issue does not justify terminating and restarting the in-progress KG2.10.2pre build.

[saramsey]

The number of edges affected is 20,171:

grep 'MED-RT' /home/ubuntu/kg2-build/kg2-umls-edges.jsonl | grep has_parent | wc -l
20171

[saramsey]

Confirmed this bug is still present in KG2.10.2pre.

Trying to reconstruct the chronology from the above messages, it looks like this patch was too late for the KG2.10.2pre build. Thus, I think we will need to test it in the KG2.10.3pre build.

[ecwood]

This seems to be solved in KG2.10.3:

Closing.