Clarification on CURIEs not normalizing correctly?

[piotrkan]

Hi, first of all thanks for this great tool, it helped us a lot with mapping some of the concepts between different biomedical networks 🚀

We were doing some EDA analysis and found some entities being semantically the same but not mapping to the same IDs after running nodeNorm with /get_normalized_nodes endpoint (conflation=True; drug_chemical=False). Is there any reason for that?

Please find the 'divergent' nodes below; if either curie_1 or curie_2 is sent as a request to /get_normalized_nodes we get two different CURIE outputs while in theory they should be the same:

entity_name	curie_1	curie_1_label	curie_2	curie_2_label
lovotibeglogene autotemcel	UMLS:C5447474	lovotibeglogene autotemcel	DRUGBANK:DB18680	Lovotibeglogene autotemcel
coagulation factor xiii	NCBIGene:100153504	F13A1	UniProtKB:A0A8D0SF31	A0A8D0SF31_PIG Coagulation factor XIII A chain (trembl)
lebrikizumab	UNII:U9JLP7V031	lebrikizumab	DRUGBANK:DB11914	Lebrikizumab
haem arginate	UNII:R1B526117P	Heme arginate	PUBCHEM.COMPOUND:135564839	Haem arginate
nogapendekin alfa	CHEMBL.COMPOUND:CHEMBL4297690	NOGAPENDEKIN ALFA	RXCUI:2682677	nogapendekin alfa inbakicept-pmln 1 MG/ML
botulinum toxin type a	CHEBI:3160	botulinum toxin type A	DRUGBANK:DB00083	Botulinum toxin type A
samarium sm-153 lexidronam	UNII:745X144DZY	Unii-745X144dzy	UNII:7389YR3OOV	Samarium Sm-153 Lexidronam Pentasodium
atidarsagene autotemcel	UMLS:C5556517	atidarsagene autotemcel	DRUGBANK:DB17538	Atidarsagene autotemcel
idecabtagene vicleucel	UMLS:C3899973	idecabtagene vicleucel	DRUGBANK:DB16665	Idecabtagene vicleucel
afamitresgene autoleucel	UMLS:C5448035	afamitresgene autoleucel	DRUGBANK:DB18592	Afamitresgene autoleucel
vilobelimab	UNII:F5T0RF9ZJA	vilobelimab	PUBCHEM.COMPOUND:137254326	Vilobelimab
pancrelipase	CHEBI:81916	Pancrelipase	UNII:FQ3DRG0N5K	Pancrelipase
donislecel	UMLS:C5707881	donislecel	DRUGBANK:DB17961	Donislecel
toripalimab	UNII:8JXN261VVA	toripalimab	DRUGBANK:DB15043	Toripalimab
ziconotide	CHEBI:142406	ziconotide	PUBCHEM.COMPOUND:16135415	Ziconotide
berdazimer	UNII:ORT9SID4QY	berdazimer sodium	DRUGBANK:DB18712	Berdazimer
fitusiran	CHEMBL.COMPOUND:CHEMBL4297754	FITUSIRAN	DRUGBANK:DB15002	Fitusiran
cosibelimab	CHEMBL.COMPOUND:CHEMBL4297729	COSIBELIMAB	GTOPDB:13676	cosibelimab
dibucaine	CHEBI:247956	Cinchocaine	UNII:J31043J63M	Dibucaine benzoate
lifileucel	UMLS:C4053624	lifileucel	DRUGBANK:DB17107	Lifileucel
coagulation factor x	PR:000007294	coagulation factor X	DRUGBANK:DB13148	Coagulation factor X human
marstacimab	UNII:0UB3OA67O7	marstacimab	DRUGBANK:DB17725	Marstacimab
thymoglobulin	UNII:D7RD81HE4W	thymoglobulin	UMLS:C0359156	Thymoglobulin
pegzilarginase	CHEMBL.COMPOUND:CHEMBL4297803	PEGZILARGINASE	UNII:4YV4KW88GD	PEGZILARGINASE
anti-d immunoglobulin	DRUGBANK:DB09312	Antilymphocyte immunoglobulin (horse)	UMLS:C5192004	Human anti-D immunoglobulin-containing product
tarlatamab	GTOPDB:12969	tarlatamab	DRUGBANK:DB17256	Tarlatamab
axatilimab	UNII:R96Z451BMC	axatilimab	DRUGBANK:DB16388	Axatilimab
interferon gamma	PR:000000017	interferon gamma	DRUGBANK:DB15753	Interferon Gamma
bleomycin	CHEBI:3139	bleomycin	CHEBI:22907	Bleomycin
pericyazine	CHEBI:31981	Periciazine	PUBCHEM.COMPOUND:71751521	Pericyazine-d4

[eKathleenCarter]

lovotibeglogene autotemcel	UMLS:C5447474	lovotibeglogene autotemcel	DRUGBANK:DB18680

I was just curious about this and looked at the first one. It's interesting especially because the UMLS term is a biolink:Cell and the Drugbank term is a biolink:ChemicalEntity

UMLS lists the Drugbank term as an equivalent identifier. I think the problem is that it is a drug, but the drug is a preparation of cells.

@gaurav @cbizon

{
  "UMLS:C5447474": {
    "id": {
      "identifier": "UMLS:C5447474",
      "label": "lovotibeglogene autotemcel"
    },
    "equivalent_identifiers": [
      {
        "identifier": "UMLS:C5447474",
        "label": "lovotibeglogene autotemcel"
      }
    ],
    "type": [
      "biolink:Cell",
      "biolink:AnatomicalEntity",
      "biolink:PhysicalEssence",
      "biolink:OrganismalEntity",
      "biolink:SubjectOfInvestigation",
      "biolink:BiologicalEntity",
      "biolink:ThingWithTaxon",
      "biolink:NamedThing",
      "biolink:PhysicalEssenceOrOccurrent"
    ]
  },
  "DRUGBANK:DB18680": {
    "id": {
      "identifier": "DRUGBANK:DB18680",
      "label": "Lovotibeglogene autotemcel"
    },
    "equivalent_identifiers": [
      {
        "identifier": "DRUGBANK:DB18680",
        "label": "Lovotibeglogene autotemcel"
      },
      {
        "identifier": "RXCUI:2671958",
        "label": ""
      }
    ],
    "type": [
      "biolink:ChemicalEntity",
      "biolink:PhysicalEssence",
      "biolink:ChemicalOrDrugOrTreatment",
      "biolink:ChemicalEntityOrGeneOrGeneProduct",
      "biolink:ChemicalEntityOrProteinOrPolypeptide",
      "biolink:NamedThing",
      "biolink:PhysicalEssenceOrOccurrent"
    ]
  }
}

[cbizon]

Entity Name	Evaluation	Notes
afamitresgene autoleucel	Different types (cell/chemical)	—
anti-d immunoglobulin	Different species	Horse / Human
atidarsagene autotemcel	Different types (cell/chemical)	—
axatilimab	Different types (chemical/protein)	MESH:C000711669 is in both cliques so this looks like maybe a conflation error
berdazimer	Should not merge, different salt	These are different structures and the drug/chemical doesn't happen to pull them together
bleomycin	Should not merge	Bleomycin is a family, Bleomycin A2 is a member of the family (along with several others)
botulinum toxin type a	Different types (chemical/protein)	UMLS:0006050 is in both, this appears to be a conflation failure
coagulation factor x	Should merge	There's probably also another clique floating around with the gene in it
coagulation factor xiii	Requires further investigation	These are both pig proteins, but the UniProt one is obsolete (or nearly so). So probably that clique will just go away, or will just be a bunch of ensembls, probably isoforms
cosibelimab	Dangling CHEMBL	This one is pretty tough - all we have is a chembl and a gtopdb. Is there another clique out there? And no structure :P
dibucaine	Should not merge, different salt	Different salt (one is parent, one is benzoate)
donislecel	Different types (cell/chemical)	—
fitusiran	Dangling CHEMBL	Straggling CHEMBL Compound (no structure)
haem arginate	Should not merge, different salt	The inchikeys are different here, but one of the cliques doesn't have any members with a picture...
idecabtagene vicleucel	Different types (cell/chemical)	—
interferon gamma	Requires further investigation	I think that these are different species (the PR is species independent, I THINK that the MESH is human)
lebrikizumab	Different types (chemical/protein)	Conflation failure - mesh is in both.
lifileucel	Different types (cell/chemical)	—
lovotibeglogene autotemcel	Different types (cell/chemical)	—
marstacimab	Different types (chemical/protein)	Conflation Failure
nogapendekin alfa	Dangling CHEMBL	Straggling CHEMBL
pancrelipase	Should merge	Probably should merge? One clique has disodium salt, but the other is no structure.
pegzilarginase	Dangling CHEMBL	UNII/CHEMBL only
pericyazine	Should not merge	One has deuteriums instead of hydrogens.
samarium sm-153 lexidronam	Should not merge, different salt	Different salts
tarlatamab	Should merge	conflation problem? Dangling gtopdb?
thymoglobulin	Different types (chemical/protein)	conflations failure, mesh in both
toripalimab	Different types (chemical/protein)	Conflation problem (share mesh)
vilobelimab	Should merge	Why doesn't the mesh merge with the UMLS? I guess because there isn't a UMLS? But there is a RXCUI?
ziconotide	Requires further investigation	There are different inchikeys, so I guess they shouldn't merge, but not totally sure why

Thanks for the examples @piotrkan ! It looks like there are a few different cases:

• Cell based treatments. (6) This is what @eKathleenCarter is pointing out above. This is one that frankly we hadn't considered yet, so we'll need to think it through
• Chemicals / Proteins (6): These are cases where there are two cliques for the same concept. One's called a chemical and one's called a Protein. This is something that we've been working on. Probably @gaurav could comment on the current situation.
• Should not merge (different salt): Our chemical merging can be pretty strict - it's (as much as possible) structure based. So different structures = different cliques. This includes some subtle differences like whether the structure is a different salt. In many cases this gets handled by drug/chemical conflation which has the side effect of causing different salt forms to be brought into a clique. But it's not guaranteed, and it won't happen that way unless there is a drug.
• There are a few others that should merge but don't, mostly indicating that we need another mapping somewhere. This includes both the ones that say "should merge" and the "dangling ChEMBL" - it looks to me like our ChEMBL mappings might be a place to take a bit of a look at expanding.
• There's one where the proteins are not the same species
• And finally there are a few where I am not quite sure what's going on :)

[jxneli]

Hello, just chiming in from a medical perspective. Anti-d immunoglobulin is a horse antibody against a human protein (a red blood cell surface marker - Rhesus D) and is given to humans as a therapeutic so both of those IDs are describing the the same thing!

[cbizon]

Thanks @jxneli, that's very interesting! I just want to double check, though. One of the cliques is what you are describing for sure. Here's a description from UMLS:

A purified, concentrated and sterile solution of horse-derived gamma globulin obtained by the immunization of horses with human T-lymphocytes, with T-cell depleting and immunosuppressive activities. Upon administration, horse anti-thymocyte globulin (hATG) specifically recognizes, modulates and destroys T-lymphocytes. Although the exact mechanism of action by which hATG causes immunosuppression is not completely understood, it is likely caused by a combination of T-lymphocyte depletion, reduction in T-cell activation and modulation of their cytotoxic activities. Administering of hATG prior to transplantation may reduce the risk of graft-versus-host disease (GvHD). (NCI)"

The other clique though just has a single UMLS term made from SNOMED with this description:

Product containing human anti-D immunoglobulin (medicinal product)

I interpret the "human" there as saying that the antibody is human? But I don't really know what I'm talking about here so let me know if you think that's wrong.

[jxneli]

Thanks @cbizon for clarifying. I have just re-read and realised the horse antibody described is actually targeting T lymphocytes (rather than red blood cells) so they are definitely different concepts! The latter one is the one we were trying to normalise to although I'm surprised there is only one ID in the clique!

[piotrkan]

thanks all for explaining these examples, I really enjoy learning about this stuff

For the IDs which do get merged - will these be added to nodenorm on a case-by-case basis or rather one root cause of the divergence is found & fixed? if I get it correctly, several of those examples seem to be related to databases (eg UMLS which I know tends to be noisy/painful to work with) so wondering whats the right approach here

[cbizon]

In general, we prefer to find a small set of root causes that fixes a whole bunch of cliques at once. So for instance, the chemical/protein or cell/drug ones are going to need to have some strategic fixes implemented that will solve all of them. Sometimes, a problem is due to data that either missing or wrong in an an upstream source and in those cases we do have a way to handle fixes one-by-one.