QA of ChEMBL KGX ingest

[vdancik]

QA Resources:

• KGX Summary report: (chembl_KGX_SUMMARY)
• RIG: (chembl_rig.yaml)
• QC update PR: update ChEMBL ingest RIG #278

Issues to consider / address for January updated release:

• ChemicalEntity is_substrate_of GeneOrGeneProduct
  • canonical predicate has_substrate should be used
    • fixed in ec813a0

Issues to consider longer term:

• MacromolecularComplex has_part Protein

  • failed normalization - ChEMBL complexes are not included in NN
     

• Biolink dependency: add species_context_qualifier and context_qualifier to ChemicalToChemicalAssociation biolink/biolink-model#1631

• Biolink dependency: add injection to DrugDeliveryEnum biolink/biolink-model#1633

• Biolink dependency: add routes of delivery to ChemicalEntity and its children biolink/biolink-model#1634

Questions:

[mbrush]

• Generally, the RIG is not up to date (it as predicates and edge properties not in the KGX summary reports of the data, and is missing some values I do see in the data, etc). This makes it hard for me to complete my QA, as I don’t know if discrepancies are due to the RIG being out of date, or the ingest code not doing things correctly.

  • (VD) @mbrush I have updated the RIG, additional comments below:
  • MHB: @vdancik I recorded some feedback in the update ChEMBL ingest RIG #278 as comments with the prefix "MHB" - where the spec therei seems out of sync with the data - per the summary report here. Please review / address all of these comments and check this as completed when resolved.

• Where confidence scores in the data come from? Not in RIG, and I don’t recall these in the Chembl data?

  • (VD) these are coming from ChEMBL database and are listed on line 96 of RIG for biolink:affects; I added it to biolink:directly_physically_interacts_with. It can have the following values:
    0 Default value - Target unknown or has yet to be assigned Unassigned
    1 Target assigned is non-molecular Non-molecular
    2 Target assigned is subcellular fraction Subcellular fraction
    3 Target assigned is molecular non-protein target Molecular (non-protein)
    4 Multiple homologous protein targets may be assigned Multiple homologous proteins
    5 Multiple direct protein targets may be assigned Multiple proteins
    6 Homologous protein complex subunits assigned Homologous protein complex
    7 Direct protein complex subunits assigned Protein complex
    8 Homologous single protein target assigned Homologous protein
    9 Direct single protein target assigned Protein

    MHB: the convention for our initial ingests is to create a dedicated new edge property in the attributes.yaml file for all bespoke confidence properties provided by a given source. Name this as "chembl_[name of property in chembl]" (see examples of other bespoke confidence score properties in the file). Use this new property in the data for now, rather than the existing has_confidence_score biolink property.

• Where does the species and anatomical context qualifier values in the data come from? Not mentioned in RIG, and I don’t recall this being provided by chembl.

  • (VD) it is provided by ChEMBL database, added to RIG

• 'has part' edges are specified in the RIG - but I don’t see these in the data?

  • (VD) this is a failure of Node Normalizer - ChEMBL IDs for protein complexes are not included in NN

• The KGX summary report includes 'automated agent' as a value from many edge types - but this is never mentioned in the RIG . . . Where is this coming from?

  • (VD) ChEMBL specifies a level of curation, added to RIG:
    agent_type=AgentTypeEnum.automated_agent if record["curated_by"] == "Autocuration" else AgentTypeEnum.manual_agent

• I don’t see degradation as an object_aspect value in the summary report, but our interaction type to predicate/qualifier mappings specified this would be used.

  • (VD) It is specified as causal_mechanism_qualifier:

    "DEGRADER": {
        "association": "ChemicalAffectsGeneAssociation",
        "predicate": "biolink:affects",
        "qualifiers": {
            "qualified_predicate": "biolink:causes",
            "object_aspect_qualifier": "abundance",
            "object_direction_qualifier": "increased",
            "causal_mechanism_qualifier": "degradation"
        }
    },

[mbrush]

Hi Vlado - I did my review of the KGX summary report and updated RIG. A couple additional questions below:

• The Chembl action type to SPOQ + mechanism mapping table here (column H) suggests that we should see some edges using the following aspects - but I don't see them in the KGX summary report:
  - expression
  - methylation
  - oxidation
  - reduction

• The mapping table also suggests I should see some directly_physically_interacts_with edges with a sequestration mechanism - but I do not (maybe none in the data?)

• There is way for me to confirm from the summary reports if additional 'directly physically interacts with' edges are being created from affects edges that use the mechanisms we decided warrant this. Can you check / confirm on your end? I see ~14K I directly physically interacts with in the graph, and almost 250K affects edges. Intuitively I would expect more 'directly physically interacts with' edges than this given that - as almost half (13/34) the mapping rows on the s/s would result in such an edge.