Is KG2 working correctly?

[edeutsch]

I've been having some problems with one of my smoke tests for KG2 since we upgraded and finally now trying to fix it. It's not returning expected answers.

Today when you ask production KG2 our standard Example 1 query, the results seem very strange to me. Please go to: https://arax.ncats.io/kg2/ And issue the Example 1 JSON query: Which proteins physically interact with acetaminophen. Usually with past KG2s we expect PTGS1 and 2 and friends. The top hits now are ANAKINRA and GOLIMUMAB and gets weirder from there. Is this sensible or is something off the rails here? @saramsey

[saramsey]

Yes, this is because the Biolink model says to map DRUGBANK:drug-interaction to biolink:physically_interacts_with:

https://github.com/biolink/biolink-model/blob/450cb767f03da6f171b20809dd4a45aed20bb6a4/biolink-model.yaml#L2314-L2341

so that is what we are doing in KG2:

https://github.com/RTXteam/RTX-KG2/blob/4262741b1805aa50eed616583e617b6aa7ffc59b/predicate-remap.yaml#L715-L717

There is, in fact, a drug-drug interaction between acetaminophen (DRUGBANK:DB00316) and the biologic therapeutic agent Anakinra (DRUGBANK:DB00026):

and since the latter is a biologic, it gets tagged with a category label of biolink:Protein. I have already put in an issue with Biolink about that we cannot be mapping DRUGBANK:drug-interaction to biolink:physically_interacts_with; see biolink/biolink-model issue 1371.

[saramsey]

Per @dkoslicki: it looks better if you use biolink:affects

[saramsey]

The situation looks better on kg2.test.transltr.io. I'm going to close this out.