What format does Expander expect for subclassing?

[edeutsch]

From: Willow Callaghan (DOGSURF)
1:20 PM
Hi all, wanted to double-check something about the upcoming subclass handling:
Our intent is to have Retriever condense subclasses into support graphs, as BTE did/as was the intended approach going forward in the last phase. This has a few implications:
Retriever will produce constructed edges for which infores:retriever is the primary_knowledge_source, with the underlying subclass edge and knowledge edge primary sources as supporting data sources
Constructed edges that Retriever responds with will have a support graph, linking to an aux graph with the subclass edge(s) and actual knowledge edge(s) connecting the original query nodes via the intermediate subclass
Retriever will not make use of query_id as it will be binding the original nodes, not the subclass nodes.
Is this what everyone expects? Is this acceptable? Will downstream tools work with this format/be updated to work with this format? Particularly looking for feedback from Shepherd/ARA, ARS, UI teams

Does anyone with knowledge of the inner workings of Expander know the answer to this question?

[dkoslicki]

Dec 10 team meeting: Plover is the reference implementation; no one else knows; each thinks quantitative test is required

[saramsey]

I have reached out to Willow by DM to make sure I understand the question.

[saramsey]

Am assuming that by "subclass-of" reasoning, Willow is referring to the fact that PloverDB, if given a query graph like

drug-[treats]->diabetes

will return triples like

metformin-[treats]->type_2_diabetes

where type_2_diabetes related to the concept node diabetes in the original query graph by a subclass_of relationship.

[saramsey]

Willow actually posted a query graph in the #general channel in the NCATSTranslator Slack workspace, along these lines:

"query_graph": {
      "nodes": {
        "sbj": {
          "categories": [ "biolink:Drug" ],
          "ids": [ "CHEBI:6801" ]
        },
        "obj": {
          "categories": [ "biolink:Disease" ],
          "ids": [ "MONDO:0005015" ]
        }
      },
      "edges": {
        "e1": {
          "subject": "sbj",
          "predicates": [ "biolink:treats_or_applied_or_studied_to_treat" ],
          "object": "obj"
        }
      }

[saramsey]

Note, MONDO:0005015 is "diabetes mellitus" and CHEBI:6801 is "metformin"

[saramsey]

I took Willow's query graph (wrapped in an object with a message key as expected by PloverDB) and put it in a file qg-2592.json,

{"message": {
    "query_graph": {
        "nodes": {
            "sbj": {
                "categories": [ "biolink:Drug" ],
                "ids": [ "CHEBI:6801" ]
            },
            "obj": {
                "categories": [ "biolink:Disease" ],
                "ids": [ "MONDO:0005015" ]
            }
        },
        "edges": {
            "e1": {
                "subject": "sbj",
                "predicates": [ "biolink:treats_or_applied_or_studied_to_treat" ],
                "object": "obj"
            }
        }
    }
}
}

and wrote a little shell script run-test.sh that runs it through the CI KG2 PloverDB,

curl -s \
     -X POST\
     'https://kg2cploverdb.ci.transltr.io/query' \
     -H "accept: application/json" \
     -H "Content-Type: application/json" \
     -d @qg-2592.json \
    | jq . \
         >results-2592.json

and ran it. The result is a 1,952 line JSON file, results-2592.json, which I will attach here.

[saramsey]

results-2592.json

[saramsey]

I'm inspecting the results from the PloverDB query. I observe that:

There are no subclass-of edges that appear explicitly in the knowledge_graph returned, though clearly PloverDB did subclass-of reasoning on the "diabetes" (MONDO:0005015) node, since I see MONDO:0005148 (type 2 diabetes) in the graph.

[saramsey]

The query_graph is pretty much unchanged:

other than the trivial change that the edge keys have been renamed.

[saramsey]

Now let's inspect an object from the results section that refers to MONDO:0005148 (type 2 diabetes). Result 1 seems to fit the bill,

we can see that that result node-binds MONDO:0005148 to MONDO:0005015.

So PloverDB seems to indicate subclass-of reasoning, at least for objects like MONDO:0005015 (here I am specifically not talking about Biolink predicates or categories), through node binding in the results section.

I do not see any aux graphs or supporting graphs in the PloverDB output.

Note: this is not ARAX output. This is PloverDB output.

[saramsey]

TL;DR: ARAX-Expand should be happy with results that indicate subclass-of reasoning through node binding

[edeutsch]

But it sounds like Willow intends to do something different (use of aux graphs) with Retriever, right?

So this would not be compatible with ARAX? Or would ARAX tolerate the aux graphs just fine?

Do we want to raise a fuss? Or just let it happen and we'll adjust ARAX if needed?

[saramsey]

I have pointed out in the Slack thread, that the proposed result format seems to violate Architecture Principle 10.i in the NCATSTranslator/TranslatorEngineering README:

[saramsey]

If Retriever is going to implement this proposed way of representing subclass-of reasoning in results, I suggest we may want to do some empirical testing to see how ARAX will react to it.