colleenXu
commented
3 years ago TDLR:
- bug with predicate restriction. When BTE correctly flips an edge for querying, some fix is needed so the predicate restriction is then done correctly.
- Potential fixes: Either (1) the predicate restriction has to use the inverse predicate for filtering/restriction OR (2) the answer edges all need to flip predicates first, then predicate restriction using the query edge's predicate can be done.
- [EDIT: andrew says we should be able to handle this even with the latest biolink. I think this involves checking, possible changes to the biolink JS package]
- potential problem with (1) is that the latest version of biolink model only lists inverses on 1 predicate of the pair. In that situation, I'm then not sure if BTE can recognize that gene_associated_with_condition is the inverse of condition_associated_with_gene AND vice versa.
- potential problem with (2) is that this may involve changed logic for what order things are done in...
- [EDIT: andrew says we should be able to handle this even with the latest biolink. I think this involves checking, possible changes to the biolink JS package]
- once this bug is fixed, we want to see if results look as-expected after #164 is fixed.
Comparing:
- "classic Predict": KCNMA1 (HGNC:6284) -(gene_associated_with_condition)-> Disease, then Disease -(condition_associated_with_gene)-> Gene (structure: A -> B -> C)
- "modified Predict for canonical predicate direction): KCNMA1-(gene_associated_with_condition)-> Disease, then Disease <-(gene_associated_with_condition)- Gene (structure: A -> B <- C).
Expected behavior:
These two scenarios should return the same knowledge_graph nodes / go through the same sub-queries. They would return different edges/results because the edges for the second hop (between B and C) would have different directions/predicates.
Current behavior:
Scenario 2 doesn't work as planned. It looks like this is because of the predicate restriction in the second step.
BTE does the querying correctly:
First: it queries from KCNMA1 -> Disease (set of IDs), then does predicate restriction to only keep nodes/edges/results with the predicate "gene_associated_with_condition". The logs for the predicate restriction are below:
{
"timestamp": "2021-06-24T16:12:18.959Z",
"level": "DEBUG",
"message": "query_graph_handler: Total number of results returned for this query is 344.",
"code": null
},
{
"timestamp": "2021-06-24T16:12:18.959Z",
"level": "DEBUG",
"message": "query_graph_handler: Successfully applied post-query predicate restriction with 42 results.",
"code": null
},Second: it correctly flips the query edge from Gene <- filtered set of Disease IDs to filtered set of Disease IDs -> Gene. It then does that set of queries.
However, it looks like it then tries to restrict the answer edges (Disease -> Gene) with the predicate "gene_associated_with_condition" which none of them have. Instead some have the flipped/inverse predicate condition_associated_with_gene (like Query 1).
See the logs below:
{
"timestamp": "2021-06-24T16:13:41.941Z",
"level": "DEBUG",
"message": "query_graph_handler: Total number of results returned for this query is 73000.",
"code": null
},
{
"timestamp": "2021-06-24T16:13:41.941Z",
"level": "DEBUG",
"message": "query_graph_handler: Successfully applied post-query predicate restriction with 0 results.",
"code": null
}So when BTE correctly flips an edge for querying, some fix is needed so the predicate restriction is then done correctly.
I could think of two ways to fix this:
- the predicate restriction uses the inverse predicate for filtering/restriction
- [EDIT: andrew says we should be able to handle this even with the latest biolink. I think this involves checking, possible changes to the biolink JS package]
- potential problem: the latest version of biolink model only lists inverses on 1 predicate of the pair. In that situation, I'm then not sure if BTE can recognize that gene_associated_with_condition is the inverse of condition_associated_with_gene AND vice versa.
- then after this....BTE needs to flip/invert the predicates on these edges to match the query graph. I can't tell if it's doing it correctly right now
- [EDIT: andrew says we should be able to handle this even with the latest biolink. I think this involves checking, possible changes to the biolink JS package]
- the answer edges all need to flip predicates first, then predicate restriction using the query edge's predicate can be done.
- potential problem: the problem with (2) is that this may involve changed logic for what order things are done in...
Query 2: the "modified" predict that doesn't work
{
"message": {
"query_graph": {
"nodes": {
"n0": {
"ids": ["HGNC:6284"],
"categories":["biolink:Gene"]
},
"n1": {
"categories": ["biolink:Disease"]
},
"n2": {
"categories": ["biolink:Gene"]
}
},
"edges": {
"e0": {
"subject": "n0",
"object": "n1",
"predicates": ["biolink:gene_associated_with_condition"]
},
"e1": {
"subject": "n2",
"object": "n1",
"predicates": ["biolink:gene_associated_with_condition"]
}
}
}
}
}Query 1: "classic" predict that works
{
"message": {
"query_graph": {
"nodes": {
"n0": {
"ids": ["HGNC:6284"],
"categories":["biolink:Gene"]
},
"n1": {
"categories": ["biolink:Disease"]
},
"n2": {
"categories": ["biolink:Gene"]
}
},
"edges": {
"e0": {
"subject": "n0",
"object": "n1",
"predicates": ["biolink:gene_associated_with_condition"]
},
"e1": {
"subject": "n1",
"object": "n2",
"predicates": ["biolink:condition_associated_with_gene"]
}
}
}
}
}
andrewsu
We should check whether BTE is correctly normalizing alternate representations of one query. For example I would expect that
ChemicalSubstance[id:XXXX]-treats-Diseasewould return the same results asDisease-treated_by->ChemicalSubstance[id:XXXX].We should also test longer query paths, ie
A[id:XXX]->B->Cshould be the same asA[id:XXXX]->B<-C.Assuming these BTE is using the expected logic to equate these various queries, we should create formal tests for these cases.