Merged graph sometimes has classes under both benign and malignant neoplasms; consider disjointness

[cmungall]

Example:

This is the source: https://github.com/DiseaseOntology/HumanDiseaseOntology/issues/231

Nevertheless, the kboom algorithm should be resilient to this and find the most likely configuration.

We need to add knowledge in terms of disjointness axioms, this will solve a general class of issues

[cmungall]

Turns out the disjointness axioms solution may not be so straightforward.

In NCIT, Histiocytoma is malignant:

But note it's a subclass of a benign neoplasm

This causes unsatisfiable classes when we convert the APs to logical axioms:

[mellybelly]

histiocytoma can be benign or malignant, need parent class? We should do check for any classes with Benign in parent’s label and Neoplastic_Status of Malignant (and vice versa)

On Jan 21, 2017, at 3:38 PM, Chris Mungall notifications@github.com<mailto:notifications@github.com> wrote:

Turns out the disjointness axioms solution may not be so straightforward.

In NCIT, Histiocytoma is malignant:

[image]https://cloud.githubusercontent.com/assets/50745/22178586/79e2be82-dfef-11e6-8a5e-0f5b06ffb31b.png

But note it's a subclass of a benign neoplasm

This causes unsatisfiable classes when we convert the APs to logical axioms:

[image]https://cloud.githubusercontent.com/assets/50745/22178577/33cfac34-dfef-11e6-8b39-d4d5657ee1fa.png

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHubhttps://github.com/monarch-initiative/monarch-disease-ontology/issues/169#issuecomment-274295721, or mute the threadhttps://github.com/notifications/unsubscribe-auth/AAwn3DQwYt8AHPnNlyezpYY3YbcOuYfIks5rUpcKgaJpZM4LqLwY.

[cmungall]

labels appear to be consistent with APs.

however, the APs seem inconsistent between sub/superclass pairs. See the commit for sparql queries and tsv results (the sparql can be run in jena or protege)

[cmungall]

• https://github.com/monarch-initiative/monarch-disease-ontology/blob/master/src/ncit/inconsistent-benign-isa-malignant.tsv
  • https://github.com/monarch-initiative/monarch-disease-ontology/blob/master/src/ncit/inconsistent-malignant-isa-benign.tsv

note a lot of these are dominated by hemangiomas. The malignancy confusion has been noted before: #122

[pnrobinson]

The benign-malignant dichotomy is not always as simple as it may sound, since it may be difficult to determine malignancy in any given sample, and since tumors go through evolution from benign to malignant and the transition from one state to another may also be difficult to capture. Finally, tumors are heterogeneous, with different subclones of the same tumor having different mutations and ergo different potentials for various cancerous behavior. It would definitely be a mistake to have some benign vs malignant criterion very high up in the ontological hierarchy. Medlically speaking it doesn't seem wrong for a malignant tumor to be a particular type of a benign tumor, since it has developed from it. Perhaps the "is_a" relation needs to be replaced by another relation that does not imply transitivity of "malignant".

[cmungall]

Precisely, the disjointness axiom only applies to subclasses. The temporal case is handled by develops_from relationships, the mixture case by part_of.

[cmungall]

Two that aren't caught by the above SPARQL (as the sparql doesn't check for classes that may be subclasses of both malignant and benign)

[cmungall]

This seems to be the classic case of a prototypical example standing in for a generic class.

"Ossifying Fibromyxoid Tumor" seems to be typically benign, hence it's classification.

Yet it has a single subclass, "Malignant Ossifying Fibromyxoid Tumor". There is no sibling class for Non-malignant.

A standard OWL design pattern would be:

• generic X
  • X with property A
  • X with property B

But when property A happens 99% of the time, it is tempting to assume A for the generic. This may make for a more domain-scientist friendly ontology, but it simply doesn't work in a monotonic logic framework like OWL.

[cmungall]

Confirmation from Larry W:

Neoplastic Status values can be Undetermined, Malignant, or Benign. If a parent is Undetermined, children can be any of the three, but once you hit a concept that is Malignant or Benign, all children of it on down should be the same.

This is consistent with our strategy of translating M or B annotations to logical axioms and implementing disjointness between these two.

However, it an Undetermined annotation should not be translated to a logical axiom, this must remain an annotation.

[cmungall]

Also from Larry:

Histiocytoma should have Neoplastic Status = Benign. I've reported this one so it can be fixed for next month's [NCIT] release

[mellybelly]

When I looked earlier I found examples of both benign and neoplastic histiocytoma in both NCIT and in the literature, for example 'Inflammatory Malignant Fibrous Histiocytoma NCI Grade 2' but these are not related to the histiocytoma term. Note too that there are a lot of seemingly malignant non-human histiocytomas (that we want NCIT to continue to support, though should be the subject of a different ticket to describe taxon constraints). (for future reference excellent review here: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC1978291/)

Should histiocytoma be an undetermined parent with subclasses?

[cmungall]

Note that we now inject disjointness axioms into the local version of ncit we use:

https://github.com/monarch-initiative/monarch-disease-ontology/blob/master/src/ncit/construct-malignant.sparql

and

https://github.com/monarch-initiative/monarch-disease-ontology/blob/master/src/ncit/upper.obo

The two PATO classes are

• PATO:0002097 ! neoplastic, malignant
• PATO:0002096 ! neoplastic, non-malignant ("benign" in NCIT)

This has worked out very well and improved resolutions. One clique could not be solved #170