Resolve linkages for Ehlers-Danlos type 1

[cmungall]

Note: this ticket is intended primarily to document the curation procedure for the new kboom-based mondo. The curation methodology is to look at all sub-module solutions in which the graph with the most likely posterior probability involves the selection of edges with a very low prior probability.

This example involves ED-type 1 (this turns up the top of the 'suspect' list). With the current priors the most likely graph is:

• solid black lines are given axioms. In this case Ordo is telling us that type-1 is a subclass of classic type, we assume this is true with Pr=1.0
• probabilistic axioms: priors are indicated by numbers (thickness correlates with probability)
  • red: equivalence axioms (inferred equivalent cliques are enclosed in boxes)
  • blue: subClass
• dotted: mappings supplied by one of the source ontologies

Note that this graph was chosen out of all possible configurations (where we treat each mapping as sub/super/equiv/wrong) as the most likely. However, the most likely involves an edge with a low prior:

• Ordo ED-type-1 SubClassOf Mesh ED-type-1

this has a low prior (Pr=0.028) because lexically we really expect these to be equivalent. Why was this configuration chosen then? In part because other priors are derived from rules such as the fact we don't usually classify things under OMIMs, and possibly some lexical misleadingness too. Overall the graph has a low probability and a low confidence, which in the current report looks like this:

• Pr(G)=0.14305135538487054 CONFIDENCE=0.10093617021276602 Success:true
• SIZE=5 (6 new axioms)
• MEMBER: MESH_C536194 Ehlers-Danlos syndrome type 1
• MEMBER: Orphanet_287 Ehlers-Danlos syndrome, classic type
• MEMBER: Orphanet_90309 Ehlers-Danlos syndrome type 1
• MEMBER: OMIM_130000 Ehlers-Danlos Syndrome, Classic Type
• MEMBER: DOID_14720 type I Ehlers-Danlos syndrome
• OMIM_130000 Ehlers-Danlos Syndrome, Classic Type SubClassOf MESH_C536194 Ehlers-Danlos syndrome type 1 Pr= 0.8518518518518519
• DOID_14720 type I Ehlers-Danlos syndrome EquivalentTo MESH_C536194 Ehlers-Danlos syndrome type 1 Pr= 0.9627906976744186
• Orphanet_90309 Ehlers-Danlos syndrome type 1 SubClassOf MESH_C536194 Ehlers-Danlos syndrome type 1 Pr= 0.028
• OMIM_130000 Ehlers-Danlos Syndrome, Classic Type EquivalentTo Orphanet_287 Ehlers-Danlos syndrome, classic type Pr= 0.8783269961977186
• OMIM_130000 Ehlers-Danlos Syndrome, Classic Type SubClassOf DOID_14720 type I Ehlers-Danlos syndrome Pr= 0.5
• Orphanet_90309 Ehlers-Danlos syndrome type 1 SubClassOf OMIM_130000 Ehlers-Danlos Syndrome, Classic Type Pr= 0.25

If the confidence is high, or if the most likely sub-ontology did not involve selecting an edge with a low prior probability there is no urgency for the curator to vet. However, in this case, the prior is exceedingly low so we can look at this and decide what to do next.

There are a variety of small nudges that will push the most likely configuration into being the correct one:

• One global change would be to weight lexical matches of the "X type Y" form higher (these are already quite high)
• Another would be for DO to add a mapping between the DO type-1 ED and the Ordo one. This would score a high prior due to the lexical rules and would make the above configuration even less likely
• The curator can simple assert all the correct axiom types in the bridge ontology; or for less, work they can do a subset of this:
  • The curator can simply place an "X" next to the low confidence mapping, giving it Pr=0.0 (or Pr=0.000...1).
  • They can also X the SubClassOf between the OMIM classic type and the MESH type 1
  • The curator can give a higher confidence to an axiom that was not selected: e.g. between the Ordo type-1 and any of the others.

I will illustrate this in the next commit.

[cmungall]

Another global fix would be to add an extra level of distrust for RELATED synonyms from MESH and OMIM. We have a rule that lexical substring (weakly) suggests subclassOf. If we look at some of the synonyms we can see this is ripe for false positives:

[Term]
id: OMIM:130000
name: Ehlers-Danlos Syndrome, Classic Type
synonym: "EHLERS-DANLOS SYNDROME, CLASSIC TYPE" RELATED []
synonym: "Eds I, Formerly" RELATED []
synonym: "Eds Ii, Formerly" RELATED []
synonym: "Ehlers Danlos Syndrome, Mild Classic Type, Formerly" RELATED []
synonym: "Ehlers Danlos Syndrome, Mitis Type, Formerly" RELATED []
synonym: "Ehlers-Danlos Syndrome, Gravis Type, Formerly" RELATED []
synonym: "Ehlers-Danlos Syndrome, Severe Classic Type, Formerly" RELATED []
synonym: "Ehlers-Danlos Syndrome, Type I, Formerly" RELATED []
synonym: "Ehlers-Danlos Syndrome, Type Ii, Formerly" RELATED []
xref: Orphanet:287
xref: Orphanet:90309
xref: UMLS:C0268335

[Term]
id: MESH:C536194
name: Ehlers-Danlos syndrome type 1
namespace: CTD_disease_ontology
alt_id: OMIM:130000
synonym: "EDS I, FORMERLY" RELATED []
synonym: "EDS II" RELATED []
synonym: "EDS1, FORMERLY" RELATED []
synonym: "EDS2, FORMERLY" RELATED []
synonym: "EHLERS DANLOS SYNDROME, MILD CLASSIC TYPE" RELATED []
synonym: "EHLERS DANLOS SYNDROME, MITIS TYPE, FORMERLY" RELATED []
synonym: "EHLERS-DANLOS SYNDROME, CLASSIC TYPE" RELATED []
synonym: "EHLERS-DANLOS SYNDROME, GRAVIS TYPE, FORMERLY" RELATED []
synonym: "EHLERS-DANLOS SYNDROME, SEVERE CLASSIC TYPE, FORMERLY" RELATED []
synonym: "EHLERS-DANLOS SYNDROME, TYPE I, FORMERLY" RELATED []
synonym: "EHLERS-DANLOS SYNDROME, TYPE II, FORMERLY" RELATED []
synonym: "Ehlers-Danlos Syndrome, Severe Classic Type" RELATED []
synonym: "Ehlers-Danlos Syndrome, Type I" RELATED []
synonym: "Ehlers-Danlos syndrome, Gravis type" RELATED []
synonym: "Ehlers-Danlos syndrome, classic severe form" RELATED []
synonym: "FORMERLY" RELATED []
is_a: MESH:D004535  ! Ehlers-Danlos Syndrome

Of course, lexical methods will always be flawed, the point of the bayesian method here is to weight them appropriately. In this particular case it seems we have a 'perfect storm' of misleading prior probabilities. In theory this should be fixed with one small tweak of one of the priors...

[cmungall]

The (deliberately) very minor tweak is overriding a single prior, in https://github.com/monarch-initiative/monarch-disease-ontology/commit/1b24d76776783744a9d7bb142314a82b8de7edd9

I will check the results in the a.m

[cmungall]

This resolves the issue:

Note how it still really doesn't like solutions that place diseases underneath an OMIM. We can look at the priors here, but this is perhaps a wider discussion

[cmungall]

And for completeness, here is the full classic-type subhierarchy: