Difference between revisions of "Novel reasoning strategies"

From phenoscape
(The default property)
(The default property)
Line 48: Line 48:
  
 
<math>\forall</math>c <math>\exists</math>n : ''instance_of''(c, Cell) <math>\and</math> ''instance_of(n, Nucleus) <math>\and</math> ''contains''(c. n) -- (1)
 
<math>\forall</math>c <math>\exists</math>n : ''instance_of''(c, Cell) <math>\and</math> ''instance_of(n, Nucleus) <math>\and</math> ''contains''(c. n) -- (1)
 +
 
<math>\forall</math>rbc<math>\not</math> <math>\exists</math>n : ''instance-of''(rbc, Cell)<math>\and</math> ''instance_of(n, Nucleus) <math>\and</math> ''contains''(c. n)--(2)
 
<math>\forall</math>rbc<math>\not</math> <math>\exists</math>n : ''instance-of''(rbc, Cell)<math>\and</math> ''instance_of(n, Nucleus) <math>\and</math> ''contains''(c. n)--(2)

Revision as of 19:12, 30 January 2009

Potential strategies to implement novel inference procedures introduced in the OBD Reasoner page. These comprise: the Balhoff rule and the Mabee rule for inferring the absence of features in taxa.

Prelude: Instance relations and class relations

It is useful to conceptualize two entities in a domain of knowledge such as biology viz. classes and instances. Every entity in the domain is an instance. Instances can be grouped together on the basis of the properties they exhibit, such as the things that are colorless. Classes are groupings or sets of instances on the basis of common properties that they exhibit. An example of a class would be car, which includes all the entities (instances) which comprise a chassis resting (typically) on 4 wheels which is powered by an internal combustion engine, such as my Porsche 911 Turbo or the car you drive.

Subsumption semantics (as inherent in the is_a relation) hold between sets of real-world instances or classes. For example, the set of all left-handed people above the age 35 is subsumed by the set of all the people above the age 35 (who may be left- or right-handed or ambidextrous). The subsumption semantics of the is_a relation are defined in the OBO relations ontology as follows: "...C is_a C' if and only if: given any c that instantiates C at a time t, c instantiates C' at t..." In plain English, this means: the set of all left-handed people above the age 35 is subsumed by (or included by) the set of people above 35 if and only if, a person who belongs to the group of left handed people above the age of 35 also belongs to the set of people above the age of 35. This is trivially valid in the real world.

The Balhoff rule

The Balhoff rule can be summarized as follows. "IF SOME instances of a species A exhibits a phenotype P, AND ALL instances of species A are also instances of species B, THEN SOME instances of species B exhibit phenotype P." Using FOL constructs, this can be represented as shown below.

Rule: <math>\forall</math>A, B<math>\exists </math>X: is_a(A, B) <math>\and </math>R(A, X) <math>\Rightarrow</math> R(B, X)

In our use-case, A and B are terms from a species taxonomy, and R is the exhibits relation. Note that this rule is applicable to all relations that R represents. In the real world, if we know that some combustion engines have four cylinders (while some other have six), we can correctly deduce that some engines (which include both 4 cylinder and 6 cylinder engines) have four cylinders. The only constraint on this rule is that some kind of subsumption relationship hold between A and B. The is_a name of the relation can be conceivably replaced by kind-of for example; as long as the semantics of the relation implied subsumption or inclusion, the Balhoff rule would hold.

The inclusion of the Balhoff rule would lead to the extraction of inferences such as "SOME Ostariophysians exhibit increased thickness of hyoid bones and SOME Ostariophysians exhibit decreased thickness of hyoid bones." All these inferences are extracted from phenotype assertions about different species like Gonorhynchiformes and Cypriniformes that are subsumed by Ostariophysians.

Implementation Strategy

The OBD reasoner, upon invocation, creates inferences by manipulating existing records in the Phenoscape database and adding new inferences to the database itself. The implementation the Balhoff rule can be done trivially by adding a new rule to the OBD reasoner, which will add new inferred records to the Phenoscape database with a SQL INSERT command.

The absence of features problem

Canonical ontologies define properties of typical or ideal classes in the domain of knowledge. In the domain of fishes, it is typical for "wild" fish specimens belonging to the Ostariophysian clade to exhibit basihyal cartilage. However, fishes belonging to the Siluriformes clade, which is subsumed by the Ostariphysian clade are characterized by the absence of basihyal cartilage. The Mabee rule is a proposal to resolve the conflicting assertions formally stated as "ALL instances of a clade C1 typically exhibit a feature F1 AND ALL instances of a clade C2 which is a subclass of C1 do not exhibit F1." This creates what is termed a logical inconsistency. These are shown in (1) and (2) below.

<math>\forall</math>A, B, F: is_a(A, B) <math>\and </math>exhibits(A, F) <math>\Rightarrow</math> exhibits(B, F) -- (1)

NOT exhibits(B, F) -- (2)

The Mabee rule involves inferences propagated to the lower levels of a taxonomy from assertions (or definitions) in the higher levels. Given that ALL instances of fishes belonging to the Ostariophysian clade are defined to possess basihyal cartilage, it can be inferred that instances of fishes belonging to the Siluriformes clade (which is subsumed by the Ostariophysian clade) would possess basihyal cartilage as well. An assertion to the contrary creates the contradiction or inconsistency.

  • NOTE
    • The NOT operator in equation (2) is the negation operator from first order logic. In (2), it is used to state the entity B does not exhibit the feature F. The semantics of the negation operator and its pertinence to the Phenoscape project will be discussed in another section. For now, it is enough to understand that the inference derived in (1) (exhibits(B, F)) and the assertion in (2) directly contradict one another.

Implementation Strategies

A few potential strategies to address this situation are presented here.

The move to existential quantification

If instead of constraining every instance of A to exhibit the feature F, the constraint was loosened to only some instances of A, the contradiction scenario can be avoided. If only some instances of A exhibited F and all A are B, then it cannot be inferred that all B exhibit F. When the assertion in (2) is added, a contradiction is averted.

This strategy will require careful curation of the top level classes in the taxonomy. Definitions of higher level classes and their properties will have to be constructed carefully to specify existential (SOME) and universal (ALL) constraints, so that inconsistencies are not entailed.

The default property

This strategy is based on the proposals by Hoehndorf et al and Ceusters et al. Property values as specified by definitions in canonical ontologies can be transformed into default values. For example, a canonical ontology would define all animal cells to contain at least one nucleus in them. However, red blood corpuscles lack nuclei, the exception to this rule. This situation is represented in FOL as shown below

<math>\forall</math>c <math>\exists</math>n : instance_of(c, Cell) <math>\and</math> instance_of(n, Nucleus) <math>\and</math> contains(c. n) -- (1)

<math>\forall</math>rbc<math>\not</math> <math>\exists</math>n : instance-of(rbc, Cell)<math>\and</math> instance_of(n, Nucleus) <math>\and</math> contains(c. n)--(2)