OntoGPT

The BC5CDR evaluation uses a specific schema template designed to represent associations between chemicals and diseases.

Key Components:

• ChemicalToDiseaseDocument: A document containing ChemicalToDiseaseRelationship triples.
• ChemicalToDiseaseRelationship: A triple consisting of a subject (Chemical), object (Disease), and predicate (ChemicalToDiseasePredicate, e.g., INDUCES or TREATS).
• Chemical: A NamedEntity with MESH ID prefixes.
• Disease: A NamedEntity with MESH ID prefixes.
• ChemicalToDiseasePredicate: An enum where, for the purposes of BC5CDR evaluation, any predicate other than INDUCES is ignored.

This template is used to evaluate Semantic Llama against the BC5CDR task.

id: http://w3id.org/ontogpt/ctd
name: ctd
title: Chemical to Disease Template
description: >-
  A template for Chemical to Disease associations.
...
classes:
  ChemicalToDiseaseDocument:
    is_a: TextWithTriples
    slot_usage:
      triples:
        range: ChemicalToDiseaseRelationship
  ChemicalToDiseaseRelationship:
    is_a: Triple
    slot_usage:
      subject:
        range: Chemical
      object:
        range: Disease
      predicate:
        range: ChemicalToDiseasePredicate
...

To support complex logic or interoperability with ontology editors like Protege, you can include OWL (Web Ontology Language) annotations within your LinkML schema. This allows you to define how schema components relate to OWL syntax, such as mapping attributes to AnnotationProperty or ObjectProperty.

Key features for OWL integration include:

• slot_uri: Defines the URI for the slot in existing vocabularies (e.g., rdfs:label, dcterms:description).
• annotations: Uses the owl key to specify OWL axiom types (e.g., owl: Class, owl: ObjectProperty).
• owl.template: Uses Jinja-style templating to define complex OWL axioms, such as EquivalentClasses. You can use the tr() function (e.g., {{tr(step)}}) to translate inputs into valid OWL entities.
• close_mappings: Indicates that a class is similar to, but not necessarily identical to, an existing ontology term.

Important: When using a schema with OWL annotations, you must export the results using the -O owl option to preserve this functionality.

classes:
  Recipe:
    tree_root: true
    close_mappings:
      - FOODON:00004081
    attributes:
      label:
        description: the name of the recipe
        slot_uri: rdfs:label
        annotations:
          owl: AnnotationProperty, AnnotationAssertion
      ingredients:
        description: a semicolon separated list of the ingredients
        multivalued: true
        range: Ingredient
        slot_uri: FOODON:00002420
        annotations:
          owl: ObjectProperty, ObjectSomeValuesFrom
    annotations:
      owl: Class
      owl.template: |
        EquivalentClasses(
          {{url}}
          ObjectIntersectionOf(
            recipe:Recipe
            {% for step in steps %}
            ObjectSomeValuesFrom(recipe:steps {{tr(step)}})
            {% endfor %}
          )
        )

If you want to use an OWL ontology as a structural schema (to extract concepts and their relationships rather than grounding to individuals), you can transform it into a LinkML YAML schema using the schema-automator tool.

Workflow

1. Convert to Functional Syntax (if needed): If schemauto fails, use the robot tool to convert your OWL file to .ofn (functional syntax) format.
2. Import with schema-automator: Use schemauto import-owl to generate the LinkML YAML.
3. Refine: The resulting YAML will require manual addition of description and annotation slots to each class to be effective for extraction operations.

Warning: This process may fail if the input ontology relies heavily on external imports.

# Step 1: Convert to functional syntax using robot
robot convert -i fruit.owl -o fruit.ofn

# Step 2: Import using schemauto
schemauto import-owl fruit.ofn

OntoGPT uses the SPIRES mechanism to extract structured information from text based on a provided data model (schema).

Working Mechanism

1. Data Model: You provide a schema (e.g., LinkML) describing the desired output structure.
2. Grounding: You provide preferred annotations for grounding NamedEntity fields.
3. Process: OntoGPT generates a prompt, queries a language model, parses the results into a dictionary, and grounds the results using an annotator (like an ontology).

Extraction Command

Use the extract command with the following flags:

• -t / --template: The base name of a predefined LinkML schema available in OntoGPT, or a path to a custom .yaml schema file.
• -i / --input: Path to the input text file.
• -o / --output: (Optional) Path to redirect the output to a file.

To see all available templates, use ontogpt list-templates.

ontogpt extract -t gocam.GoCamAnnotations -i ~/path/to/abstract.txt