FAQ: Modeling#

How do I get started defining a data model?#

The tutorial walks you through some basic data models.

After that the section on schemas guides you through some of the core features of the modeling framework.

What is the difference between is_a and mixins?#

LinkML allows any class to have:

zero or one is_a parents, declared using the is_a slot
any number of mixin parents, declared using the mixins slot

Semantically these are the same - all inheritable slots are inherited through is_a and mixins.

Classes should have a single inheritance backbone, with is_a representing the “main” parent. Generally the mixin and is_a hierarchies should be stratified.

See these wikipedia pages for more information.

Didn’t you know composition is favored over inheritance these days?#

For background, see the Wikipedia article on composition over inheritance

We have certainly seen cases where inheritance is abused in programming languages, especially when it comes to behavioral classes.

However, in our experience inheritance is still very useful when used for data classes. We trust the users of LinkML to create and design schemas carefully. But you can avoid it altogether if you like, and use composition entirely!

When should I use attributes vs slots?#

The attributes metamodel slot is really just a convenient shorthand for being able to declare slots “inline”.

LinkML treats slots as first class entities. They are defined in their own section of a schema, and can be reused by any number of classes (and refined, using slot_usage). This can be very powerful for reuse.

However, this can also be slightly inconvenient for simple schemas, especially those where we have classes with slots that are completely “owned” by that class. The attribute slot can be used to avoid having to specify the slot separately

What are induced slots?#

Because the same slot can be reused in different classes (with each class potentially refining semantics using slot_usage), it can be useful to give a new “name” for the implicit class-specific version of that slot.

For example, if you have a slot name, and this is used in classes Person and Organization, and these are refined for each class (for example, “Organization” may refine the name slot to follow a particular regular expression). In some generators such as the markdown generator, you will see “induced” slots such as Organization_name.

The extent to which these are made visible is currently the subject of some discussion, see GitHub for details.

Induced slots can be materialized as attributes using the linkml generator

See:

derived models

Why would I need to define my own types?#

Types are scalar values such as integers or strings. They are also known as “literals” in RDF.

Strictly speaking it is not necessary to define your own types, you can just use the builtin types (string, integer, etc).

However, defining your own types can be good practice, as it can make your intent clearer. For example, if you have a slot description you may want to specify the range as your own type NarrativeText that maps to string behind the scenes. But this provides additional cues, e.g. that the value of this field is intended to be human-readable text.

An example of a type section might be:

types:
  CountType:
    uri: xsd:integer
    base: int
    minimum_value: 0
    description: An integer that specifies cardinality
  SymbolType:
    uri: xsd:string
    base: str
    pattern: "^\\w+$"
    description: A symbol is a string used as a shorthand identifier that is restricted to a subset of characters

Some applications may choose to interpret this in particular ways. E.g. you may want to define all narrative text fields as being amenable to spellchecking, or machine learning natural language processing, or special kinds of indexing in ElasticSearch/Solr

Why would I want to use enums over strings?#

Enums provide a more controlled vocabulary than strings. You can validate categorical fields using enums, whereas with basic strings you don’t have a built in way of validating if a string is valid.

Enums also give you hooks into ontologies and vocabularies.

How do I constrain the value of a slot using an ontology or vocabulary?#

There are a variety of ways of tackling this in LinkML.

The fundamental question is whether you want to either:

define a fixed set of terms in the schema in advance
specify the set of terms via a query (e.g. a particular ontology branch)

See the two questions below for answers to each

How do I constrain a slot to a fixed set of ontology terms?#

You can do this using an enum.

For example, if we wanted to model a DNA sequence variant type as being a fixed set of terms from SO:

prefixes:
  SO: http://purl.obolibrary.org/obo/SO_

classes:
  variant:
    slots:
       - ...
       - variant type
        
slots:
  variant type:
    range: variant_type_enum
  ...:

enums:
  variant_type_enum:
    permissible_values: 
      point_mutation:
          meaning: SO:1000008
      deletion:
          meaning: SO:0000159
      insertion: 
          meaning: SO:0000667

Note that we are mapping each permissible value to an ontology term. Mapping to ontology/vocabulary terms is optional, but if you can do it, we strongly recommend it. It provides interoperation hooks - others with different data models may have their own enumerations, by making the meaning of each permissible value explicit, data can be merged automatically.

How do I constrain a slot to a branch of an ontology or a whole ontology?#

LinkML basic enums allow you to restrict a value to a fixed set of terms. This works well, if

the vocabulary is known in advance
it is a relatively small number of terms

However, this does not work so well if you want to constrain something to a very large vocabulary - for example, any job code from an occupation ontology, any body part from an anatomical ontology.

In this case, you can use dynamic enums:

slots:
  cell_type:
    range: CellTypeEnum

enums:
  CellTypeEnum:
    reachable_from:
      source_ontology: obo:cl
      source_nodes:
        - CL:0000000
      include_self: false
      relationship_types:
        - rdfs:subClassOf

This restricts to any subclass (transitive, non-self) of the term cell in the cell ontology.

An alternative pattern is to use a regular expression:

types:
  CellTypeId:
    typeof: uriorcurie
    pattern: '^CL:\d+$'

slots:
  cell_type:
    range: CellTypeId

However, this has a number of limitations

You can also model ontology terms directly, e.g

default_prefix: my_schema

classes:
  variant:
    slots:
       - variant type
  ontology term:
     slots:
        - name
        - id
        - ontology namespace
        - synonyms
        - secondary ids


slots:
  variant type:
    range: ontology term
  name:
  id:
     type: uriorcurie
  ontology namespace:
  synonyms:
  secondary ids:

declare id_prefixes for a class that constrain the kinds of identifiers used to describe the class

default_prefix: my_schema

classes:
  variant:
    slots:
       - variant type
  sequence ontology term:
     slots:
        - name
        - id
        - ontology namespace
        - synonyms
        - secondary ids
    id_prefixes:
       - SO


slots:
  variant type:
    range: sequence ontology term
  name:
  id:
    identifier: true
    type: uriorcurie
  ontology namespace:
  synonyms:
  secondary ids:

See dynamic enums for more details.

Can I use LinkML to develop ontologies?#

LinkML is intended as a schema modeling framework, rather than an ontology modeling framework. Schemas are intended for modeling and constraining the structure of data, whereas ontologies and ontology modeling frameworks like OWL are for modeling and constraining models of the world.

LinkML is intended to be combined with OWLs and other controlled vocabularies, using terms from these resources as mappings in enumerations.

However, the distinction here is frequently blurred, and there are many examples of schemas that have been modeled using OWL - e.g. BioPAX, FOAF.

LinkML allows any schema to be translated to OWL using the gen-owl <https://linkml.io/linkml/generators/owl>_ generator. There are a number of reasons to do this:

take advantage of ontology exploration and browsing tools such as BioPortal and Protege
use OWL reasoning over schemas and data (with the caveat that OWL uses Open World reasoning)

You can also use Schema Automator <https://linkml.io/schema-automator>_ to do the reverse translation.

It doesn’t make sense to develop a large terminological-style ontology such as an OBO ontology as LinkML classes, since LinkML is intended for data modeling.

One option is to use the linkml-owl <https://linkml.io/linkml-owl>_ framework to generate OWL classes from LinkML data

Are CURIEs used in schema definitions checked for expandability and resolution?#

No, not at this time. However, linkml_runtime does have methods to help you expand the CURIEs in your data using the prefixes in your model (see: linkml_runtime.utils.namespaces.py) into URIs. In addition, the curies python package which provides a standalone CURIE expansion service. There are many ways to check if a URI is resolvable. One open source python package to do this is: LinkChecker.

Are CURIEs used in data that validates against a given LinkML schema checked for expandability and resolution?#

No, not at this time. However, linkml_runtime does have methods to help you expand the CURIEs in your data using the prefixes in your model (see: linkml_runtime.utils.namespaces.py) into URIs. Specifying a regular expression to constrain the CURIEs in your data to a particular pattern is also possible.
See the regular expression metaslot. However, validating a CURIE that matches the regular expression, but is invalid in some other way (e.g. is an obsolete ontology term) is not currently supported.

Is it possible for us to import only a subset of an existing LinkML model?#

Not yet, but we are working on a tool to this, please check out linkml-transformer for more details.

Can I combine dynamic enums using boolean expressions#

Yes, this is possible.

See enum documentation

Can I use regular expressions to constrain values?#

Yes, regular expressions can be used either on slots, or on types.

This is done using the pattern metaslot

Can I reuse regular expression patterns?#

Yes, you can do this via the structured_pattern metaslot

First you declare the patterns to be reused in the top level of your schema:

settings:
  float: "\\d+[\\.\\d+]"
  unit: "\\S+"
  email: "\\S+@\\S+{\\.\\w}+"

You can then use this inside a structured pattern:

  height:
    range: string
    structured_pattern:
      syntax: "{float} {unit.length}"
      interpolated: true
      partial_match: false

How do I do the equivalent of JSON-Schema composition?#

See: Schema Composition in JSON-Schema docs

LinkML provides the following analogous concepts:

In some cases, the use of schema composition can be avoided by using simple inheritance patterns.

Note that these constructs may not be supported by all generators. See advanced features for current documentation.

Why are my class names translated to CamelCase?#

LinkML allows you to use any convention you like when authoring schemas. However, when translating to other formalisms such as JSON-Schema, RDF, Python then those naming conventions are applied.

For example, if you define a class:

default_prefix: my_schema

classes:
  my class:
    attributes:
      my slot:

Then this will translate as follows:

Python, JSON-Schema
- MyClass
- my_slot
RDF/OWL URIs
- my_schema:MyClass
- my_schema:my_slot

Note in the RDF/OWL representation, separate rdfs:label triples will be generated retaining the original human-friendly name.

This has the advantage of keeping human-friendly nomenclature in the appropriate places without specifying redundant computer names and human names

However, the automatic translation can be confusing, so some schemas opt to follow standard naming conventions in the schema:

default_prefix: my_schema

classes:
  MyClass:
    attributes:
      my_slot:

you have the option of specifying human-friendly titles for each element:

default_prefix: my_schema

classes:
  MyClass:
    title: my class
    attributes:
      my_slot:
        title: my slot

Note that one current limitation of the LinkML generator framework is that it does not protect you from using keywords that are reserved in certain formalisms.

For example, if you define a slot in, then this conflicts with the Python keyword, and the generated python code will raise errors. For now the recommendation is to avoid these as they arise.

In future, the LinkML framework will

warn if a reserved term is used
provide a mechanism for transparent mapping between a schema element and a “safe” version of the element

When two data classes are linked by a slot in one class definition, how is the reciprocal association expressed in LinkML?#

Relationships between classes can be defined in a few ways:

via slots that dictate the link via domain and range constraints.
via objects that capture the two objects and the relationship between those concepts.

via slots that dictate the link via domain and range constraints

default_prefix: my_schema

classes:
  allele:
    slots:
       - allele of
  gene:
     
slots:
  allele of: 
     type: uriorcurie
     domain: allele
     range: gene

via objects that capture the two objects and the relationship between those concepts

default_prefix: my_schema

classes:
  allele:
  gene:
  allele gene relation:
     slots:
        - subject
        - object
        - predicate
      
slots:
  predicate: 
     range: predicate_enum
  subject:
     range: allele
  object:
     range: gene

enums:
  predicate_enum:
    permissible_values:
      allele_of:

You can further annotate your schema with information that two of your classes represent entities and one represents a relationship:

default_prefix: my_schema

classes:
  allele:
  gene:
  allele gene relation:
     represents_relationship: true
     slots:
        - subject
        - object
        - predicate
      
slots:
  predicate: 
     range: predicate_enum
     relational_role: PREDICATE
  subject:
     range: allele
     relational_role: SUBJECT
  object:
     range: gene
     relational_role: OBJECT

enums:
  predicate_enum:
    permissible_values:
      allele_of:

Applications can make use of this metadata - e.g for compact property graph representations, ER-style visualizations of the schema, auto-inferring convenient shortcut slots.

How do I avoid name clashes when importing a schema?#

Currently the assumption of existing LinkML tools is that all element names are unique, both within an individual schema and across imports.

This means if you want to import a schema personinfo, and personinfo includes a class Person, or another imported schema or your own schema has a class Person, there will be an element clash, and you will need to either remove the import, change the imported schema, or change your own schema.

Historically this has not been a major issue, as imports are typically used sparingly, and the assumption is that the imported schema is orthogonal to the importing one. In many cases the apparent issue is resolved simply by not using an import and instead reusing class_uris.

However, we recognize the unique element restriction can be limiting, and we are currently exploring mechanisms that provide more flexibility in reuse, including:

the use of structured_imports to selectively import elements from a schema
alternatives to imports and inheritance, such as using implements
using linkml-transformer to transform upstream schemas rather than import them

What is the prefixes section at the start of a schema?#

The prefixes section can be used to provide CURIE abbreviations for entities. Under the hood, all elements in a LinkML schema are identified by a URI, but we typically expose these as CURIEs.

For example linkml:SchemaDefinition is a CURIE that expands to https://w3id.org/linkml/SchemaDefinition.

The prefixes section allows you to define a set of prefixes that can be used throughout the schema, for example:

prefixes:
  linkml: https://w3id.org/linkml/
  biolink: https://w3id.org/biolink/
  schema: http://schema.org/

Is there a standard registry of prefixes?#

LinkML is closely aligned with the bioregistry, a community-driven, curated, hierarchical collection of prefix namespaces for use in data resources. Bioregistry is used commonly in the life sciences, but it is not restricted to this domain.

We recommend using prefixes that align with bioregistry.

See also Unifying the identification of biomedical entities with the Bioregistry (2022), doi:10.1038/s41597-022-01807-3

What are id_prefixes used for?#

The LinkML meta modeling element, id_prefixes can be applied to any Class. This is used to specify which prefixes should be used on the identifiers for that class.

The id_prefixes are listed in decreasing priority order, with the “preferred” prefix listed first.

Downstream software components can use this field to constrain data entry to a particular kind of identifier.

To see examples, Biolink uses id_prefixes extensively. For example, the MolecularEntity class shows that identifiers for this class can be drawn from PubChem, CHEBI, DrugBank, etc.

For more, see URIs and Mappings

When is it important to have mappings?#

Any element in a LinkML schema can have any number of mappings associated with it

Mappings are useful in a variety of ways including:

they make your data and your schema more FAIR (Findable, Accessible, Reusable, and Interoperable)
when people use data that conforms to your model, and integrated with data that conforms to another model, they can use mappings between models to help automate data harmonization.
mappings can provide links to other documentation sources for your model, allowing expertise to be shared between projects and not duplicated
mappings allow advanced users to reason over your model.

Mappings can be established for exact equivalences, close, related, narrow and broad equivalences For more detail on the kinds of mappings (and their mappings to SKOS), see linkml:mappings

Mappings are an entire optional feature, you can create a schema without any mappings. However, we encourage their use, and we encourage adding them prospectively as you build your data model, rather than doing this retrospectively. Thinking about mappings will help you think about how your modeling relates to the modeling done by others as part of other databases or standards.

How do I represent relationships in LinkML?#

For some use cases, objects described using LinkML can stand in isolation, and do not need to be related. For example, for a simple database of material samples (biosamples, geosamples, etc), each sample may be considered a standalone entity described with an identifier and various properties.

However, more often then not, your objects need to be inter-related. Here there are a number of modeling questions that you will need to answer:

can my objects be related in different ways, or is there a uniform relationship type
do I need to store information/metadata about the relationship itself?
are the related objects somehow “external”, or are they within my dataset – and if so, should this be enforced

Depending on the answer, LinkML has different modeling constructs to help you, including:

range constraints, which can refer to classes
the ability to assign a slot as an identifier, allowing other objects to link to it
inlining, which determines how relationships are serialized in formats like JSON

Other more advanced constructs are also possible to allow you to treat relationships as first-class entities.