LinkML specification


This document specifies the LinkML language. LinkML is a modeling language that allows the construction of a model (aka schema) which specifies a set of constraints over a object graph.

A LinkML model is typically expressed as a YAML structure, and instance trees are also typically specified as JSON or YAML structures. However, the specification is independent of any particular serialization.

LinkML Models

A LinkML model M consists of the following model elements/definitions:

  • classes MC = {c1,...}, which group objects

  • class references MR = {r1,...}, which group pointers to objects

  • slots MS = {s1, ...}, which describe how objects relate to other objects

  • enums ME = {e1, ...}, enumerated values (value sets)

  • types MT = {t1, ...}, scalar/atomic types, such as integers, strings

  • subsets MP = {p1, ..}, which partition model elements into groupings or views

Note: class references are not declared directly, but are induced (see below)

Example Model: Organization Schema

Classes: Person, Organization, Address, …

Slots: id, name, date_of_birth, employed_at, lives_at, …

Enums: JobCode, …

Types: Date, String, …

To help understand the basic concepts, it can be helpful to think about analogous structures in other frameworks. However, it should be understood these are not equivalents.

  • Classes are analogous to classes in object-oriented languages, tables in relational databases and spreadsheets, and classes in RDFS/OWL

  • Slots are analogous to attributes in object oriented languages, columns or fields in relational databases and spreadsheets, and properties in RDFS/OWL

  • Enums are analogous to enums in programming languages and some relational systems. However, in LinkML enums are optionally backed by stronger semantics with enum elements mapped to vocabularies or ontologies

  • Types are analogous to builtin types in most OO languages and database systems, or extensible types in some systems. They correspond to literals in RDF/OWL

LinkML Objects

A LinkML object o ranges over a Domain of Discourse O, i.e. o O. Each object directly instantiates exactly one of the following:

  • A class in MC

  • A class reference in MR

  • A scalar type (literal) in MT

  • An enum in ME

  • The distinguished value None, indicating a lack of a value

The function instantiates maps each object to exactly one model element, i.e instantiates(o) MC MR MT ME

Note: in LinkML, everything is an object. The string "foo" is an instance of a string type. The LinkML concept of an object maps to the term “element” in the JSON spec. Here we use the term compound object to denote an object that instantiates a class.

Each o O has a list of slot-value assignments A(o), where each assignment is a pair s, V, with the following constraints.

  • s MS, i.e. s is a slot declared in the model

  • V is either a list of objects v1, ..., vn (each of which can be any LinkML object) for any non-negative integer n, or a singular object v, or the atom None.

Formally, A(o) MS x O P(O), where P denotes the set of all possible lists formed from elements of O.

Here we use dot notation to indicate assignments, e.g. o.s indicates the value or values of slot s in object o

Example Object Graph: Organization Schema

Objects: Jane instantiates Person, Acme instantiates Organization, JanesAddress instantiates Address = “Jane Li” (instance of String type)

Jane.employed_at = IdentifierOf(Acme) (instance of Object Reference type)

Jane.lives_at = [JanesAddress] (list of instances of Class type)

“IdentifierOf” – in Python constructed from ClassName[classid] (e.g. If “Person” has the identifier “Foo”, Ref is “PersonFoo”)

An object tree is a collection of objects organized in a tree structure such that each object is a node in the tree, and each assignment forms a child edge, where the edge is labeled with the slot. The tree is a strict hierarchy and has a single root.

A resolved object graph is an object graph in which every (resolvable) instance of a reference in MR is replaced by the corresponding instance of a class in MC. For example, in the example above, the node for Acme and Ref(Acme) would be merged. Unlike the native object graph this need not form a tree structure.

LinkML Metamodel

Every LinkML Model is itself an object, instantiating the SchemaDefinition class in the LinkML metamodel. The object tree rooted in the model is formed of objects that instantiate metamodel classes.

Each of these model objects, including the model itself, can have a set of slot-value assignments, where the slots are drawn from the LinkML metamodel.

The complete LinkML metamodel is spelled out in the metamodel documentation; see SchemaDefinition

We introduce here key metamodel components necessary for understanding structural constraints:

  • Each class c has a collection of slots, c.slots, which are the slots that are applicable to that class.

  • Each slot s has a range indicated s.range, which identifies the set of valid assignments for a slot.

  • Slots may have other assignments for indicating cardinality, value constraints, etc

  • Each enum o has a collection of permissible values o.permissible_values, each corresponding to a controlled term.

Example Model: Organization Schema

Person.slots = [id, name, data_of_birth, employed_at, resides_in, …]

Organization.slots = [id, name, …]

name.range = String

employed_at.range = Organization

resides_in.range = States

states.permissible_values = “AK”, “AL”, “AR”, “AZ”, …

Induced Models

Instance graphs are validated against induced models. An induced model is derived from an asserted model. We describe this process in terms of copy operations but it is up to individual implementations how to handle this - the induced model could be a virtual view over the asserted model.

First the asserted model is copied to create the induced model seed, then the following rules are applied. The rules are applied recursively, in any order. Then post-processing steps are applied.

Rule: Model Import Copying

Each model imports zero to many imports, indicated by the imports metaslot on the model.

When calculating the induced model, all elements from the imported model are copied to the importing model. this is applied recursively, such that if m1 imports m2, and m2 imports m3, all elements of m3 will be copied to m1.

Note: If two or more models import the same target (e.g. m1 imports m2 and m3 and m2 imports m3), m3 will be only be resolved once.

Note: Two models are considered to be “identical” if they both have the same id. If m2 and m3 both have id:, it is assumed that, despite the different location, they represent the same thing. LinkML will check the model version field and will raise an error if m2 has version: 1.0.0 and m3 has version: 1.0.1

Rule: Induced Class-Slots

Each class c can have zero to many slots declared using the slots metaslot.

If a slot s is used in a class c via slots declarations on that class, then an induced slot c!s is created.

When an induced slot c!s is created, then all metaslots from c are copied to c!s.

A class c can declare any number of slot_usages, which will allow assignment of metaslots in c!s. All metaslots from the slot_usage object is copied to the induced slot.

Rule: Inheritance

Each class has zero or one is_a parent, and zero to many mixin parents.

Parents(c) = { c.is_a } c.mixins

All metaslots of the parent are copied to the child, if the metaslot is declared inheritable.

This is applied recursively, such that a class inherits all slots of its ancestors

For example, if the model contains class Person such that Person.slots = [name, ], and Employee isa Person, then Employee will have also that slot.

Asserted Model

Induced Model

Person.slots = [name, ...]
Employee.slots = [employed_by]

Employee.slots = [employed_by, name, ...]

the same applies to slots. heritable metaslots are copied from parent slots.

semantics of copying from parents:

If a metaslot s is declared multivalued then when copying s from a parent to a child, the values are appended.

If a metaslot s is declared multivalued

if a slot is multi valued then copying will append, unless the element already exists.

if the slot is single valued, and intersection rules can be applied to the slot, then these are performed on all values

if the slot is single valued, and intersection rules cannot be applied to the slot, then the following precedence rules are applied:

  • metaslot values from slot_usage take the highest priority

  • metaslot values from the slot definition take the next highest priority

  • direct mixins take the next highest priority. where multiple direct mixins are provided as a list, the last element takes highest priority

  • direct is_as take the next highest priority

  • the above two rules are applied one level up, and then recursively applied

Intersection rules










IF subsumes(v1,v2) then v2
ELSE IF subsumes(v2,v1) then v2 ELSE UNDEFINED

If the result of applying any intersection rule is UNDEFINED then we fall back on precedence rules

Structural conformance of object trees

Any object o can be evaluated for structural conformance to an induced model M

  • For each o O, where o directly instantiates C:

    • For each s,V A(o):

      • Let c!s be the class slot of s for class C

      • Perform structural checks on c!s for V

Multivalued check

  • if c!s.multivalued is True, then V must be a list

  • If c!s.multivalued is False, then V must be a single value

Minimum and Maximum Values checks

  • if c!s.maximum_value is assigned, then V must be a singular integer and must be less that or equal to the maximum value

  • if c!s.minimum_value is assigned, then V must be a singular integer and must be greater that or equal to the minimum value

Range checks

Depending on the range of c!s a particular check is applied on each c V (if V is multivalued) or v=V



c!s.range MC

class instantiation check

c!s.range MT

type check

c!s.range ME

enum check

Range class instantiation check

if s.range MC: v must be either:

  • an object that instantiates a class c DESC*(s.range)

  • a reference to an object that instantiates a class c DESC*(s.range)

Here DESC* is the reflexive closure of the child function, where child(p) = { c : c.is_a = p or p   c.mixins }

Additional checks MAY be performed based on whether s.inlined is True

  • if s.inlined, then v SHOULD NOT be a Reference

  • if s.inlined is False, then EITHER:

    • v SHOULD be a Reference

    • OR v instantiates a class R such that R has no slot rs that is declared to be an identifier. i.e. rs.identifier = True

Range type check

if s.range MT: v must be a literal of type MT

Range enum check

if s.range MO: v must be a member of the permissible value in MO

TODO: other enum types

abstract classes and slots

a class instance SHOULD not instantiate a class that is declared abstract.

an assignment s,V SHOULD not have s declared abstract.

Required value checks

  • For each o O, where o directly instantiates C:

    • For each s C.slots:

      • If s.required is True, then there MUST exist a value o.s

      • If s.recommended is True, then there SHOULD exist a value o.s

Constraints on instantiation of classes

  • For each o O, where o directly instantiates C:

    • c.abstract MUST be false

    • c.mixin SHOULD be false

Note in future versions we may introduce model level metaslots that allow these to be relaxred

constraints on models

stratification of isa and mixin hierarchies

is_a and mixin hierarchies SHOULD be stratified:

  • if c.mixin = True then all parents p of c SHOULD have p.mixin = True

  • if c.mixin = False and c.is_a is provided, then c.is_a.mixin SHOULD be set to True

  • for all p in c.mixins, p.mixin SHOULD be set to True

Glossary of terms (Information)

  • assignment

  • class

  • class slot

  • element (of schema):

  • enum

  • inheritance

  • is_a

  • model (aka schema)

  • mixin

  • object (aka instance):

  • range

  • required

  • slot

  • type (aka scalar value, literal, atom):