Tool Implementer Guide#

This guide is for developers of generic LinkML tools that operate over the linkml metamodel, for example:

  • Generic Web-based data submission tools driven by a LinkML schema

  • Tools that produce or modify LinkML schemas

  • Tools that analyze schemas or align schemas or schema elements

  • Tools that auto-genete APIs from schemas

  • Generic faceted data browsers that flexibly operate over multiple schemas (see this draft

  • Schema editors for LinkML itself

Note

This guide is not intended for developers of specific applications that are built for any one particular schema. For example, if I am building an application geared towards entering phenotypic data about patients then I may want to build that application around a schema for patients. In this case, my application isn’t intended to work with other schemas, and hence does not need to know anything about the linkml metamodel.

An example of a generic-schema driven editing application is json-editor https://github.com/json-editor/json-editor, which provides a way to create instance data for an arbitrary JSON-Schema.

[react-jsonschema-form](https://rjsf-team.github.io/react-jsonschema-form/) is based on the same idea.

Another example is DataHarmonizer which provides a spreadsheet-like data entry interface for any “flat” LinkML schema.

General Considerations#

Any generic LinkML application will be driven by a LinkML model (schema), which itself conforms to the LinkML metamodel. It follows that these applications will itself need to conform to that model.

One challenge is that the LinkML metamodel includes many different features, not all of which may be required for a particular application. See the section on Schema Conformance Levels for guidance here, as this will help you target your application to a particular profile.

Another challenge is working with schemas themselves. The native syntax for LinkML is YAML, and YAML parsers are provided with all major languages. However, you will likely need to do more than parse the model. There is a lot of “business logic” associated with a model. For example, the rules that govern inheritance and refinement of slots.

For Python applications, the SchemaView library provides this business logic, but currently this logic must be re-implemented for different languages. This may be a concern if you are building web applications in JavaScript. The section on Working with Other Languages addresses this.

Another challenge with building generic applications is making these configurable at the schema level for different domains. The section on customization and application hints deals with this.

Finally, you may wish to target a different framework than LinkML itself. For semantic web applications, ShEx or SHACL may be a good choice to build applications from. One advantage of LinkML is that you can author in LinkML and compile to these other frameworks. But there are trade-offs here. See the section on Other Frameworks.

Schema Conformance Levels#

LinkML is a rich language supporting extensible types, inheritance, use of Linked Data IRIs, ontological enums, imports, rules, boolean constructs. It can be daunting to consider building a tool that supports all of this. Thankfully this is not necessary.

We are including support for the notion of “conformance levels” or “profiles” to LinkML. This will allow applications to clearly state what parts of the specification are supported, and for people deploying the application to adapt accordingly.

.note:: An example use case is a simple questionnaire UI/application that is

driven by a LinkML schema, where each question is a slot and each answer either free text or in the case of multiple choice, an enum. There is no need for inheritance in this case, so the application only needs to pay attention to elements in the schema such as slots and permissible_values.

Currently specific conformance levels in LinkML have not been defined, but current thinking is to have different categories:

  • A structural conformance level that pertains to logical parts of the schema

    • The minimal level would include classes, attributes, range and basic builtin types, as well as cardinalities

    • The next level would include class inheritance and slots, and basic enums

    • The level above this would include slot inheritance, and slot overrides with slot_usage

    • The highest level would include boolean expressions and rules

  • A metadata conformance level that pertains to informational aspects of the schema that do not drive validation or inference

    • minimal metadata would include metaslots for description, comments

    • above this, a minimal semantic level would include class and slot uris, and mappings

    • the full level would include all possible metaslots

The LinkML framework provides some functionality for being able to do transformations “down” levels. For example, a rich schema that uses inheritance and slot_usage can be “rolled down” to a schema that includes only leaf elements, with inferred slots materialized as attributes. The resulting schema will be less rich, but may be sufficient for certain kinds of applications.

See the YAML Generator for one such tool.

Other Languages#

LinkML is programming language neutral. However, currently much of the stack for introspecting schemas is written in Python.

For example, in Python, if you want a list of slots for that class you can use class_slots. But what if you are building a generic JavaScript data entry widget that will work for any LinkML class?

There are a few broad approaches here:

  • 1. Use the direct output of your language’s YAML/JSON parser, and implement logic ad-hoc

  • 2. Develop a full blown analog of the Python SchemaView class for your language

  • 3. Write Python code to transform LinkML YAML into a native form geared for your application

The first approach may be easiest, but we would only recommend this for simple schemas (low structural conformance – see above).

If you are interested in approach 2, please make an issue on our GitHub repo so we can coordinate! We may have suggestions for general strategies for bootstrapping. And we may already have prototype implementations for your language.

For example, we have a prototype Java code generator that can create Java classes for any LinkML model. The LinkML metamodel is in LinkML, so this can be used to make a Java object model for LinkML itself, which can help bootstrap efforts to make domain logical libraries.

Materializing imports and inferences#

The linkml generator can be used to materialize imports closure and to materialize inferred/induced slots as attributes. This frees the client logic from needing to implement this logic locally.

General Guidelines for applications#

These guidelines apply to how applications should use elements of the LinkML metamodel.

  • Applications should use title to obtain the user-friendly name for a slot. For example a spreadsheet-like data entry tool should display these as column headers

  • If title field is not available, use the name slot

  • The description slot should be used to provide information to users, e.g. as tool-tips

  • The pattern slot should be used to constrain values entered by the user

  • The required slot should be used to indicate to users if a field is not filled in. Applications MAY choose to still allow such data to be saved, e.g. if the user is in an intermediate state

  • The multivalued slot should be used to indicate whether data should be inputted or displayed as a list/set

  • The range slot should be used to constrain values for a slot. The application should also apply relevant logic to this calculation depending on conformance level supported.

  • The URI of a construct may be used to provide links for a user to find more information in an element. For example, in a data entry form a column may have a hyperlink to complete documentation on that data dictionary element

  • An application may use the primary URI of a concept or its mappings to look up semantics for the type and behave appropriately (see examples with geolocation below)

  • Minimally, data entry applications should treat enums as dropdowns or radio-button selectors

  • In a data entry application, a slot that is multivalued and has a range of an enum may be implemented as a multi-select.

  • Enums that take exactly two possible permissible values may be displayed as toggles

  • Applications may choose to use standard ontology browsers such as OLS or BioPortal, or standard query endpoints to obtain more information on enums. See below.

Slots that may potentially be added to provide applications with hints:

  • precedence order

  • grouping categories

Customizing and application hints#

One challenge with generic applications is that they often look and feel… generic. This can have some advantages, e.g. consistent look and feel. But in general UX can be improved by customizing things.

There are three broad approaches:

    1. Make a custom application, with custom domain logic implemented programmatically

    1. Define configuration files

    1. Add schema hints

The first approach is out of scope for this guide – but if you do go down this route, the LinkML framework provides various utilities that may help, such as the ability to generate custom language bindings.

The other two approaches are fairly similar and involve providing a mechanism for a generic application to customize look, feel, and behavior in a way that doesn’t require changing software/code.

External Configuration#

External configuration files may be best for “style sheet” type configurations for controlling colors, shapes, sizes, etc. These could potentially be tweaked by an individual user.

It is easy to roll your own configuration format, but we would recommend creating a schema for your configuration data model. An example of this is KGViz Schema which is a stylesheet language for visualizing ontology graphs, based on Graphviz<https://graphviz.org/>.

Schema Hints#

Schema hints embed additional information in the schema itself. In contrast to external configurations, this is harder for a user to change, and so is best suited for ‘centralized’ configuration.

A simple example might be a slot that takes a string as range. A generic data entry application has no way of knowing how big a text entry box to provide, and whether this should accept single-line or multi-line output. The application could ‘play it safe’ and give the user a large multi-line box, but this would be poor UX if the string field is always a 3-letter code.

LinkML allows for slots and types to be annotated with information that would serve as hints for applications. It is up to you the level of granularity you provide here. However, specifying the precise number of rows and columns may be embedding too much application logic in the schema. Instead we encourage thinking of “semantic types”. For example, you could define two types:

types:

  NameString:
    typeof: string
    pattern: "^[^\\n]$"
    description: A description that holds a human readable name
    comments:
     - This is designed to support different styles of names from
       multiple languages, but certain characters such as newlines are
       never in names

  FormattedString:
    typeof: string
    description: >-
      A string in which characters such as newlines are
      permitted and used for formatting

slots:
  full_name:
    range: NameString
  address:
    range: FormattedString

And then hardcode these types into the application.

A more flexible approach would be instead to use annotations on the types:

types:

  NameString:
    typeof: string
    pattern: "^[^\\n]$"
    description: ...
    annotations:
      dash.singleLine: true

  FormattedString:
    typeof: string
    description: ...
    annotations:
      dash.singleLine: false

This is better as you can reuse the same vocabulary on different types, and you introduce decoupling between specific schemas and your application.

In this case, we are reusing the dash vocabulary which is intended for exactly this kind of purpose. Furthermore, if you compile your schema to SHACL then it will have the dash annotations, allowing you to leverage generic SHACL applications (next section).

Using ontologies and standard vocabularies to drive behavior#

Consider a schema that reuses standard vocabularies such as wgs84 for slots:

prefixes:
  wgs: http://www.w3.org/2003/01/geo/wgs84_pos#
  schema: http://schema.org/

slots:
  latitude:
    domain: geolocation value
    range: decimal degree
    description: >-
      latitude
    slot_uri: wgs:lat
    exact_mappings:
      - schema:latitude

  longitude:
    domain: geolocation value
    range: decimal degree
    description: >-
      longitude
    slot_uri: wgs:long
    exact_mappings:
      - schema:longitude

Applications may choose to have specific behavior for lat-long fields, for example, including a map widget. Applications may also choose to use mappings as well as the primary URI.

Handling enums#

In addition to the general guidance above, applications may allow for custom behavior with enums.

Applications may choose to display enum permissible values as a hierarchy, especially if there are many permissible values. The hierarchy is not provided in the schema itself, but additional APIs or ontology files can be used. The choice of which relationship types to display in the hierarchy may be ontology or application dependent but applications are encouraged to use standard annotations from an ontology like OMO.

For open-ended enums or enums with very many permissible values, applications may choose to use an autocomplete service from an existing ontology. This has the advantage of lookup on multiple different aliases. However, note the autocomplete service may return more values than are present in the permissible value list.

Handling units and quantities#

There are a wide variety of ways to model quantities, and these are use case dependent. Is it important to capture ranges or precision/error bars? Is the unit baked in to the slot, or does the user specify this? Is the quantity captured as a single parseable text string, or is a complex object used?

The modeling decisions will vary based on the use case. However, if certain conventions are followed, then generic applications can be made ‘smart’.

For example, if we model quantity values as classes and reuse the concept from the standard qudt<http://qudt.org/> vocabulary:

quantity value:
  description: >-
    A simple quantity, e.g. 2cm
  attributes:
    verbatim:
      description: >-
        Unnormalized atomic string representation, should in syntax {number} {unit}
    has unit:
      description: >-
        The unit of the quantity
      slot_uri: qudt:unit
    has numeric value:
      description: >-
        The number part of the quantity
      range:
        double
  class_uri: qudt:QuantityValue
  mappings:
    - schema:QuantityValue

Then applications can be aware of the semantics of this field and act accordingly; for example:

  • allow free text entry and use a library like quantulum<https://github.com/marcolagi/quantulum> to parse into structured form

  • allow for conversion between units

  • use sliders to allow input

  • etc

Using other frameworks#

You should also feel free to build applications that use other frameworks. You can compile to these from LinkML, but be aware that you will be restricted to the expressivity of that language–e.g. a project like json-edit can only make use of what is expressible in JSON Schema.

If considering a non-LinkML framework for form-based data entry we would strongly recommend SHACL + DASH. See Form Generation using SHACL and DASH.