SchemaLink IE Engine

Schema-guided information extraction from biomedical text using Large Language Models.

The SchemaLink IE engine takes a schema developed with the SchemaLink webapp (SchemaLink.biodata.di.unimi.it) and a text as input to extract structured entities and relations — grounded to biomedical ontologies — that conform to the schema.

---

Overview

The SchemaLink IE engine implements a dependency-aware multi-step pipeline that:

1. Parses the schema and builds a class-dependency DAG
2. Runs topologically-sorted GPT calls — each step conditioned on the results of its dependencies
3. Applies algorithmic post-processing rules (deduplication, cardinality, inheritance resolution)
4. Optionally grounds extracted entities to biomedical ontologies via OAK

   Schema     ──┐
                ├──▶  SchemaLink Engine  ──▶  Entities and Triples
Biomedical Text ┘         (LLM + DAG)           (JSON / LinkML)

---

Installation

pip install schemalink-engine

Or from source:

git clone https://github.com/BioDataUniMI/schemalink-engine.git
cd schemalink-engine
pip install -e .

Set your OpenAI API key:

schemalink api-key set sk-your-key-here

Note: If the OPENAI_API_KEY environment variable is set in your shell, it will always take priority over the key saved with api-key set. Run schemalink api-key check to see which source is being used. To use the key manager, unset the environment variable first: unset OPENAI_API_KEY

---

Quick Start

Given a schema file (schema.yaml, built with SchemaLink) and a text file (text.txt):

# Standard extraction — dependency-aware by default
schemalink extract schema.yaml text.txt

# With ontology grounding (recommended when the schema defines annotators)
schemalink extract schema.yaml text.txt --ground

# Flat extraction — all classes extracted independently, ignoring dependencies
schemalink extract schema.yaml text.txt --flat

Example Schema (Drug–Disease)

id: https://example.org/ddi
name: drug_disease_schema
prefixes:
  MONDO: https://purl.obolibrary.org/obo/mondo/mondo-international.owl
  CHEBI: http://purl.obolibrary.org/obo/chebi.owl
imports:
  - ontogpt:core
  - linkml:types
classes:
  DrugTreatsDiseaseRelationship:
    is_a: Triple
    slot_usage:
      subject:
        range: Drug
      object:
        range: Disease
      predicate:
        range: DrugTreatsDiseasePredicate
  DrugTreatsDiseasePredicate:
    is_a: RelationshipType
    attributes:
      id:
        pattern: 'Treats'
  Disease:
    is_a: NamedEntity
    id_prefixes: [MONDO]
    annotations:
      annotators: sqlite:obo:mondo
  Drug:
    is_a: NamedEntity
    id_prefixes: [CHEBI]
    annotations:
      annotators: sqlite:obo:chebi

Example Output

{
  "Disease": {
    "mentions": [
      { "label": "Parkinson's disease", "id": "MONDO:0005180" }
    ]
  },
  "Drug": {
    "mentions": [
      { "label": "levodopa", "id": "CHEBI:15765" },
      { "label": "carbidopa", "id": "CHEBI:3395" }
    ]
  },
  "DrugTreatsDiseaseRelationship": {
    "mentions": [
      { "subject": "CHEBI:15765", "predicate": "Treats", "object": "MONDO:0005180" }
    ]
  }
}

---

Python API

from schemalink_engine.pipeline import run_extraction_pipeline

# Dependency-aware extraction (default)
run_extraction_pipeline(
    schema_path="schema.yaml",
    text_path="text.txt",
    with_dependencies=True,   # True by default
    show_results=True,
)

# With ontology grounding
run_extraction_pipeline(
    schema_path="schema.yaml",
    text_path="text.txt",
    ground_entities={"mode": "auto"},
    show_results=True,
)

---

Grounding

By default, extraction returns raw LLM labels with no ontology IDs. Pass --ground to enable grounding:

schemalink extract schema.yaml text.txt --ground

The grounding method is chosen automatically based on the annotators: field defined in the schema class:

Annotator in schema	Grounding method
sqlite:obo:mondo, sqlite:obo:chebi, etc.	OAK exact match against a local ontology SQLite database
cellosaurus, ncbigene, mesh_d, etc.	Lookup against a local reference table

OAK databases are downloaded automatically on first use from the bbop-sqlite S3 bucket and cached in ~/.data/oaklib/. If the OAK library is unavailable, the engine falls back to querying the .db files directly via SQLite. If no local database or lookup table is found, the entity is left ungrounded.

Classes with no annotators: field are always left ungrounded regardless of the --ground flag.

---

Web Interface

The SchemaLink IE engine ships with a Flask production server that exposes a REST API and a streaming SSE endpoint:

python production_server.py
# → http://localhost:15002/engine/api/v2/extract  (streaming)
# → http://localhost:15002/engine/api/v1/extract  (sync)

The full web application is available at SchemaLink.

---

Supported Ontologies

SchemaLink supports any ontology available as an OAK SQLite database, including:

chebi · go · mondo · hp · hgnc · pr · mesh · pw · doid · ncit · uberon · cl · ro · and 200+ more via the OBO Foundry

---

CLI Reference

schemalink extract <schema> <text> [options]

Options:
  --ground              Enable ontology grounding (method chosen automatically
                        from the annotators defined in the schema)
  --flat                Disable dependency-aware extraction — extract all
                        classes independently (default: dependency-aware ON)
  --add_guidelines      Include schema-level guidelines in prompts
  --classes A B C       Extract only specific classes
  --model <name>        Override the GPT model for this run
  --show_prompts        Print the prompts sent to the LLM (API call still made)
  --show_results        Print the extraction output to stdout
  --json_schema         Print the parsed JSON schema and exit

schemalink api-key set <key>   Save your OpenAI API key
schemalink api-key check       Check whether an API key is configured
schemalink api-key remove      Remove the saved API key
schemalink model set <name>    Set the default GPT model
schemalink models              List all supported GPT models

---

Architecture

schemalink_engine/
├── cli.py                  # CLI entry point
├── pipeline.py             # Main extraction pipeline
├── schema_convertor.py     # LinkML YAML → JSON schema parser
├── api_key_manager.py      # OpenAI key & model management
└── utils/
    ├── generate_dependencies.py          # DAG construction from schema
    ├── dag_generator.py                  # DAG visualization
    ├── extract_named_entity_classes.py   # NER class identification
    ├── process_named_entities.py         # GPT-based NER
    ├── handle_inherited_classes.py       # Inheritance resolution
    ├── process_inherited_entities.py     # Inherited class extraction
    ├── handle_relationship_classes.py    # Relation class identification
    ├── process_relationship_entities.py  # GPT-based RE
    ├── grounding.py                      # OAK-based entity grounding
    └── some_helper.py                    # Topological sort, utilities

---

Requirements

• Python ≥ 3.8
• OpenAI API key
• See requirements.txt for full dependency list

---

License

MIT License — see LICENSE for details.

---

Citation

If you use SchemaLink in your research, please cite:

@inproceedings{schemalinkIEengine2026,
author={Emanuele Cavalleri, Ali Rastegar Mojarad, J. Harry Caufield, Justin T. Reese, Christopher J. Mungall, and Marco Mesiti},
title = "{Schema-Driven Structured Information Extraction from Biomedical Literature via Large Language Models}", 
year = {2026},
publisher = {Association for Computing Machinery},
address = {New York, NY, USA},
booktitle = {Proceedings of the 35th ACM International Conference on Information and Knowledge Management},
location = {Rome, Italy},
series = {CIKM '26},
notes = {To appear.}
}