OntoGPT

Introduction

OntoGPT is a Python package for the generation of Ontologies and Knowledge Bases using GPT. It is a knowledge extraction tool that uses a Large Language Models (LLMs) to extract semantic information from text.

This makes use of so-called instruction prompts in Large Language Models (LLMs) such as GPT-4.

Currently three different strategies for knowledge extraction have been implemented in the ontogpt package:

• SPIRES: Structured Prompt Interrogation and Recursive Extraction of Semantics
  • Zero-shot learning (ZSL) approach to extracting nested semantic structures from text
  • This approach takes two inputs - 1) LinkML schema 2) free text, and outputs knowledge in a structure conformant with the supplied schema in JSON, YAML, RDF or OWL formats
  • Uses text-davinci-003 or gpt-3.5-turbo (gpt-4 untested)
• HALO: HAllucinating Latent Ontologies
  • Few-shot learning approach to generating/hallucinating a domain ontology given a few examples
  • Uses code-davinci-002
• SPINDOCTOR: Structured Prompt Interpolation of Narrative Descriptions Or Controlled Terms for Ontological Reporting
  • Summarize gene set descriptions (pseudo gene-set enrichment)
  • Uses text-davinci-003 or gpt-3.5-turbo (gpt-4 untested)

Pre-requisites

• Python 3.9+
• OpenAI account

An OpenAI key is necessary for using OpenAI's GPT models. This is a paid API and you will be charged based on usage. If you do not have an OpenAI account, you may sign up here. You will need to set your API key using the Ontology Access Kit:

poetry run runoak set-apikey -e openai <your openai api key>

You may also set additional API keys for optional resources:

• BioPortal account (for grounding). The BioPortal key is necessary for using ontologies from BioPortal. You may get a key by signing up for an account on their web site.
• NCBI E-utilities. The NCBI email address and API key are used for retrieving text and metadata from PubMed. You may still access these resources without identifying yourself, but you may encounter rate limiting and errors.
• HuggingFace Hub. This API key is necessary to retrieve models from the HuggingFace Hub service.

These optional keys may be set as follows:

poetry run runoak set-apikey -e bioportal <your bioportal api key>
poetry run runoak set-apikey -e ncbi-email <your email address>
poetry run runoak set-apikey -e ncbi-key <your NCBI api key>
poetry run runoak set-apikey -e hfhub-key <your HuggingFace Hub api key>

Setup

For feature development and contributing to the package:

git clone https://github.com/monarch-initiative/ontogpt.git
cd ~/path/to/ontogpt
poetry install

To simply start using the package in your workspace:

pip install ontogpt

Note that some features require installing additional, optional dependencies.

These may be installed as:

poetry install --extras extra_name
# OR
pip install ontogpt[extra_name]

where extra_name is one of the following:

• docs - dependencies for building documentation
• web - dependencies for the web application
• recipes - dependencies for recipe scraping and parsing
• gpt4all - dependencies for loading LLMs from GPT4All
• textract - the textract plugin
• huggingface - dependencies for accessing LLMs from HuggingFace Hub, remotely or locally

Examples

Strategy 1: Knowledge extraction using SPIRES

Input

Consider some text from one of the input files being used in the ontogpt test suite. You can find the text file here. You can download the raw file from the GitHub link to that input text file, or copy its contents over into another file, say, abstract.txt. An excerpt

The cGAS/STING-mediated DNA-sensing signaling pathway is crucial for interferon (IFN) production and host antiviral responses

... [snip] ...

The underlying mechanism was the interaction of US3 with β-catenin and its hyperphosphorylation of β-catenin at Thr556 to block its nuclear translocation ... ...

We can extract knowledge from the above text this into the GO pathway datamodel by running the following command:

Command

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

Note: The value accepted by the -t / --template argument is the base name of one of the LinkML schema / data model which can be found in the templates folder.

Output

The output returned from the above command can be optionally redirected into an output file using the -o / --output.

The following is a small part of what the larger schema-compliant output looks like:

genes:
- HGNC:2514
- HGNC:21367
- HGNC:27962
- US3
- FPLX:Interferon
- ISG
gene_gene_interactions:
- gene1: US3
  gene2: HGNC:2514
gene_localizations:
- gene: HGNC:2514
  location: Nuclear
gene_functions:
- gene: HGNC:2514
  molecular_activity: Transcription
- gene: HGNC:21367
  molecular_activity: Production
...

Working Mechanism

1. You provide an arbitrary data model, describing the structure you want to extract text into
   • This can be nested (but see limitations below)
2. Provide your preferred annotations for grounding NamedEntity fields
3. OntoGPT will:
   • Generate a prompt
   • Feed the prompt to a language model (currently OpenAI GPT models)
   • Parse the results into a dictionary structure
   • Ground the results using a preferred annotator

Strategy 2: HALO

Documentation to come

Strategy 3: Gene Enrichment using SPINDOCTOR

Given a set of genes, OntoGPT can find similarities among them.

Ex.:

ontogpt enrichment -U tests/input/genesets/sensory-ataxia.yaml

The default is to use ontological gene function synopses (via the Alliance API).

• To use narrative/RefSeq summaries, use the --no-ontological-synopses flag
• To run without any gene descriptions, use the --no-annotations flag

Features

Define your own extraction model using LinkML

There are a number of pre-defined LinkML data models already developed here - src/ontogpt/templates/ which you can use as reference when creating your own data models.

Define a schema (using a subset of LinkML) that describes the structure in which you want to extract knowledge from your text.

Gene:
  is_a: NamedThing
  id_prefixes:
    - HGNC
  annotations:
    annotators: gilda:, bioportal:hgnc-nr

Symptom:
  is_a: NamedThing
  id_prefixes:
    - HP
  annotations:
    annotators: sqlite:obo:hp

Inheritance:
  is_a: NamedThing
  annotations:
    annotators: sqlite:obo:hp
```

• Prompt hints can be specified using the prompt annotation (otherwise description is used)
• Multivalued fields are supported
• The default range is string — these are not grounded. Ex.: disease name, synonyms
• Define a class for each NamedEntity
• For any NamedEntity, you can specify a preferred annotator using the annotators annotation

We recommend following an established schema like BioLink Model, but you can define your own.

Next step is to compile the schema. For that, you should place the schema YAML in the directory src/ontogpt/templates/. Then, run the make command at the top level. This will compile the schema to Python (Pydantic classes).

Once you have defined your own schema / data model and placed in the correct directory, you can run the extract command.

Ex.:

ontogpt extract -t mendelian_disease.MendelianDisease -i marfan-wikipedia.txt

Multiple levels of nesting

Currently no more than two levels of nesting are recommended.

If a field has a range which is itself a class and not a primitive, it will attempt to nest.

Ex. the gocam schema has an attribute:

  attributes:
      ...
      gene_functions:
        description: semicolon-separated list of gene to molecular activity relationships
        multivalued: true
        range: GeneMolecularActivityRelationship

The range GeneMolecularActivityRelationship has been specified inline, so it will nest.

The generated prompt is:

gene_functions : <semicolon-separated list of gene to molecular activities relationships>

The output of this is then passed through further SPIRES iterations.

Text length limit

Currently SPIRES must use text-davinci-003, which has a total 4k token limit (prompt + completion).

You can pass in a parameter to split the text into chunks. Returned results will be recombined automatically, but more experiments need to be done to determined how reliable this is.

Schema tips

It helps to have an understanding of the LinkML schema language, but it should be possible to define your own schemas using the examples in src/ontogpt/templates as a guide.

OntoGPT-specific extensions are specified as annotations.

You can specify a set of annotators for a field using the annotators annotation.

Ex.:

  Gene:
    is_a: NamedThing
    id_prefixes:
      - HGNC
    annotations:
      annotators: gilda:, bioportal:hgnc-nr, obo:pr

The annotators are applied in order.

Additionally, when performing grounding, the following measures can be taken to improve accuracy:

• Specify the valid set of ID prefixes using id_prefixes
• Some vocabularies have structural IDs that are amenable to regexes, you can specify these using pattern
• You can make use of values_from slot to specify a Dynamic Value Set
  • For example, you can constrain the set of valid locations for a gene product to be subclasses of cellular_component in GO or cell in CL

Ex.:

classes:
  ...
  GeneLocation:
    is_a: NamedEntity
    id_prefixes:
      - GO
      - CL
    annotations:
      annotators: "sqlite:obo:go, sqlite:obo:cl"
    slot_usage:
      id:
        values_from:
          - GOCellComponentType
          - CellType

enums:
  GOCellComponentType:
    reachable_from:
      source_ontology: obo:go
      source_nodes:
        - GO:0005575 ## cellular_component
  CellType:
    reachable_from:
      source_ontology: obo:cl
      source_nodes:
        - CL:0000000 ## cell

OWL Exports

The extract command will let you export the results as OWL axioms, utilizing linkml-owl mappings in the schema.

Ex.:

ontogpt extract -t recipe -i recipe-spaghetti.txt -o recipe-spaghetti.owl -O owl

src/ontogpt/templates/recipe.yaml is an example schema that uses linkml-owl mappings.

See the Makefile for a full pipeline that involves using robot to extract a subset of FOODON and merge in the extracted results. This uses recipe-scrapers.

OWL output: recipe-all-merged.owl

Classification:

Web Application Setup

There is a bare bones web application for running OntoGPT and viewing results.

Install the required dependencies by running the following command:

poetry install -E web

Then run this command to start the web application:

poetry run web-ontogpt

Note: The agent running uvicorn must have the API key set, so for obvious reasons don't host this publicly without authentication, unless you want your credits drained.

OntoGPT Limitations

1. Non-deterministic

• This relies on an existing LLM, and LLMs can be fickle in their responses

2. Coupled to OpenAI

• You will need an OpenAI account to use their API. In theory any LLM can be used but in practice the parser is tuned for OpenAI's models

SPINDOCTOR web app

To start:

poetry run streamlit run src/ontogpt/streamlit/spindoctor.py

HuggingFace Hub

A select number of LLMs may be accessed through HuggingFace Hub. See the full list using ontogpt list-models

Specify a model name with the -m option.

Example:

ontogpt extract -t mendelian_disease.MendelianDisease -i tests/input/cases/mendelian-disease-sly.txt -m FLAN_T5_BASE

Using local models

OntoGPT supports using language models released by GPT4All.

Specify the name of a model when using the extract command with the -m or --model option and OntoGPT will retrieve the model.

For example:

ontogpt --verbose extract -t mendelian_disease.MendelianDisease -i mendelian-disease-sly.txt -m ggml-gpt4all-j-v1.3-groovy

will download the ggml-gpt4all-j-v1.3-groovy.bin file, generate a prompt, and try that prompt against the specified model.

Citation

SPIRES is described further in: Caufield JH, Hegde H, Emonet V, Harris NL, Joachimiak MP, Matentzoglu N, et al. Structured prompt interrogation and recursive extraction of semantics (SPIRES): A method for populating knowledge bases using zero-shot learning.

arXiv publication: http://arxiv.org/abs/2304.02711

Contributing

Contributions on recipes to test welcome from anyone! Just make a PR here. See this list for accepted URLs

Acknowledgements

We gratefully acknowledge Bosch Research for their support of this research project.